CN116980471A - Content pushing method, device, apparatus, storage medium and computer program product - Google Patents
Content pushing method, device, apparatus, storage medium and computer program product Download PDFInfo
- Publication number
- CN116980471A CN116980471A CN202310100967.6A CN202310100967A CN116980471A CN 116980471 A CN116980471 A CN 116980471A CN 202310100967 A CN202310100967 A CN 202310100967A CN 116980471 A CN116980471 A CN 116980471A
- Authority
- CN
- China
- Prior art keywords
- content
- screening
- exposure position
- model
- fusion
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 76
- 238000004590 computer program Methods 0.000 title claims abstract description 28
- 238000012216 screening Methods 0.000 claims abstract description 341
- 230000004927 fusion Effects 0.000 claims description 229
- 230000006399 behavior Effects 0.000 claims description 124
- 238000012360 testing method Methods 0.000 claims description 74
- 238000005192 partition Methods 0.000 claims description 61
- 238000012545 processing Methods 0.000 claims description 19
- 238000001914 filtration Methods 0.000 claims description 12
- 238000005516 engineering process Methods 0.000 abstract description 16
- 230000000694 effects Effects 0.000 abstract description 12
- 238000013473 artificial intelligence Methods 0.000 abstract description 9
- 230000006870 function Effects 0.000 description 24
- 230000003993 interaction Effects 0.000 description 20
- 230000008569 process Effects 0.000 description 14
- 230000008859 change Effects 0.000 description 8
- 230000035699 permeability Effects 0.000 description 6
- 238000012795 verification Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 5
- 230000003542 behavioural effect Effects 0.000 description 4
- 230000008901 benefit Effects 0.000 description 4
- 238000000638 solvent extraction Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000004891 communication Methods 0.000 description 3
- 230000014759 maintenance of location Effects 0.000 description 3
- 238000003062 neural network model Methods 0.000 description 3
- 238000012549 training Methods 0.000 description 3
- 238000013500 data storage Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 230000002452 interceptive effect Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000001105 regulatory effect Effects 0.000 description 2
- 230000003068 static effect Effects 0.000 description 2
- OKTJSMMVPCPJKN-UHFFFAOYSA-N Carbon Chemical compound [C] OKTJSMMVPCPJKN-UHFFFAOYSA-N 0.000 description 1
- 230000003213 activating effect Effects 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000033228 biological regulation Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 229910021389 graphene Inorganic materials 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/20—Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
- H04N21/25—Management operations performed by the server for facilitating the content distribution or administrating data related to end-users or client devices, e.g. end-user or client device authentication, learning user preferences for recommending movies
- H04N21/258—Client or end-user data management, e.g. managing client capabilities, user preferences or demographics, processing of multiple end-users preferences to derive collaborative data
- H04N21/25866—Management of end-user data
- H04N21/25891—Management of end-user data being end-user preferences
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/20—Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
- H04N21/23—Processing of content or additional data; Elementary server operations; Server middleware
- H04N21/235—Processing of additional data, e.g. scrambling of additional data or processing content descriptors
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Multimedia (AREA)
- Signal Processing (AREA)
- Computer Graphics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present application relates to a content pushing method, apparatus, computer device, storage medium and computer program product. The method relates to artificial intelligence technology, and comprises the following steps: acquiring candidate contents for pushing the content stream of the target object; for each content type, screening out primary screening content from candidate content belonging to the content type based on the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content type; for a plurality of exposure positions for pushing the content stream, starting from the first exposure position in the plurality of exposure positions, determining the position characteristics of the current exposure position; the position features comprise preamble content statistical features of the current exposure position; based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types, the exposure content of the current exposure position is screened from the preliminary screening content; and pushing the content stream to the target object according to the exposure content of each exposure position, thereby improving the pushing effect.
Description
Technical Field
The present application relates to the field of internet technology, and in particular, to a content pushing method, a content pushing device, a computer device, a storage medium, and a computer program product.
Background
With the rapid development of internet technology, various contents are pushed to objects through a recommendation system, and in a common recommendation system, multi-index fusion has wide technical application and plays a significant role in various recommendation scenes. The multi-index fusion utilizes the estimated division of the indexes to fuse, and aims to promote various service indexes.
Push systems often generate a push content stream by fusing multiple indexes into objects, the push content stream comprises a plurality of push contents, however, the current push system does not have scene self-adaption and crowd individuation based on the push content stream pushed by fusing multiple indexes, cannot meet different crowd interest demands and favorites for different content types, cannot meet different business targets, and has poor push effect.
Disclosure of Invention
In view of the foregoing, it is desirable to provide a content pushing method, apparatus, computer device, computer-readable storage medium, and computer program product that can enhance the pushing effect.
The application provides a content pushing method. The method comprises the following steps:
acquiring candidate contents for pushing the content stream of the target object; each candidate content has a corresponding content type;
for each content type, screening out primary screening content from candidate content belonging to the content type based on the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content type;
for a plurality of exposure positions for content stream pushing, starting from the first exposure position in the plurality of exposure positions, determining the position characteristics of the current exposure position; the position features comprise preamble content statistical features of the current exposure position;
based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content types, the exposure content of the current exposure position is screened out from the preliminary screening content;
and pushing a content stream to the target object according to the exposure content of each exposure position.
The application also provides a content pushing device. The device comprises:
the acquisition module is used for acquiring candidate contents for pushing the content stream of the target object; each candidate content has a corresponding content type;
The first screening module is used for screening the primary screening content from candidate content belonging to the content type based on the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content type;
a determining module, configured to determine, for a plurality of exposure positions for content stream pushing, a position feature of a current exposure position from a first exposure position of the plurality of exposure positions; the position features comprise preamble content statistical features of the current exposure position;
the second screening module is used for screening exposure content of the current exposure position from the preliminary screening content based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content type;
and the pushing module is used for pushing the content stream to the target object according to the exposure content of each exposure position.
The application also provides computer equipment. The computer device comprises a memory storing a computer program and a processor which when executing the computer program performs the steps of:
acquiring candidate contents for pushing the content stream of the target object; each candidate content has a corresponding content type;
For each content type, screening out primary screening content from candidate content belonging to the content type based on the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content type;
for a plurality of exposure positions for content stream pushing, starting from the first exposure position in the plurality of exposure positions, determining the position characteristics of the current exposure position; the position features comprise preamble content statistical features of the current exposure position;
based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content types, the exposure content of the current exposure position is screened out from the preliminary screening content;
and pushing a content stream to the target object according to the exposure content of each exposure position.
The application also provides a computer readable storage medium. The computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of:
acquiring candidate contents for pushing the content stream of the target object; each candidate content has a corresponding content type;
for each content type, screening out primary screening content from candidate content belonging to the content type based on the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content type;
For a plurality of exposure positions for content stream pushing, starting from the first exposure position in the plurality of exposure positions, determining the position characteristics of the current exposure position; the position features comprise preamble content statistical features of the current exposure position;
based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content types, the exposure content of the current exposure position is screened out from the preliminary screening content;
and pushing a content stream to the target object according to the exposure content of each exposure position.
The application also provides a computer program product. The computer program product comprises a computer program which, when executed by a processor, implements the steps of:
acquiring candidate contents for pushing the content stream of the target object; each candidate content has a corresponding content type;
for each content type, screening out primary screening content from candidate content belonging to the content type based on the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content type;
for a plurality of exposure positions for content stream pushing, starting from the first exposure position in the plurality of exposure positions, determining the position characteristics of the current exposure position; the position features comprise preamble content statistical features of the current exposure position;
Based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content types, the exposure content of the current exposure position is screened out from the preliminary screening content;
and pushing a content stream to the target object according to the exposure content of each exposure position.
The content pushing method, the device, the computer equipment, the storage medium and the computer program product are used for pushing the content stream of the target object by acquiring candidate content; each candidate content has a corresponding content type; for each content type, based on the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content type, the primary screening content is screened from the candidate content belonging to the content type, namely, the candidate content of the same content type is accurately and primarily screened by utilizing the object attribute characteristics and the behavior statistical characteristics which can highlight the individuation of the crowd of the target object. For a plurality of exposure positions for pushing the content stream, starting from the first exposure position in the plurality of exposure positions, determining the position characteristics of the current exposure position; the position features comprise preamble content statistical features of the current exposure position, namely, the position features of the current exposure position are adaptively adjusted according to related content capable of reflecting the preamble exposure position; and screening the exposure content of the current exposure position from the preliminary screening content based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types. Therefore, the exposure content of each exposure position matched with the interest requirement of the target object is screened out for the second time by combining the object attribute characteristics and the behavior statistical characteristics with crowd individuation and the current exposure position characteristics with self-adaption. According to the exposure content of each exposure position, content streams with higher suitability can be pushed to a target object, so that the pushing effect is improved.
Drawings
FIG. 1 is an application environment diagram of a content pushing method in one embodiment;
FIG. 2 is a flow chart of a content pushing method in one embodiment;
FIG. 3 is a flow chart of the steps of screening primary screen content in one embodiment;
FIG. 4 is a schematic flow chart of determining the number of primary screens in one embodiment;
FIG. 5 is a flowchart illustrating a step of selecting exposure content of a current exposure position according to an embodiment;
FIG. 6 is a flowchart illustrating a step of selecting an exposure content of a current exposure position according to another embodiment;
FIG. 7 is a flow chart of determining exposure content for a current exposure position in one embodiment;
FIG. 8 is a flowchart illustrating the update steps of the first trained object partition model in one embodiment;
FIG. 9 is a schematic diagram of model training of a first trained object partition model;
FIG. 10 is a flowchart illustrating the update steps of the second trained object partition model in one embodiment;
FIG. 11 is a block diagram of a content pushing device in one embodiment;
fig. 12 is an internal structural diagram of a computer device in one embodiment.
Detailed Description
The present application will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present application more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the application.
The embodiment of the application provides a content pushing method, which relates to an artificial intelligence (Artificial Intelligence, AI) technology, wherein the artificial intelligence is a theory, a method, a technology and an application system which simulate, extend and expand human intelligence by using a digital computer or a machine controlled by the digital computer, sense environment, acquire knowledge and acquire an optimal result by using the knowledge. In other words, artificial intelligence is an integrated technology of computer science that attempts to understand the essence of intelligence and to produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence, i.e. research on design principles and implementation methods of various intelligent machines, enables the machines to have functions of sensing, reasoning and decision.
The artificial intelligence technology is a comprehensive subject, and relates to the technology with wide fields, namely the technology with a hardware level and the technology with a software level. Artificial intelligence infrastructure technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and other directions.
In the content pushing process of the recommendation system, after finishing the fine-ranking stage, based on each candidate content obtained in the fine-ranking stage, the recommendation system needs to determine a pushing content stream from the candidate content through multi-index fusion in order to further push the content of the target object. The multi-index comprises estimated click rate, estimated duration and estimated interaction rate. However, in the process of determining the content stream through multi-index fusion, personalized characteristics of objects to be pushed in different crowds are not fully considered, and a determination process of how to adaptively adjust the pushed content in a mixed-scheduling scene of different content types is not considered. Therefore, content pushing in the related art cannot accurately push the content to be pushed in different people, that is, the pushing effect is poor.
Based on the content, the embodiment of the application provides a content pushing method, which comprises the steps of obtaining candidate content for pushing a content stream of a target object; each candidate content has a corresponding content type; for each content type, based on the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content type, the primary screening content is screened from the candidate content belonging to the content type, namely, the candidate content of the same content type is accurately and primarily screened by utilizing the object attribute characteristics and the behavior statistical characteristics which can highlight the individuation of the crowd of the target object. For a plurality of exposure positions for pushing the content stream, starting from the first exposure position in the plurality of exposure positions, determining the position characteristics of the current exposure position; the position features comprise preamble content statistical features of the current exposure position, namely, the position features of the current exposure position are adaptively adjusted according to related content capable of reflecting the preamble exposure position; and screening the exposure content of the current exposure position from the preliminary screening content based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types. Therefore, the exposure content of each exposure position matched with the interest requirement of the target object is screened out for the second time by combining the object attribute characteristics and the behavior statistical characteristics with crowd individuation and the current exposure position characteristics with self-adaption. According to the exposure content of each exposure position, content streams with higher suitability can be pushed to a target object, so that the pushing effect is improved.
The content pushing method provided by the embodiment of the application can be applied to an application environment shown in fig. 1. Wherein the terminal 102 communicates with the server 104 via a network. The data storage system may store data that the server 104 needs to process. The data storage system may be integrated on the server 104 or may be located on the cloud or other servers.
In some embodiments, server 104 obtains candidate content for content stream pushing on the target object; each candidate content has a corresponding content type; for each content type, the server 104 screens out the primary screening content from the candidate content belonging to the content type based on the object attribute feature of the target object and the behavior statistical feature of the corresponding content type; for a plurality of exposure positions for content stream pushing, the server 104 determines a position feature of a current exposure position from a first exposure position of the plurality of exposure positions; the position features comprise preamble content statistical features of the current exposure position; based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types, the server 104 screens the exposure content of the current exposure position from the preliminary screening content; according to the exposure content of each exposure position, the server 104 pushes a content stream to the target object, for example, after the server 104 determines the content stream according to the exposure content of each exposure position, the server 104 sends the content stream to a client that the target object logs in with the target account, where the client is deployed in the terminal 102, and the client may be a news client, a video client, or the like, which is not limited specifically.
The terminal 102 may be, but not limited to, various personal computers, notebook computers, smart phones, tablet computers, internet of things devices, and portable wearable devices, where the internet of things devices may be smart speakers, smart televisions, smart air conditioners, smart vehicle devices, and the like. The portable wearable device may be a smart watch, smart bracelet, headset, or the like. The server 104 may be implemented as a stand-alone server or as a server cluster of multiple servers.
In one embodiment, as shown in fig. 2, a content pushing method is provided, in which the method is applied to a computer device (e.g., server 104 in fig. 1) for explanation, and includes the following steps:
step 202, obtaining candidate contents for pushing the content stream of the target object; each candidate content has a corresponding content type.
The target object is an object to be subjected to content pushing, and the content stream is used for pushing exposure content to the target object, so that the target object can click, browse and comment on the corresponding exposure content in the client. The exposure content may be video, articles, etc. Alternatively, the content stream displays the exposure content at the exposure positions on the display screen of the terminal running the client in the order of the exposure positions.
In the content pushing process, the recalled content needs to be subjected to coarse arrangement and fine arrangement in sequence to obtain candidate content, namely the candidate content is selected in the fine arrangement stage in the content pushing process, and the content types of the candidate content are multiple, for example, the content types can be article types, video types, topic types and the like.
Optionally, the computer device obtains a plurality of candidate contents for content stream pushing on the target object, each candidate content having a corresponding content type, and at least two content types exist. The number of candidate contents belonging to the content type may be one or a plurality of candidate contents for each content type, and is not particularly limited.
Illustratively, a computer device obtains a plurality of candidate content in a target push scenario for a target object, each candidate content having a corresponding content type. The target pushing scene can be a news pushing scene, an advertisement pushing scene, a film and television pushing scene and the like, and is not particularly limited.
Step 204, for each content type, screening out the primary screening content from the candidate content belonging to the content type based on the object attribute feature of the target object and the behavior statistical feature of the corresponding content type.
Wherein the object attribute features characterize the underlying attribute features of the target object. The object attribute characteristics of the target object may be age, sex, user liveness in each client, and the like of the target object. For any crowd formed by a plurality of objects, the crowd contains target objects, and the object attribute characteristics of each object in the crowd are different, in other words, the object attribute characteristics can also represent the crowd characteristics of the target objects in the crowd, and the object attribute characteristics of each of any two objects in the crowd are different, so that the personalized characteristics of the target objects in the crowd can be reflected according to the object attribute characteristics.
The behavior statistical feature characterizes historical behavior features of the target object in different content types, wherein the behavior statistical feature is a feature obtained based on real-time portrait of the target object and real-time behavior sequence statistics, and the behavior statistical feature can be, for example, historical playing time length, historical clicking and the like of the target object in different content types.
Optionally, for each content type, the computer device screens out the prescreened content from at least one candidate content belonging to the targeted content type based on the object attribute characteristics of the target object and the behavioral statistics characteristics of the targeted content type.
Illustratively, for each content type, the computer device determines a preliminary screening score for each screened content belonging to the targeted content type based on the object attribute characteristics of the target object and the behavioral statistics characteristics of the targeted content type. The computer device determines the preliminary screening content from at least one screening content belonging to the targeted content type.
Step 206, for a plurality of exposure positions for pushing the content stream, starting from the first exposure position in the plurality of exposure positions, determining the position characteristics of the current exposure position; the location features include preamble statistics of the current exposure location.
Wherein the preamble content statistics of the current exposure position are determined based on at least one preamble exposure position located before the current exposure position, the preamble content statistics including a number of content types of the preamble exposure position and a diversity statistics of the preamble exposure position. Illustratively, the current exposure position is the 5 th exposure position, the current exposure position corresponds to 4 preamble exposure positions, the 4 exposure positions relate to 2 content types, and the number of content types exposed by the preamble positions is 2. Diversity statistics characterizes the diversity nature of the exposure locations of the preamble locations, e.g., for article types, diversity refers to the diversity of some article attributes, such as how much recommended by different categories (entertainment, life, etc.), etc. It follows that the position feature of the current exposure position can be determined after the exposure content of the preamble exposure position is determined.
Optionally, for a plurality of exposure positions for content stream pushing, the computer device determines a current exposure position starting from a first exposure position of the plurality of exposure positions. If the computer equipment verifies that the current exposure position is the first exposure position, the computer equipment determines the position characteristic of the current exposure position based on the position information of the current exposure position. If the computer equipment verifies that the current exposure position is not the first exposure position, the computer equipment determines at least one preamble position before the current exposure position and determines preamble content statistical characteristics of the current exposure position according to the at least one preamble position. The computer device determines the position characteristics of the current exposure position according to the preamble content statistical characteristics of the current exposure position.
Illustratively, after the computer device determines the current exposure position, if the computer device verifies that the current exposure position is the first exposure position, the computer device determines a location feature of the current exposure position based on location information of the current exposure position. If the computer equipment verifies that the current exposure position is not the first exposure position, the computer equipment determines at least one preamble position before the current exposure position and determines preamble content statistical characteristics of the current exposure position according to the at least one preamble position. The computer device determines the position information of the current exposure position and determines the position characteristics of the current exposure position according to the preamble content statistical characteristics and the position information of the current exposure position. The position information of the current exposure position includes a position (e.g., what exposure position the current exposure position is) where the current exposure position is in the content stream, and also includes a content type for which the current exposure position allows exposure.
And step 208, screening the exposure content of the current exposure position from the preliminary screening content based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types.
Optionally, the computer device screens the exposure content of the current exposure position from the preliminary screening content of each content type based on the position feature of the current exposure position, the object attribute feature of the target object and the behavior statistical feature of the corresponding content type.
As previously described, the determination of the position feature of the current exposure position takes into account the preamble exposure position preceding the current exposure position. In the process of determining the exposure content of the current exposure position, not only the current exposure position is considered, but also the content information of the preamble exposure position is considered, so that the accuracy and the reliability of the exposure content of each exposure position are improved.
Step 210, pushing a content stream to the target object according to the exposure content of each exposure position.
Optionally, the computer device obtains the exposure content of each exposure location and pushes a content stream to the target object according to the exposure content of each exposure location.
Illustratively, the computer device obtains the exposure content for each exposure position, and determines a content stream in accordance with the position order of each exposure position and the exposure content for each exposure position, and pushes the content stream to the target object.
In the content pushing method, candidate content for pushing the content stream of the target object is obtained; each candidate content has a corresponding content type; for each content type, based on the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content type, the primary screening content is screened from the candidate content belonging to the content type, namely, the candidate content of the same content type is accurately and primarily screened by utilizing the object attribute characteristics and the behavior statistical characteristics which can highlight the individuation of the crowd of the target object. For a plurality of exposure positions for pushing the content stream, starting from the first exposure position in the plurality of exposure positions, determining the position characteristics of the current exposure position; the position features comprise preamble content statistical features of the current exposure position, namely, the position features of the current exposure position are adaptively adjusted according to related content capable of reflecting the preamble exposure position; and screening the exposure content of the current exposure position from the preliminary screening content based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types. Therefore, the exposure content of each exposure position matched with the interest requirement of the target object is screened out for the second time by combining the object attribute characteristics and the behavior statistical characteristics with crowd individuation and the current exposure position characteristics with self-adaption. According to the exposure content of each exposure position, content streams with higher suitability can be pushed to a target object, so that the pushing effect is improved.
For each candidate content, the target object may click on the candidate content according to the interest degree or the requirement of the target object, or browse for a while after clicking on the candidate content, or forward the candidate content, that is, the target object may generate different behaviors for the same candidate content. Therefore, in order to ensure the accuracy of content pushing, each content may be comprehensively evaluated by a plurality of different preset indexes.
To this end, in some embodiments, for each content type, the screening of the preliminary content from candidate content belonging to the content type based on the object attribute features of the target object and the behavior statistics of the corresponding content type, respectively, includes: aiming at a target object, obtaining the estimated score of each candidate content corresponding to each preset index; and for each content type, respectively fusing the estimated scores of the preset indexes corresponding to the candidate contents belonging to the content type based on the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content type, and screening the primary screening content from the candidate contents belonging to the content type according to the fusion result.
The preset indexes can be duration, interaction rate, consumption, click rate, forwarding quantity, praise, comment quantity and the like. The pre-estimated score of the preset index is an estimate of the preset index. The estimated score may be regarded as the score obtained in the fine-pitch stage.
Optionally, for each candidate content, the computer device obtains a pre-estimated score of each preset index corresponding to the candidate content for the target object. For each candidate content, the computer equipment determines the content type of the targeted candidate content, and based on the object attribute characteristics of the target object and the behavior statistical characteristics of the content type, the computer equipment fuses the estimated scores of the preset indexes corresponding to the targeted candidate content, and then determines the fusion result of the targeted candidate content. For each content type, the computer device filters at least one candidate content belonging to the targeted content type based on the fusion result of each candidate content belonging to the targeted content type, and determines primary screening content.
For example, there are two types of content, namely an article type and a video type, and there are 2 candidate contents belonging to the article type, namely an article 1 and an article 2, and 1 candidate content belonging to the video type, namely a video 1. Taking article 1 as an example, the duration of the article 1 is 30 minutes, the click rate is 80%, namely the pre-estimated score of the preset index is the duration is 30 minutes, the pre-estimated score of the preset index is 0.8, based on the pre-estimated score of the click rate, the computer equipment fuses the pre-estimated score of the duration and the pre-estimated score of the click rate based on the object attribute characteristics of the target object and the behavior statistical characteristics of the article types, and then a fusion result of the article 1 is obtained. Likewise, the fusion result of article 2 and the fusion result of video 1 are determined. The computer equipment screens the article 1 and the article 2 according to the fusion result of the article 1 and the fusion result of the article 2, and determines the primary screening content of the article type; similarly, the computer device determines the initial content of the video type according to the fusion result of the video 1.
Of course, in practical application, the number of candidate contents related to the application is more, the number of content types is also more, the number of preset indexes is also more, and the application is not limited in particular.
In this embodiment, for each content type, based on the object attribute feature of the target object and the behavior statistical feature corresponding to the content type, the preset scores of the preset indexes corresponding to the candidate content belonging to the content type can be fused, that is, different preset indexes are fused, and each candidate content belonging to the same content type can be effectively and accurately evaluated, so as to obtain the fusion result. Based on the above, according to the fusion result of each candidate content belonging to the same content type, the effective screening of the candidate content of each content type is realized. Meanwhile, the object attribute characteristics and the behavior statistical characteristics of the target object are combined, so that a screening process aiming at the target object can be performed, and the pertinence of the follow-up content pushing is further ensured.
In some embodiments, for a target object, obtaining a pre-estimated score of each candidate content corresponding to each preset index includes: determining a preset index for evaluating the candidate content; and for any candidate content, estimating the estimated score of the preset index corresponding to the candidate content for the target object according to the object attribute characteristics of the target object and the content characteristics of the candidate content by a trained estimated model corresponding to the preset index.
The trained pre-estimated model is a trained pre-estimated model, the trained pre-estimated model corresponding to a preset index is used for pre-estimating the pre-estimated score of the preset index, for example, if the preset index is a duration, the trained pre-estimated model corresponding to the preset index is a duration pre-estimated model; for another example, if the preset index is the click rate, the trained prediction model corresponding to the preset index is the click rate prediction model. The content characteristics of the candidate content characterize the content characteristics of the candidate content.
Optionally, the computer device determines a preset index for evaluating the candidate content. For any candidate content, the computer equipment determines trained pre-estimated models corresponding to all preset indexes respectively. For each candidate content and each preset index, the computer equipment predicts the estimated score of the corresponding preset index of the candidate content for the target object according to the object attribute characteristics of the target object and the content characteristics of the candidate content through the trained estimated model corresponding to the preset index.
Taking an example of estimating the estimated score of the duration corresponding to the article 1, for the article 1, the computer device obtains a trained estimated model a corresponding to the duration, where the trained model a is used for estimating the estimated score of the duration as the preset index. The computer equipment predicts the estimated score of the duration corresponding to the article 1 for the target object according to the object attribute characteristics of the target object and the content characteristics of the article 1 through the trained estimated model A, namely, the duration of browsing the article 1 for the target object is estimated for the target object.
In this embodiment, after determining the preset indexes for evaluating the candidate contents, for each candidate content, the object attribute feature and the content feature of the candidate content are respectively input into the trained prediction model corresponding to each preset index, so that for each candidate content, the predicted score of the target object in each preset index can be rapidly and accurately predicted.
In some embodiments, as shown in fig. 3, a flow chart of the step of screening the primary screen content in one embodiment is shown. Based on the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types, after the pre-estimated scores of the corresponding preset indexes of the candidate content belonging to the content types are fused, the primary screening content is screened from the candidate content belonging to the content types according to the fusion result, and the method comprises the following steps:
step 302, obtaining the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types.
Step 304, determining a first fusion parameter corresponding to the content type based on the object attribute feature and the behavior statistical feature.
The first fusion parameter is a fusion parameter corresponding to the primary screening content and is used for fusing the estimated scores of a plurality of preset indexes belonging to the same content type. The first fusion parameters comprise fusion parameters respectively corresponding to preset indexes, each fusion parameter corresponding to each preset index comprises a power weight and a bias constant corresponding to the preset index, for example, the preset index is a duration, the corresponding fusion parameter is the power weight and the duration bias constant of the duration, the preset index is a click rate, the corresponding fusion parameter is the power weight and the click rate bias constant of the click rate, the preset index is a long click rate, the corresponding fusion parameter is the power weight and the long click bias constant of the long click rate, the long click is the target object click, and the click of the preset duration is stopped.
Optionally, for each content type, the computer device determines a first fusion parameter corresponding to the content type for which the object attribute feature of the target object is intended based on the behavior statistics feature of the content type for which the object attribute feature is intended.
And 306, fusing the estimated scores of the preset indexes corresponding to the candidate contents belonging to the content type according to the first fusion parameters corresponding to the content type to obtain the preliminary screening score of the candidate contents.
Optionally, for each content type, the computer device determines a first screening function corresponding to the content type according to a first fusion parameter corresponding to the content type, and fuses the estimated scores of the preset indexes corresponding to the candidate content belonging to the content type by using the first screening function to obtain the primary screening score of the candidate content.
For each content type, the computer device determines a fusion parameter corresponding to each preset index based on a first fusion parameter corresponding to the content type, constructs a first screening function, determines the first screening function corresponding to the content type, and fuses the estimated scores of the preset indexes corresponding to the candidate content belonging to the content type by using the first screening function to obtain a primary screening score of the candidate content.
The step of constructing a first screening function includes: for each preset index, the computer equipment determines factors corresponding to the preset index according to the power weight, the bias constant and the pre-estimated score of the preset index, and constructs a first screening function based on the product of the factors corresponding to the preset indexes.
For example, for each preset index, the computer device determines a sum of the bias constant and the pre-estimated score of the preset index for which it is intended, and determines, by a power operation, a factor corresponding to the preset index for which it is intended, based on the sum and the power weight of the preset index for which it is intended. Wherein the bias constant is a small number, and avoids zero primary screening score caused by zero factor of corresponding preset index.
For example, there are four preset indexes, namely, a duration, a click rate, a long click rate and an interaction rate, and considering that before a user does not click, corresponding content is automatically played in some clients, for example, in a video client, when the user browses a video list, the user is in a preset browsing position on the video list, so that the user can quickly browse, based on the preset indexes of the duration, an automatic playing duration is further set, and based on the automatic playing duration, a first screening function is as follows:
Wherein score merge1 For the primary screening, stayTime is the duration, autoPlayTime is the automatic play duration, beta time For a duration bias constant, alpha time Is the power of duration weight; realClick is click Rate, beta RealClick For click rate bias constant, α RealClick DeepClick is the power of click rate, beta DeepClick Bias constant for long click rate, alpha DeepClick Power of long click rate; interaction is Interaction rate, beta Interaction For the interaction rate bias parameter, α Interaction Is the power of the interaction rate.
And step 308, screening the primary screening content from the candidate content belonging to the content type according to the primary screening number.
Optionally, for each content type, the computer device determines at least one candidate content belonging to the targeted content type and screens out initial content from among the at least one candidate content belonging to the targeted content type according to a preliminary screening score of the at least one candidate content belonging to the targeted content type.
For each content type, the computer device determines at least one candidate content belonging to the targeted content type, and screens out initial content from the at least one candidate content belonging to the targeted content type according to a preliminary screening rule corresponding to the targeted content type according to a preliminary screening score corresponding to the targeted content type. The primary screening rule is to sort primary screening scores from large to small, and select a preset number of candidate contents from a first sequence number; illustratively, T0PN (i.e., the top N candidate content) is selected, e.g., the top 5 candidate content is selected. Of course, it is also possible to sort the initial scores from small to large, and select a preset number of candidate contents, such as selecting the last 5 candidate contents, from the last sequence number.
Considering different recommendations, different crowds and different consumption scenes, the requirements for different content types are different, and based on the requirements, corresponding prescreening rules can be set according to the requirements. For example, for an object whose terminal is connected to a video preference class of wifi (wireless fidelity, wireless network), longer videos are preferred, so that for the object, the number of video types of content in a content stream that is ultimately pushed to the object is higher than other content types, based on which a rule of prescreening of each content type, for example, the number of prescreened content of a video type is higher than other content types, can be adaptively determined. Likewise, for an object that resides in a first-line city and is interested in capturing the graphic preference class of news information in fragmented time, reducing video medium recommendations tends to recommend short graphic class news, so that the number of chapter-type content in the content stream that is ultimately pushed to the object is higher for that object than for other content types.
In this embodiment, a first fusion parameter corresponding to a content type is determined according to an object attribute feature and a behavior statistical feature of the content type, based on this, for each candidate content belonging to the content type, the estimated scores of a plurality of preset indexes of the candidate content are fused according to the corresponding first fusion parameter, that is, for each content type, comprehensive evaluation is performed based on a plurality of preset indexes of each content, so as to complete the screening process comprehensively and accurately, and accuracy and effectiveness of the primary screening content are ensured.
In some embodiments, determining a first fusion parameter corresponding to the content type based on the object attribute feature and the behavior statistics feature includes: inputting the object attribute characteristics and the behavior statistical characteristics of the corresponding content types into a first trained object division model; outputting first fusion parameters corresponding to the content types aiming at the target object through a first trained object division model, wherein the first fusion parameters comprise fusion parameters respectively corresponding to preset indexes, and the first fusion parameters are used for fusing estimated scores of the preset indexes corresponding to candidate contents estimated aiming at the target object.
The first trained object partition model is a trained model and is used for determining a first fusion parameter corresponding to the content type. The first trained object partition model is a deep neural network model (Deep Neural Networks, DNN for short).
Optionally, the computer device obtains a first trained object partition model, and inputs the object attribute features and the behavior statistical features of the corresponding content types into the first trained object partition model. And outputting a first fusion parameter corresponding to the content type aiming at the target object through the first trained object division model.
It should be noted that, the determination of the first fusion parameter needs to know the object attribute characteristics of the target object. The object attribute characteristics are different for different objects, and then the first fusion parameters corresponding to the respective objects are different. That is, the first trained object partition model can be adaptively adjusted in real time according to different objects and different content types, so that the first fusion parameters are obtained by considering the objects and the content types, thousands of people and thousands of models are realized, and the first fusion parameters are more accurate.
Illustratively, the computer device obtains a first trained object partition model and inputs object attribute features and behavioral statistics of corresponding content types into the first trained object partition model. And outputting fusion parameters corresponding to each preset index respectively aiming at the target object through the first trained object division model. That is, after the object attribute features and the behavior statistical features are input to the first trained object partition model, there are a plurality of outputs, such as a simultaneous output duration, a click rate, an interaction rate, and the like.
In this embodiment, based on the object attribute feature and the behavior statistical feature of the content type, the first fusion parameter considering both the target object and the content type can be obtained through the first trained object partition model. Therefore, the first fusion parameters can be automatically adjusted according to the target object and the content type without manual guidance adjustment, and the crowd individuation and self-adaption are realized.
In some embodiments, as shown in FIG. 4, a schematic flow chart of determining the number of primary screens in one embodiment is shown. The computer equipment inputs the object attribute characteristics and the behavior statistical characteristics of the corresponding content types into a first trained object division model; outputting first parameters to be processed corresponding to the content type aiming at the target object through a first trained object division model, wherein the first parameters to be processed comprise fusion parameters respectively corresponding to preset indexes. And the computer equipment obtains a first fusion parameter corresponding to the content type through activating the function according to the first parameter to be processed corresponding to the content type. For each content type, the computer equipment determines fusion parameters corresponding to all preset indexes based on first fusion parameters corresponding to the content type, constructs a first screening function, determines the first screening function corresponding to the content type, and fuses the estimated scores of all preset indexes corresponding to the candidate content by using the first screening function to obtain the primary screening score of the candidate content.
In this embodiment, based on the object attribute feature and the behavior statistical feature of the content type, the first fusion parameter considering both the target object and the content type can be obtained through the first trained object partition model. Based on the above, for each content type, for each candidate content belonging to the content type, the pre-estimated scores of the multiple preset indexes of the candidate content are fused according to the corresponding first fusion parameters, that is, for each content type, comprehensive evaluation is performed based on the multiple preset indexes of each content, so that the screening process is completed comprehensively and accurately, and the accuracy and effectiveness of the primary screening content are ensured.
In some embodiments, as shown in fig. 5, a flow chart of the step of screening the exposure content of the current exposure position in one embodiment is shown. Based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types, the exposure content of the current exposure position is screened from the preliminary screening content, and the method comprises the following steps:
step 502, determine the above content characteristics of the current exposure position.
The above content feature of the current exposure position is a content feature of the content of the previous exposure position, and for example, the above content feature of the current exposure position may be a content feature of the exposure content of the previous exposure position adjacent to the current exposure position, or may be a content feature of the exposure content of a plurality of previous exposure positions, which is not particularly limited.
Optionally, the computer device determines a current exposure position and determines at least one preamble exposure position of the current exposure position. The computer device determines the above content features for the current exposure position based on the content features of the exposure content for the at least one preamble exposure position.
Illustratively, after the computer device determines at least one of the preamble exposure positions of the current exposure position, the computer device takes the preamble exposure position adjacent to the current exposure position as a target preamble exposure position and takes the content characteristics of the exposure content of the target preamble exposure position as the above content characteristics of the current exposure position. Alternatively, the computer device takes the content characteristics of the exposure content of each of the preamble exposure positions as the above content characteristics of the current exposure position.
Step 504, based on the above content characteristics and the content characteristics of the primary screening content, a re-estimation score of the primary screening content corresponding to each preset index is determined.
The re-estimation score of each preset index is used for screening out the exposure content of the current exposure position.
Optionally, for each of the primary screening contents, the computer device obtains content characteristics of the targeted primary screening content, and determines a re-estimation score of the targeted primary screening content corresponding to each preset index based on the above content characteristics of the current exposure position and the content characteristics of the targeted primary screening content.
Step 506, for each content type, based on the position feature, the object attribute feature of the target object and the behavior statistical feature of the corresponding content type, the overestimated scores of the preset indexes corresponding to the preliminary screening content belonging to the content type are fused, and then the exposure content of the current exposure position is screened from the preliminary screening content according to the fusion result.
Optionally, for each primary screening content, the computer device determines a content type to which the targeted primary screening content belongs, and based on the position feature, the object attribute feature of the target object and the behavior statistical feature of the content type, determines a fusion result of the targeted primary screening content after fusing the re-estimation scores of preset indexes corresponding to the targeted primary screening content. And the computer equipment screens out the exposure content of the current exposure position according to the fusion result of the primary screening content of each content type.
Illustratively, for all of the primary screened content, there are 2 content types, 3 for the article type and the video type, and 4 for the primary screened content belonging to the article type. After determining the fusion results for each of the primary screen content. The computer equipment determines target preset screening conditions corresponding to the preamble exposure positions, and screens out exposure contents of the current exposure positions according to the target preset screening conditions and the fusion result of the primary screening contents.
It should be noted that, the fusion result of each primary screen content may be regarded as a fusion score. According to the actual service requirement, a plurality of preset screening conditions exist, for example, the preset screening conditions can be to take the content with the highest fusion score in the current exposure position as the exposure content of the current exposure position. The preset screening condition may be to randomly select one preliminary screening content from at least one preliminary screening content of the current exposure position as the exposure content of the current exposure position. The target preset screening condition of the current exposure position and the target preset screening condition of the preamble exposure position of the current exposure position are the same, that is, before determining the exposure content of the first exposure position, the computer device sequentially determines the exposure content of each exposure position from a plurality of preset screening conditions based on the target preset screening conditions.
In this embodiment, according to the above content features of the current exposure position and the content features of the preliminary screening content, the content features of the context are comprehensively considered, and the re-estimation number of the preliminary screening content corresponding to each preset index is re-estimated, so that the effectiveness and accuracy of the exposure content of the current exposure position can be further ensured to be determined subsequently. Based on the position characteristics, the object attribute characteristics and the behavior statistical characteristics of the corresponding content types, the overestimated scores of the preset indexes corresponding to the preliminary screening content belonging to the content types are fused, and then the exposure content of the current exposure position is accurately determined according to the fusion result.
In some embodiments, determining the overestimated score of the preliminary screening content for each of the preset indicators based on the above content features and the content features of the preliminary screening content includes: determining a preset index for evaluating the primary screening content; and for any primary screening content, estimating the re-estimation number of the preset index corresponding to the primary screening content according to the content characteristics of the primary screening content and the content characteristics of the primary screening content through a trained re-estimation model corresponding to the preset index.
The trained re-estimation model is a trained re-estimation model, and the trained estimation model corresponding to the preset index is used for re-estimating the score of the preset index. For example, if the preset index is a duration, the trained re-estimation model corresponding to the preset index is a duration re-estimation model; for another example, if the preset index is the click rate, the trained re-estimation model corresponding to the preset index is the click rate re-estimation model.
Optionally, the computer device determines a preset index for evaluating the primary screening content. For any primary screening content, the computer equipment determines trained re-estimation models corresponding to all preset indexes respectively. For each primary screening content and each preset index, the computer equipment predicts the re-estimation score of the preset index corresponding to the primary screening content according to the content characteristics of the primary screening content and the content characteristics of the primary screening content by a trained re-estimation model corresponding to the preset index.
Taking article 3 as an example, for the article 3, the computer device obtains a trained re-estimation model a corresponding to the duration, where the trained model a is used for estimating a re-estimation score with a preset index as the duration. The computer equipment predicts the re-estimation score of the duration corresponding to the article 3 according to the content characteristics of the article 3 and the content characteristics of the article 3 through the trained pre-estimation model a, namely, the re-estimation of the duration of the article 3 browsed by the target object is performed according to the content characteristics of the article 3 and the content characteristics of the article 3.
In this embodiment, for any primary screening content, based on the above content features and the content features of the primary screening content, the content features of the context are comprehensively considered, and the re-estimation number of each preset index corresponding to the primary screening content is re-estimated, so that the validity and accuracy of the exposure content of the current exposure position can be further ensured to be determined subsequently.
In some embodiments, as shown in fig. 6, a flowchart of a step of screening exposure content of a current exposure position in another embodiment is shown. Based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types, after the overestimation scores of the corresponding preset indexes of the preliminary screening content belonging to the content types are fused, the exposure content of the current exposure position is screened from the preliminary screening content according to the fusion result, and the method comprises the following steps:
step 602, determining a second fusion parameter corresponding to the content type at the current exposure position based on the position feature, the object attribute feature of the target object and the behavior statistical feature of the corresponding content type.
The second fusion parameters are used for fusing the re-estimation scores of the primary screening content corresponding to the preset indexes. The second fusion parameters comprise fusion parameters respectively corresponding to the preset indexes, and the fusion parameters corresponding to each preset index comprise power weights and bias constants corresponding to the preset indexes. For example, if the preset index is a duration, the corresponding fusion parameter is a power weight and a duration bias constant of the duration, if the preset index is a click rate, the corresponding fusion parameter is a power weight and a click rate bias constant of the click rate, if the preset index is a long click rate, the corresponding fusion parameter is a power weight and a long click bias constant of the long click rate, and the long click is a target object click, and the click of the preset duration is stopped.
Optionally, for each content type, the computer device determines a second fusion parameter corresponding to the targeted content type at the current exposure position based on the position feature of the current exposure position, the object attribute feature of the target object, and the behavior statistics feature of the targeted content type.
Step 604, according to the second fusion parameters corresponding to the content type at the current exposure position, the re-estimation scores of the pre-screening content belonging to the content type corresponding to the preset indexes are fused, so as to obtain the fusion score corresponding to the pre-screening content belonging to the content type at the current exposure position.
Optionally, for each content type, the computer device determines a second screening function corresponding to the content type according to a second fusion parameter corresponding to the content type at the current exposure position, and facilitates the second screening function to fuse the re-estimated scores of the pre-screened content corresponding to the content type to obtain the fusion score corresponding to the screened content at the current exposure position. The second fusion parameters further include fusion parameters corresponding to the new indexes, for example, the new indexes include diversity indexes and exploration indexes, and the fusion parameters corresponding to each new index include power weights and bias constants corresponding to the new indexes. The diversity index reflects the category of content belonging to the same content type, e.g., for article types, there are a number of different categories, such as entertainment, life, etc. The search index reflects a search for an interest in the object, for example, a search for a degree of interest in each content by the object based on an existing interest representation of the object.
For each content type, the computer device determines a fusion parameter corresponding to each preset index and a fusion parameter corresponding to each new index according to a second fusion parameter corresponding to the content type, constructs a second screening function, determines a second screening function corresponding to the content type, and fuses the overestimated number of each preset index corresponding to the primary screening content belonging to the content type and the new index number corresponding to each new index by using the second screening function to obtain a fusion number of the primary screening content.
The new index score of the diversity index is determined by a diversity strategy, wherein the diversity strategy refers to how much diversity entropy increase can be brought by the current content of the current exposure position compared with the content exposed in the previous period. The current content is the same type of content as the exposed content. The new index score of the exploration index is determined through an exploration strategy, wherein the exploration strategy refers to the exploration degree of the current content on other contents compared with the existing interest portrait of the object, and the current content is the same as the content type of the other contents.
The step of constructing a second screening function includes: for each preset index, the computer device determines a first factor corresponding to the preset index according to the power weight, the bias constant and the re-estimation number of the preset index. For each new indicator, the computer device determines a second factor corresponding to the preset indicator for which it is intended, based on the power weight, the bias constant, and the new indicator score for which it is intended. The computer equipment constructs a second screening function based on the product of the first factors corresponding to the preset indexes and the product of the second factors corresponding to the new indexes.
For example, for each preset indicator, the computer device determines a sum of the bias constant and the pre-estimated score of the preset indicator for which it is intended, and determines, by a power operation, a first factor corresponding to the preset indicator for which it is intended, according to the sum and the power weight of the preset indicator for which it is intended. For each new indicator, the computer device determines a sum of the bias constant and the score of the new indicator for each new indicator, and determines a second factor corresponding to the new indicator for each new indicator by power operation according to the sum and the power weight of the new indicator for each new indicator.
Wherein the bias constant is a small number, and avoids the fusion score being zero due to zero of the first factor of the corresponding preset index or zero of the second factor of the corresponding new index.
For example, four preset indexes are time length, click rate, long click rate and interaction rate, and various indexes and exploration indexes. Considering that before the user does not click, in some clients, corresponding content is automatically played, for example, in a video client, when the user browses a video list, the user is enabled to quickly browse the video at a preset browsing position on the video list, and based on the preset index of the duration, the automatic playing duration is also set. Based on this, the second screening function is as follows:
wherein score merge2 To merge the scores, stayTime is the duration, autoPlayTime is the automatic play duration, β time For a duration bias constant, alpha time Is the power of duration weight; realClick is click Rate, beta RealClick For click rate bias constant, α RealClick DeepClick is the power of click rate, beta DeepClick Bias constant for long click rate, alpha DeepClick Power of long click rate; interaction is Interaction rate, beta Interaction For the interaction rate bias parameter, α Interaction Is the power of the interaction rate; diversity is Diversity, beta Diversity For diversity bias parameter, alpha Diversity Power of diversity weights; explore is explored, beta Explore To explore the bias parameters, α Explore Is the power of exploration weight.
And 608, screening the exposure content of the current exposure position from the preliminary screening content according to the fusion score.
Optionally, the computer device determines a target preset screening condition corresponding to the preamble exposure position, and screens out the exposure content of the current exposure position according to the target preset screening condition according to the fusion score of each primary screening content.
For example, if the target preset screening condition is that the content with the highest fusion score in the current exposure position is taken as the exposure content of the current exposure position, the computer equipment takes the corresponding primary screening content with the highest fusion score as the exposure content of the current exposure position according to the fusion score of each primary screening content of the current exposure position.
In this embodiment, to further ensure accuracy and effectiveness of the second fusion parameter, the second fusion parameter is determined by a location feature capable of reflecting context information, an object attribute feature reflecting crowd personalization, and a behavior statistical feature reflecting content type. In this way, the pre-estimated scores of a plurality of preset indexes of the preliminary screening content are fused through the second fusion parameters corresponding to the content type at the current exposure position, so that the exposure content at the current exposure position can be accurately determined.
In some embodiments, determining the second fusion parameter corresponding to the content type at the current exposure position based on the location feature, the object attribute feature of the target object, and the behavior statistical feature of the corresponding content type includes: and inputting the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types into a second trained object division model. And outputting second fusion parameters corresponding to the content type at the current exposure position aiming at the target object through a second trained object division model, wherein the second fusion parameters comprise fusion parameters respectively corresponding to preset indexes, and the second fusion parameters are used for fusing the re-estimation scores of the preset indexes corresponding to the preliminary screening content re-estimated aiming at the target object.
The second trained object partition model is a trained model and is used for determining a second fusion parameter corresponding to the content type. The second trained object partition model is a deep neural network model (Deep Neural Networks, DNN for short).
Optionally, the computer device obtains a second trained object partition model, and inputs the position feature, the object attribute feature and the behavior statistical feature of the corresponding content type of the current exposure position into the second trained object partition model. And outputting a second fusion parameter corresponding to the content type at the current exposure position aiming at the target object through a second trained object division model.
It should be noted that, for different objects, the object attribute features are different, and then the second fusion parameters corresponding to the respective objects are different. That is, the second trained object partition model can be adaptively adjusted in real time according to different objects and different content types, so that the second fusion parameters are obtained by considering the objects and the content types, thousands of people and thousands of models are realized, and the first fusion parameters are more accurate. In addition, based on the above information of the preamble position, the accuracy of the second fusion parameter can be further improved.
Illustratively, the computer device inputs the location feature, the object attribute feature, and the behavioral statistics of the corresponding content type for the current exposure location into a second trained object partition model. And outputting fusion parameters respectively corresponding to each preset index and fusion parameters respectively corresponding to each new index aiming at the target object through a second trained object division model. That is, there are multiple outputs after the object attribute features and the behavior statistical features are input to the first trained object partition model.
In this embodiment, by using the second trained object division model, the three aspects of the current exposure position, crowd individuation and content type can be integrated, and the second fusion parameter can be adaptively adjusted, without manual guidance and adjustment, so that the reliability of the second fusion parameter is greatly improved.
In the process of determining the fusion score, in order to promote the comparability among the content types, the primary screening content belonging to each content type is regulated through the score regulating parameter corresponding to each content type so as to promote the comparability among the content types, and a multi-content type mixed arrangement of real-time fusion is formed.
To this end, in some embodiments, the method further comprises: and determining a score adjustment parameter corresponding to the content type at the current exposure position based on the position characteristic, the object attribute characteristic of the target object and the behavior statistical characteristic of the corresponding content type.
It should be noted that, for the current exposure position, after determining the fusion score of each primary screen content, the fusion score of each primary screen content of the article type may be high overall, the fusion score of the primary screen content of the video type is low overall, and a deviation exists between the article type and the video type, and the deviation is a distribution difference of estimated fusion scores from a single content type. In practice, the target object may be more video-prone, based on which the fusion score can be modified by a score adjustment parameter, i.e. unifying a criterion, to enable the scores of different content types to be compared.
In this embodiment, the score adjustment parameters corresponding to the corresponding content types at the current exposure positions can be determined in real time through the position features, the object attribute features of the target object, and the behavior statistical features of the corresponding content types.
After determining the score adjustment parameters corresponding to the content type at the current exposure position, in some embodiments, filtering the exposure content at the current exposure position from the preliminary screening content according to the fused score includes: adjusting the fusion score corresponding to the primary screening content belonging to the content type at the current exposure position according to the score adjustment parameter corresponding to the content type at the current exposure position to obtain the adjustment score corresponding to the primary screening content belonging to the content type at the current exposure position; and screening the exposure content of the current exposure position from the preliminary screening content according to the adjustment score of each preliminary screening content corresponding to the current exposure position.
Optionally, the computer device uses a product of a score adjustment parameter corresponding to the current exposure position of the content type and a fusion score corresponding to the current exposure position of the primary screening content belonging to the content type as an adjustment score corresponding to the current exposure position of the primary screening content belonging to the content type. The computer equipment determines a target preset screening condition corresponding to the preamble exposure position, and screens out the exposure content of the current exposure position according to the target preset screening condition according to the adjustment score of each preliminary screening content.
Illustratively, for the current exposure position, there is an existing article 1 and video 1, the fusion score of article 1 is 5, and the fusion score of video 1 is 0.1. It is now determined that the respective score adjustment parameters for the article type and the video type are respectively for 0.1 and 6. The score-based adjustment parameter can reflect that the target object is more prone to video type, based on which the adjustment score for article 1 is 0.5; the adjustment score for video 1 was 0.6. If the target preset screening condition is that the content with the highest score is taken as the exposure content of the current exposure position, the exposure content of the current exposure position is video 1.
In this embodiment, because there is a difference between the content types, the value ranges of the scores of the content types are also uneven, and in order to promote the comparability between the content types, according to the score adjustment parameters corresponding to the content types at the current exposure position, the fusion scores corresponding to the primary screening content belonging to the content types at the current exposure position are adjusted, so as to obtain the adjustment scores corresponding to the primary screening content belonging to the content types at the current exposure position. Therefore, the value fields of the corresponding scores of the primary screening contents of different content types can be unified, and the importance of each primary screening content can be intuitively reflected. And then according to the corresponding adjustment score of each primary screening content at the current exposure position, the exposure content at the current exposure position can be accurately screened out from the primary screening content in real time.
In some embodiments, determining the score adjustment parameter for the content type at the current exposure position based on the location feature, the object attribute feature of the target object, and the behavior statistics feature of the corresponding content type includes: inputting the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types into a second trained object division model; and outputting a score adjustment parameter corresponding to the content type at the current exposure position aiming at the target object through the second trained object division model, wherein the score adjustment parameter is used for adjusting the fusion score corresponding to the primary screening content belonging to the content type at the current exposure position.
Optionally, the computer device obtains a second trained object partition model, and inputs the position feature, the object attribute feature and the behavior statistical feature of the corresponding content type of the current exposure position into the second trained object partition model. And outputting a second fusion parameter and a fraction adjustment parameter corresponding to the content type at the current exposure position aiming at the target object through a second trained object division model.
The computer device inputs the position feature, the object attribute feature and the behavior statistical feature of the corresponding content type of the current exposure position into the second trained object division model, and outputs the fusion parameters corresponding to the preset indexes, the fusion parameters corresponding to the new indexes and the score adjustment parameters of the corresponding content type.
In this embodiment, through the second trained object partition model, not only the second fusion parameter can be automatically adjusted, but also the score adjustment parameter corresponding to each content type can be adaptively determined, so as to promote comparability between each content type.
In some embodiments, as shown in fig. 7, a flow chart of the exposure content for determining the current exposure position in one embodiment is shown. Specifically, the computer equipment inputs the position feature, the object attribute feature of the target object and the behavior statistical feature of the corresponding content type into a second trained object division model; and outputting a to-be-processed adjustment parameter and a to-be-processed second parameter corresponding to the content type at the current exposure position by the target object through a second trained object division model, and respectively processing the to-be-processed adjustment parameter and the to-be-processed second parameter by the computer equipment through an activation function to obtain a score adjustment parameter and a second fusion parameter corresponding to the content type at the current exposure position. For each content type, the computer equipment determines a second screening function corresponding to the content type according to a second fusion parameter corresponding to the content type at the current exposure position, and is beneficial to the second screening function to fuse the overestimated scores of the preset indexes corresponding to the primary screening content belonging to the content type to obtain the fusion score corresponding to the screened content at the current exposure position. Adjusting the fusion score corresponding to the primary screening content belonging to the content type at the current exposure position according to the score adjustment parameter corresponding to the content type at the current exposure position to obtain the adjustment score corresponding to the primary screening content belonging to the content type at the current exposure position; and adjusting the score corresponding to the current exposure position according to the primary screening content.
In this embodiment, according to the score adjustment parameter corresponding to the current exposure position of the content type, the fusion score corresponding to the current exposure position of the primary screening content belonging to the content type is adjusted, so as to obtain the adjustment score corresponding to the current exposure position of the primary screening content belonging to the content type. Therefore, the value fields of the corresponding scores of the primary screening contents of different content types can be unified, and the importance of each primary screening content can be intuitively reflected.
In some embodiments, filtering the exposure content for the current exposure position from the preliminary screening content by the fusion score includes: and respectively screening the exposure content of which the current exposure position is matched with each preset screening condition from the preliminary screening content according to the fusion score.
As described above, there are a plurality of preset screening conditions, based on which the computer device screens out the exposure content whose current exposure position matches each preset screening condition from the fusion score of each preliminary screening content according to the fusion score of each preliminary screening content of the current exposure position.
For each preset screening condition, the computer equipment takes the preset screening condition as a target preset screening condition, and for each current exposure position, the computer equipment screens the exposure content of which the current exposure position is matched with the target preset screening condition from the fusion scores of the primary screening content according to the fusion scores of the primary screening content of the current exposure position.
In this embodiment, exposure contents matching with each preset screening condition at the current exposure position are screened from the primary screening contents according to the fusion scores through a plurality of preset screening conditions, so that the exposure contents matching with each preset screening condition can be flexibly and accurately determined.
In other embodiments, the filtering the exposure content of the current exposure position from the preliminary screening content according to the adjustment score corresponding to the current exposure position of each preliminary screening content includes: and respectively screening the exposure content with the current exposure position matched with each preset screening condition from the preliminary screening content according to the adjustment score.
Optionally, for each preset screening condition, the computer device uses the preset screening condition as a target preset screening condition, and for each current exposure position, the computer device screens the exposure content of which the current exposure position is matched with the target preset screening condition from the fusion scores of the primary screening content according to the adjustment scores of the primary screening content of the current exposure position.
In this embodiment, after the adjustment of each fusion score, comparability between each content type is achieved. In this way, through a plurality of preset screening conditions, the exposure content of which the current exposure position is matched with each preset screening condition is screened out from the primary screening content according to the adjustment score, so that the exposure content matched with each preset screening condition can be flexibly and accurately determined.
Based on this, after screening out the exposure content whose current exposure position matches each preset screening condition, in some embodiments, pushing the content stream to the target object according to the exposure content of each exposure position includes: according to the exposure content of each exposure position corresponding to the same preset screening condition, obtaining a content stream formed by the exposure content corresponding to the same screening condition; determining a target content stream from the content streams corresponding to each preset screening condition; the target content stream is pushed to the target object.
Optionally, the computer device obtains a content stream formed by the exposure content corresponding to the same screening condition according to the exposure content of the exposure positions corresponding to the same preset screening condition. The computer equipment re-predicts the exposure probability and the click rate corresponding to each exposure position in each content stream according to the content stream of each preset screening condition, and determines a target content stream from a plurality of content streams based on the exposure probability and the click rate corresponding to each exposure position in each content stream. The computer device pushes the target content stream to the target object.
Illustratively, for each content stream, the computer device re-predicts the exposure probability and click rate for each exposure pose in the content stream for which it is intended, and calculates the average of the exposure probability and click rate in the content stream for which it is intended. The computer device determines a target content stream based on the exposure probability average and the average of click rates. Alternatively, for each content stream, the computer device calculates an exposure probability difference between a maximum exposure probability and a minimum exposure probability in the content stream for which it is intended, and a click rate difference between a maximum click rate and a minimum click rate. And the computer equipment determines a target content stream according to the exposure probability difference and the click rate difference respectively corresponding to the content streams.
In this embodiment, a plurality of preset screening conditions are preset based on different requirements and service requirements, and based on the preset screening conditions, content streams corresponding to the preset screening conditions are respectively determined. And selecting a target content stream which is more matched with the target object from the plurality of content streams, so that the content pushing effect is ensured.
In some embodiments, as shown in fig. 8, a flowchart of the updating step of the first trained object partition model in one embodiment is shown. The updating of the first trained object partition model comprises:
step 802, a first object partition model updated last time is obtained.
The first object partition model is a deep neural network model (Deep Neural Networks, abbreviated as DNN).
Step 804, obtaining a plurality of groups of disturbance parameters for current update and test objects corresponding to each group of disturbance parameters.
Wherein the perturbation parameters are used to determine the model update.
Optionally, the computer device obtains a plurality of sets of perturbation parameters for the current update and test objects corresponding to each set of perturbation parameters. The number of disturbance parameters in each disturbance group is the same as the number of parameters of the model, and each disturbance parameter is randomly generated by the computer equipment based on normal distribution.
Illustratively, during the offline service, the computer device generates multiple sets of perturbation parameters for the current update based on the normal distribution and caches the multiple sets of perturbation parameters in redis (Remote Dictionary Server, remote dictionary service). Thus, in an online service, the computer device obtains multiple sets of perturbation parameters from redis for the current update. Alternatively, during online service, the computer device generates multiple sets of disturbance parameters for the current update based on the normal distribution, and caches the multiple sets of disturbance parameters in the online model serving. Thus, in the current update process, the computer device obtains multiple sets of disturbance parameters for the current update from the online model serving.
Step 806, for each group of disturbance parameters, performing disturbance processing on the first object partition model updated in the previous time according to the disturbance parameters to obtain a disturbance model, outputting first fusion parameters corresponding to each preset index for the test object according to the object attribute characteristics and the behavior statistical characteristics of the test object through the disturbance model, and after content pushing is performed on the test object based on the first fusion parameters, calculating feedback scores about the disturbance model according to feedback data of the test object.
The feedback data includes feedback values of various feedback indexes, and the feedback indexes can be duration, whether the test object clicks, consumption duration of the test object, and the like.
Optionally, for each group of disturbance parameters, the computer device performs disturbance processing on the first object partition model updated in the previous time according to the disturbance parameters to obtain a corresponding disturbance model. For each group of disturbance parameters, the computer equipment outputs first fusion parameters corresponding to each preset index of the test object according to the object attribute characteristics and the behavior statistical characteristics of the test object corresponding to the disturbance parameters through the aimed disturbance model. For each disturbance model, after content pushing is carried out on the test object corresponding to the disturbance model based on the first fusion parameter corresponding to the disturbance model, the computer equipment calculates the feedback score of the disturbance model according to the feedback data of the test object corresponding to the disturbance model.
Further, the step of counting feedback scores for the disturbance model based on feedback data of the test object corresponding to the disturbance parameter includes:
the computer equipment determines a plurality of feedback indexes, and for each group of disturbance parameters, the computer equipment determines feedback values of feedback indexes of each test object corresponding to the disturbance parameters. For each group of disturbance parameters and each feedback index, the computer equipment takes the average value of the group of disturbance parameters corresponding to the feedback index as a first average value. The computer device determines a feedback value for each feedback indicator for each reference object determined by the first object partitioning model. The reference object is a test object of the first object partition model. For each feedback indicator, the computer device treats the mean value of the feedback indicator, determined by the first object division model, as a second mean value. For each feedback index, the computer equipment determines the difference value of the feedback index according to the first average value and the second average value corresponding to the feedback index, and determines the standard value of the difference value of the feedback index. And the computer equipment determines the feedback score corresponding to the disturbance parameter group through weighted summation according to the standard value of the difference value of each feedback index and the weight of each feedback index.
In some embodiments, determining the feedback value of the feedback indicators for each reference object determined by the first object partitioning model includes: the computer equipment outputs first fusion parameters corresponding to preset indexes of the reference object according to object attribute characteristics and behavior statistical characteristics of the reference object through the first object division model, and feedback data of the reference object are obtained after content pushing is carried out on the test object based on the first fusion parameters. And the computer equipment determines the feedback value of each feedback index of each test object according to the feedback data of the test object.
In some embodiments, for each feedback indicator, the computer device determines a difference value of the feedback indicator according to the first average value and the second average value corresponding to the feedback indicator, including: for each feedback index, the computer device uses the difference between the corresponding first average value and the corresponding second average value of the feedback index as the difference value of the feedback index.
In some embodiments, determining a standard value of the difference in the feedback indicator comprises: for each feedback index, the computer equipment takes the average value of the difference values of the feedback indexes corresponding to the disturbance parameters of each group as a third average value, and calculates the variance value of the difference values of the feedback indexes corresponding to the disturbance parameters of each group. For each feedback index, the computer equipment calculates the difference between the difference value of the feedback index and the third average value corresponding to the feedback index, and takes the ratio of the difference value to the variance value as the standard value of the difference value of the feedback index. The standard value of the difference value of the feedback index is understood to be a value obtained by normalizing the difference value of the feedback index.
For example, taking the feedback index as the duration, there are 2 groups of disturbance parameters, the first group of disturbance parameters corresponds to 100 test objects, and the second group of disturbance parameters has 80 test objects. For the duration, the computer device calculates the average value of the feedback values of the durations of 100 test objects, takes the average value as a first average value, determines the feedback values of the durations of 50 test objects through a first object division model, and obtains a second average value through calculating the average value of the feedback values of the durations of 50 test objects. And the computer equipment calculates the average value of the 180 feedback values according to the feedback values of the duration of 100 test objects and the feedback values of the duration of 80 test objects to obtain a third average value.
It should be noted that, a set of disturbance parameters is generated every preset time period, and for each set of disturbance parameters, the first average value of each feedback index corresponding to the set of disturbance parameters is determined by a plurality of test objects, so that additional noise caused by the variation of a single test object is avoided. Meanwhile, through mean value processing, the feedback score of a long distance (in a long time) is observed, but not the feedback score of a short distance (in a short time), so that the problem that noise exists at the pareto boundary on the optimization non-deterministic problem is solved, and the accuracy and the effectiveness of a subsequent model are ensured.
Step 808, determining a target disturbance model from the disturbance models according to the feedback scores of the disturbance models.
The feedback score can reflect the pushing effect generated by the corresponding disturbance model, and the higher the feedback score is, the better the generated pushing effect is reflected.
Optionally, the computer device takes the disturbance model with the highest feedback score as the target disturbance model.
Step 810, determining the update amount of the model parameters for the current update according to the update amount of the model parameters for the previous update, the disturbance parameters of the target disturbance model and the feedback score.
Optionally, for each model parameter, the computer device determines an amount of change in the model parameter based on the feedback score of the target disturbance model and the disturbance parameter corresponding to the model parameter. For each model parameter, the computer device determines the model parameter update amount for the current update corresponding to the model parameter according to the model parameter update amount and the change amount for the previous update corresponding to the model parameter.
In some embodiments, determining the amount of change in the model parameter based on the feedback score of the target disturbance model and the disturbance parameter corresponding to the model parameter includes: the computer equipment calculates the product of the feedback score of the target disturbance model and the disturbance parameter corresponding to the model parameter, obtains the number of the disturbance parameters corresponding to the target disturbance model, the learning rate and the variance value of the normal distribution when the disturbance parameter is generated, and determines the variation of the model parameter according to the product, the number of the disturbance parameters, the learning rate and the variance value of the normal distribution when the disturbance parameter is generated.
For example, the model parameter i variance can be determined by the following formula:
update amount = learning rate/number of disturbance parameters/variance value x feedback score x disturbance parameters corresponding to model parameters
In some embodiments, determining the model parameter update amount for the current update corresponding to the model parameter according to the model parameter update amount and the change amount for the previous update corresponding to the model parameter includes: the computer equipment determines the attenuation coefficient and the corresponding duty ratio coefficient of the previous update, and superimposes the product of the attenuation coefficient and the variation, the product of the model parameter update quantity used for the previous update and the duty ratio coefficient to obtain the model parameter update quantity used for the current update.
For example, the model parameter update amount v for the current update can be determined by the following formula t ：
Wherein, gamma is the corresponding duty ratio coefficient updated in the previous time and can be set manually according to the service requirement. V (v) t-1 Is the amount of model parameter update for the current update. η is the attenuation coefficient of the light-emitting diode,is the variation of model parameters, wherein +.>Is the gradient of the current update, J (θ) is the feedback score at parameter θ.
It should be noted that, the model parameter updating amount is determined based on the concept of momentum gradient updating, that is, a natural evolution strategy and a momentum gradient optimization method are combined, and in each evolution updating process, the evolution effect with faster convergence and high stability is achieved by carrying out attenuation addition on the previous updating amount. In addition, based on an evolutionary algorithm real-time updating mechanism, nonlinear relations among features are learned, fusion parameters are generated in real time on line, and thousands of people and thousands of faces in a fusion form are achieved. Furthermore, introducing momentum gradient updates can enhance the stability of model evolution. In addition, as described above, the determining of the model parameter updating amount fully considers the coordination relationship among multiple content types, considers different service requirements, and determines the first fusion parameter based on the feedback indexes of the consumption time, the video automatic playing time and the diversity, so as to perform the first-stage fusion, thereby effectively improving the first-stage fusion efficiency and ensuring the accuracy of the model.
Step 812, updating the first object partition model updated in the previous time according to the update amount of the model parameters for the current time to obtain the first object partition model updated in the current time.
Optionally, the computer device updates the first object partition model updated last time according to the update amount of the model parameter for the current update, to obtain the first object partition model updated last time. If the current updated first object division model is used, according to the object attribute characteristics and the behavior statistical characteristics of the current verification object, outputting first fusion parameters corresponding to preset indexes of the verification object, pushing the content of the verification object based on the first fusion parameters, determining feedback data of the verification object, determining benefits generated by the verification objects according to the feedback data of the verification objects by the computer equipment, and if the generated benefits reach the preset benefits, indicating that the current updated first object division model is trained, and taking the trained first object division model as a first trained object division model.
In some embodiments, as depicted in FIG. 9, a schematic of model training of the model is divided for a first trained object. Specifically, a first object partition model updated last time is acquired. And acquiring a plurality of groups of disturbance parameters for current updating and test objects corresponding to each group of disturbance parameters. And for each group of disturbance parameters, carrying out disturbance processing on the first object partition model updated in the previous time according to the disturbance parameters to obtain a disturbance model, outputting first fusion parameters corresponding to each preset index for the test object according to the object attribute characteristics and the behavior statistical characteristics of the test object through the disturbance model, and after carrying out content pushing on the test object based on the first fusion parameters, carrying out statistics on feedback scores of the disturbance model according to feedback data of the test object. And determining a target disturbance model from the disturbance models according to the feedback scores of the disturbance models. For each model parameter, the computer device determines the amount of change in the model parameter based on the feedback score of the target disturbance model and the disturbance parameter corresponding to the model parameter. For each model parameter, the computer equipment acquires the updating quantity (namely, the corresponding baseline parameter) of the model parameter for the previous updating from the storage service redis, and determines the updating quantity of the model parameter for the current updating corresponding to the model parameter according to the updating quantity and the changing quantity of the model parameter for the previous updating corresponding to the model parameter. The computer equipment updates the first object division model updated in the previous time according to the update quantity of the model parameters used for the current update to obtain the first object division model updated in the current time. The computer device acquires a plurality of groups of disturbance parameters for the next update, takes the plurality of groups of disturbance parameters for the next update and the model parameter update amount for the current update as current exploration parameters, and stores the current exploration parameters in a storage service redis. Wherein the current exploration parameter is used to generate the next disturbance model.
In this embodiment, the perturbation model capable of being updated is obtained by performing perturbation processing on the first object division model updated last time. Based on the concept of momentum gradient evolution, nonlinear relations among features are updated in real time, fusion parameters are generated in real time on line, thousands of people and thousands of faces in fusion form are achieved, and accuracy and effectiveness of a first trained object division model which is trained later are guaranteed.
In some embodiments, as shown in fig. 10, which is a flowchart illustrating a step of updating the second trained object partition model in one embodiment, the step of updating the second trained object partition model includes:
step 1002, a second object partition model updated last time is obtained.
Step 1004, obtaining a plurality of groups of disturbance parameters for current update and test objects corresponding to each group of disturbance parameters.
Step 1006, for each group of disturbance parameters, performing disturbance processing on the second object partition model updated in the previous time according to the disturbance parameters to obtain a disturbance model, outputting second fusion parameters corresponding to each preset index for the test object according to the object attribute features, the behavior statistical features and the position features of the test object through the disturbance model, and after content pushing is performed on the test object based on the second fusion parameters, counting feedback scores about the disturbance parameters according to feedback data of the test object.
The feedback data includes feedback values of various feedback indexes, and the feedback indexes can be duration, whether the test object clicks, consumption duration of the test object, and the like.
Optionally, for each group of disturbance parameters, the computer device performs disturbance processing on the second object division model updated in the previous time according to the disturbance parameters to obtain a corresponding disturbance model. For each group of disturbance parameters, the computer equipment outputs second fusion parameters corresponding to each preset index of the test object according to the position features corresponding to the disturbance parameters, the object attribute features and the behavior statistical features of the test object through the corresponding disturbance model. For each disturbance model, after content pushing is carried out on the test object corresponding to the disturbance model based on the second fusion parameter corresponding to the disturbance model, the computer equipment calculates the feedback score of the disturbance model according to the feedback data of the test object corresponding to the disturbance model.
Further, the step of counting feedback scores for the disturbance model based on feedback data of the test object corresponding to the disturbance parameter includes:
the computer equipment determines a plurality of feedback indexes, and for each group of disturbance parameters, the computer equipment determines feedback values of feedback indexes of each test object corresponding to the disturbance parameters. For each group of disturbance parameters and each feedback index, the computer equipment takes the average value of the group of disturbance parameters corresponding to the feedback index as a fourth average value. The computer device determines a feedback value for each feedback indicator for each reference object determined by the second object partitioning model. The reference object is a test object of the second object partition model. For each feedback indicator, the computer device treats the mean value of the feedback indicator determined by the second object division model as a fifth mean value. For each feedback index, the computer equipment determines the difference value of the feedback index according to the fourth average value and the fifth average value corresponding to the feedback index, and determines the standard value of the difference value of the feedback index. And the computer equipment determines the feedback score corresponding to the disturbance parameter group through weighted summation according to the standard value of the difference value of each feedback index and the weight of each feedback index.
In some embodiments, determining the feedback value of the feedback indicator for each reference object determined by the second object partitioning model includes: the computer equipment outputs second fusion parameters corresponding to each preset index aiming at the reference object according to the position features, the object attribute features and the behavior statistical features of the reference object through the second object division model, and feedback data of the reference object are obtained after content pushing is carried out on the test object based on the second fusion parameters. And the computer equipment determines the feedback value of each feedback index of each test object according to the feedback data of the test object.
In some embodiments, for each feedback indicator, the computer device determines a difference value of the feedback indicator according to the fourth mean value and the fifth mean value corresponding to the feedback indicator, including: for each feedback index, the computer device uses the difference between the fourth average value corresponding to the feedback index and the fifth average value corresponding to the feedback index as the difference value of the feedback index.
In some embodiments, determining a standard value of the difference in the feedback indicator comprises: for each feedback index, the computer equipment takes the average value of the difference values of the feedback indexes corresponding to the disturbance parameters of each group as a sixth average value, and calculates the variance value of the difference values of the feedback indexes corresponding to the disturbance parameters of each group. For each feedback index, the computer equipment calculates the difference between the difference value of the feedback index and the sixth mean value corresponding to the feedback index, and takes the ratio of the difference value to the variance value as the standard value of the difference value of the feedback index. The standard value of the difference value of the feedback index is understood to be a value obtained by normalizing the difference value of the feedback index.
Step 1008, determining a target disturbance model from the disturbance models according to the feedback scores of the disturbance models.
In step 1010, the update amount of the model parameter for the current update is determined according to the update amount of the model parameter for the previous update, the disturbance parameter of the target disturbance model and the feedback parameter.
Step 1012, updating the second object partition model updated in the previous time according to the update amount of the model parameters for the current time to obtain the second object partition model updated in the current time.
In some embodiments, a second object partition model of a previous update is obtained. And acquiring a plurality of groups of disturbance parameters for current updating and test objects corresponding to each group of disturbance parameters. And for each group of disturbance parameters, carrying out disturbance processing on the second object partition model updated in the previous time according to the disturbance parameters to obtain a disturbance model, outputting second fusion parameters corresponding to each preset index for the test object according to the object attribute characteristics, the behavior statistical characteristics and the position characteristics of the test object through the disturbance model, and after carrying out content pushing on the test object based on the second fusion parameters, carrying out statistics on feedback scores of the disturbance model according to feedback data of the test object. And determining a target disturbance model from the disturbance models according to the feedback scores of the disturbance models. For each model parameter, the computer device determines the amount of change in the model parameter based on the feedback score of the target disturbance model and the disturbance parameter corresponding to the model parameter. For each model parameter, the computer equipment acquires the updating quantity (namely, the corresponding baseline parameter) of the model parameter for the previous updating from the storage service redis, and determines the updating quantity of the model parameter for the current updating corresponding to the model parameter according to the updating quantity and the changing quantity of the model parameter for the previous updating corresponding to the model parameter. The computer equipment updates the second object division model updated in the previous time according to the update quantity of the model parameters used for the current update to obtain the second object division model updated in the current time. The computer device acquires a plurality of groups of disturbance parameters for the next update, takes the plurality of groups of disturbance parameters for the next update and the model parameter update amount for the current update as current exploration parameters, and stores the current exploration parameters in a storage service redis. Wherein the current exploration parameter is used to generate the next disturbance model.
It should be noted that, the first trained object partition model and the second trained object partition model are obtained by training independently of each other, and the frequency and duration of updating are the same.
In this embodiment, the perturbation model capable of being updated is obtained by performing perturbation processing on the second object division model updated last time. Based on the concept of momentum gradient evolution, nonlinear relations among features are updated in real time, fusion parameters are generated in real time on line, thousands of people and thousands of faces in fusion form are achieved, and accuracy and effectiveness of a second trained object division model which is trained later are guaranteed.
According to the method, the first trained object division model and the second trained object division model are combined to push the content of the target object, so that core indexes such as average residence time, next day retention, average total depth consumption, interaction rate, recommendation diversity and the like are improved in a news pushing scene, in addition, in the process of determining the first fusion parameters through the first trained object division model, multiple preset indexes are fused, compatibility of various business requirements is achieved, individuation is achieved, and user experience is effectively improved.
Specifically, in the whole end of the news scene, based on the content pushing method provided by the application, the average residence time of the news client is increased by 3.26%, the average total depth consumption of the news client is increased by 3.13%, the average depth consumption of the news client is increased by 1.74%, the average depth consumption of the news client is increased by PV (page view ) is increased by 3.67%, the next day retention rate is increased by 0.56%, the interactive permeability is increased by 1.89%, the comment permeability is increased by 3.83%, the sharing permeability is increased by 1.41%, the interactive uv (website independent visitor) is increased by 1.33% (the average residence time of the new client is increased by 1.99%, the average total depth consumption of the news client is increased by 2.05%, the average depth consumption of the news client is increased by 7.74%, the average depth consumption of the news is increased by VV is increased by 0.54%, and the next day retention rate is increased by 1.25%). In a headline channel of a news scene, the average time length of stay of the channel is increased by 6.01%, the average total depth consumption is increased by 5.53%, the average depth consumption PV is increased by 1.98%, the average depth consumption VV is increased by 8.52%, the interaction permeability is increased by 2.45%, the comment permeability is increased by 4.59%, the sharing permeability is increased by 1.97% (the average time length of stay of a new user channel is increased by 1.87%, the average total depth consumption is increased by 3.64%, and the average depth consumption PV is increased by 8.97%).
The application also provides an application scene, which applies the content pushing method. Specifically, the application of the content pushing method in the application scenario is as follows: in a news pushing scene, after a user logs in a news client, in order to ensure that the news client can push a content stream matched with the user, the content stream pushing method provided by the application can be adopted for content stream pushing. Specifically, candidate contents for pushing the content stream of the target object are obtained; each candidate content has a corresponding content type; for each content type, screening out primary screening content from candidate content belonging to the content type based on the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content type; for a plurality of exposure positions for pushing the content stream, starting from the first exposure position in the plurality of exposure positions, determining the position characteristics of the current exposure position; the position features comprise preamble content statistical features of the current exposure position; based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types, the exposure content of the current exposure position is screened from the preliminary screening content; and pushing a content stream to the target object according to the exposure content of each exposure position.
Of course, the content pushing method provided by the application is not limited to the above, and can be applied to other application scenes, for example, in advertisement pushing scenes, film and television pushing scenes and the like, and content pushing can be realized by the content pushing method provided by the application.
The above application scenario is only illustrative, and it can be understood that the application of the content pushing method provided by the embodiments of the present application is not limited to the above scenario.
In a specific embodiment, the present embodiment provides a content pushing method, which is performed by a computer device. The specific description is as follows:
the first step: acquiring candidate contents for pushing the content stream of the target object; each candidate content has a corresponding content type. Determining a preset index for evaluating the candidate content; and for any candidate content, estimating the estimated score of the preset index corresponding to the candidate content for the target object according to the object attribute characteristics of the target object and the content characteristics of the candidate content by a trained estimated model corresponding to the preset index.
And a second step of: for any content type, acquiring object attribute characteristics of a target object and behavior statistical characteristics of the corresponding content type; inputting the object attribute characteristics and the behavior statistical characteristics of the corresponding content types into a first trained object division model; outputting first fusion parameters corresponding to the content types aiming at the target object through a first trained object division model, wherein the first fusion parameters comprise fusion parameters respectively corresponding to preset indexes, and the first fusion parameters are used for fusing estimated scores of the preset indexes corresponding to candidate contents estimated aiming at the target object. And according to the first fusion parameters corresponding to the content types, fusing the estimated scores of the candidate content belonging to the content types corresponding to the preset indexes to obtain the preliminary screening score of the candidate content. And screening the primary screening content from the candidate content belonging to the content type according to the primary screening number.
And a third step of: for a plurality of exposure positions for pushing the content stream, starting from the first exposure position in the plurality of exposure positions, determining the position characteristics of the current exposure position; the location features include preamble statistics of the current exposure location.
Fourth step: the above content characteristics of the current exposure position are determined. Determining a preset index for evaluating the primary screening content; and for any primary screening content, estimating the re-estimation number of the preset index corresponding to the primary screening content according to the content characteristics of the primary screening content and the content characteristics of the primary screening content through a trained re-estimation model corresponding to the preset index.
Fifth step: inputting the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types into a second trained object division model; and outputting a second fusion parameter and a fraction adjustment parameter corresponding to the content type at the current exposure position aiming at the target object through a second trained object division model, wherein the second fusion parameter comprises fusion parameters respectively corresponding to preset indexes, and the second fusion parameter is used for fusing the re-estimation fraction of the preset indexes corresponding to the preliminary screening content re-estimated aiming at the target object. The score adjustment parameter is used for adjusting the fusion score corresponding to the current exposure position of the primary screening content belonging to the content type.
Sixth step: determining a second fusion parameter corresponding to the content type at the current exposure position based on the position characteristic, the object attribute characteristic of the target object and the behavior statistical characteristic of the corresponding content type; according to a second fusion parameter corresponding to the content type at the current exposure position, fusing the overestimated scores of the pre-screened content belonging to the content type corresponding to the preset indexes to obtain a fusion score corresponding to the pre-screened content belonging to the content type at the current exposure position; adjusting the fusion score corresponding to the primary screening content belonging to the content type at the current exposure position according to the score adjustment parameter corresponding to the content type at the current exposure position to obtain the adjustment score corresponding to the primary screening content belonging to the content type at the current exposure position; and respectively screening the exposure content of which the current exposure position is matched with each preset screening condition from the preliminary screening content according to the adjustment score of each preliminary screening content corresponding to the current exposure position. According to the exposure content of each exposure position corresponding to the same preset screening condition, obtaining a content stream formed by the exposure content corresponding to the same screening condition; determining a target content stream from the content streams corresponding to each preset screening condition; the target content stream is pushed to the target object.
In the embodiment, candidate contents for pushing the content stream of the target object are obtained; each candidate content has a corresponding content type; for each content type, based on the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content type, the primary screening content is screened from the candidate content belonging to the content type, namely, the candidate content of the same content type is accurately and primarily screened by utilizing the object attribute characteristics and the behavior statistical characteristics which can highlight the individuation of the crowd of the target object. For a plurality of exposure positions for pushing the content stream, starting from the first exposure position in the plurality of exposure positions, determining the position characteristics of the current exposure position; the position features comprise preamble content statistical features of the current exposure position, namely, the position features of the current exposure position are adaptively adjusted according to related content capable of reflecting the preamble exposure position; and screening the exposure content of the current exposure position from the preliminary screening content based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content types. Therefore, the exposure content of each exposure position matched with the interest requirement of the target object is screened out for the second time by combining the object attribute characteristics and the behavior statistical characteristics with crowd individuation and the current exposure position characteristics with self-adaption. According to the exposure content of each exposure position, content streams with higher suitability can be pushed to a target object, so that the pushing effect is improved.
It should be understood that, although the steps in the flowcharts related to the embodiments described above are sequentially shown as indicated by arrows, these steps are not necessarily sequentially performed in the order indicated by the arrows. The steps are not strictly limited to the order of execution unless explicitly recited herein, and the steps may be executed in other orders. Moreover, at least some of the steps in the flowcharts described in the above embodiments may include a plurality of steps or a plurality of stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of the steps or stages is not necessarily performed sequentially, but may be performed alternately or alternately with at least some of the other steps or stages.
Based on the same inventive concept, the embodiment of the application also provides a content pushing device for realizing the content pushing method. The implementation of the solution provided by the device is similar to the implementation described in the above method, so the specific limitation in one or more embodiments of the content pushing device provided below may refer to the limitation of the content pushing method hereinabove, and will not be repeated herein.
In one embodiment, as shown in fig. 11, there is provided a content pushing apparatus including: an acquisition module 1102, a first screening module 1104, a determination module 1106, a second screening module 1108, and a pushing module 1110, wherein:
an obtaining module 1102, configured to obtain candidate content that performs content stream pushing on a target object; each candidate content has a corresponding content type.
The first filtering module 1104 is configured to, for each content type, filter the first screened content from the candidate content belonging to the content type based on the object attribute feature of the target object and the behavior statistical feature of the corresponding content type.
A determining module 1106, configured to determine, for a plurality of exposure positions for content stream pushing, a position feature of a current exposure position from a first exposure position of the plurality of exposure positions; the location features include preamble statistics of the current exposure location.
The second filtering module 1108 is configured to filter the exposure content of the current exposure position from the preliminary filtering content based on the position feature, the object attribute feature of the target object, and the behavior statistical feature of the corresponding content type.
And a pushing module 1110, configured to push a content stream to the target object according to the exposure content of each exposure position.
In one embodiment, a first screening module is configured to obtain, for a target object, a pre-estimated score of each candidate content corresponding to each preset index; and for each content type, respectively fusing the estimated scores of the preset indexes corresponding to the candidate contents belonging to the content type based on the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content type, and screening the primary screening content from the candidate contents belonging to the content type according to the fusion result.
In one embodiment, a first screening module is configured to determine a preset index for evaluating candidate content; and for any candidate content, estimating the estimated score of the preset index corresponding to the candidate content for the target object according to the object attribute characteristics of the target object and the content characteristics of the candidate content by a trained estimated model corresponding to the preset index.
In one embodiment, a first screening module is configured to obtain an object attribute feature of a target object and a behavior statistical feature of a corresponding content type; determining a first fusion parameter corresponding to the content type based on the object attribute characteristics and the behavior statistical characteristics; according to the first fusion parameters corresponding to the content types, fusing the estimated scores of the candidate content belonging to the content types corresponding to the preset indexes to obtain the preliminary screening score of the candidate content; and screening the primary screening content from the candidate content belonging to the content type according to the primary screening number.
In one embodiment, a first filtering module is configured to input the object attribute feature and the behavior statistical feature of the corresponding content type into a first trained object partition model; outputting first fusion parameters corresponding to the content types aiming at the target object through a first trained object division model, wherein the first fusion parameters comprise fusion parameters respectively corresponding to preset indexes, and the first fusion parameters are used for fusing estimated scores of the preset indexes corresponding to candidate contents estimated aiming at the target object.
In one embodiment, a second screening module for determining the above content characteristics of the current exposure position; based on the content characteristics and the content characteristics of the primary screening content, determining the re-estimation score of the primary screening content corresponding to each preset index; and for each content type, based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics of the corresponding content type, the overestimated scores of the preset indexes corresponding to the preliminary screening content belonging to the content type are fused, and then the exposure content of the current exposure position is screened from the preliminary screening content according to the fusion result.
In one embodiment, the second screening module is configured to determine a preset index for evaluating the primary screening content; and for any primary screening content, estimating the re-estimation number of the preset index corresponding to the primary screening content according to the content characteristics of the primary screening content and the content characteristics of the primary screening content through a trained re-estimation model corresponding to the preset index.
In one embodiment, the second filtering module is configured to determine a second fusion parameter corresponding to the content type at the current exposure position based on the location feature, the object attribute feature of the target object, and the behavior statistical feature of the corresponding content type; according to a second fusion parameter corresponding to the content type at the current exposure position, fusing the overestimated scores of the pre-screened content belonging to the content type corresponding to the preset indexes to obtain a fusion score corresponding to the pre-screened content belonging to the content type at the current exposure position; and screening the exposure content of the current exposure position from the preliminary screening content according to the fusion score.
In one embodiment, the second filtering module is configured to input the location feature, the object attribute feature of the target object, and the behavior statistical feature of the corresponding content type into a second trained object partition model; and outputting second fusion parameters corresponding to the content type at the current exposure position aiming at the target object through a second trained object division model, wherein the second fusion parameters comprise fusion parameters respectively corresponding to preset indexes, and the second fusion parameters are used for fusing the re-estimation scores of the preset indexes corresponding to the preliminary screening content re-estimated aiming at the target object.
In one embodiment, the second filtering module is further configured to determine a score adjustment parameter corresponding to the content type at the current exposure position based on the location feature, the object attribute feature of the target object, and the behavior statistical feature of the corresponding content type; the second screening module is used for adjusting the fusion score corresponding to the primary screening content belonging to the content type at the current exposure position according to the score adjustment parameter corresponding to the content type at the current exposure position, so as to obtain the adjustment score corresponding to the primary screening content belonging to the content type at the current exposure position; and screening the exposure content of the current exposure position from the preliminary screening content according to the adjustment score of each preliminary screening content corresponding to the current exposure position.
In one embodiment, the second filtering module is configured to input the location feature, the object attribute feature of the target object, and the behavior statistical feature of the corresponding content type into a second trained object partition model; and outputting a score adjustment parameter corresponding to the content type at the current exposure position aiming at the target object through the second trained object division model, wherein the score adjustment parameter is used for adjusting the fusion score corresponding to the primary screening content belonging to the content type at the current exposure position.
In one embodiment, the second screening module is configured to screen, according to the fusion score, exposure contents whose current exposure positions match with preset screening conditions from the primary screening contents respectively;
the pushing module is used for obtaining a content stream formed by the exposure contents corresponding to the same screening condition according to the exposure contents of which the exposure positions correspond to the same preset screening condition; determining a target content stream from the content streams corresponding to each preset screening condition; the target content stream is pushed to the target object.
In one embodiment, the apparatus further includes a first updating module, configured to obtain a first object partition model updated last time; acquiring a plurality of groups of disturbance parameters used for current updating and test objects corresponding to each group of disturbance parameters; for each group of disturbance parameters, carrying out disturbance processing on a first object partition model updated in the previous time according to the disturbance parameters to obtain a disturbance model, outputting first fusion parameters corresponding to each preset index for the test object according to object attribute characteristics and behavior statistical characteristics of the test object through the disturbance model, and after carrying out content pushing on the test object based on the first fusion parameters, carrying out statistics on feedback scores of the disturbance model according to feedback data of the test object; determining a target disturbance model from the disturbance models according to the feedback scores of the disturbance models; determining a model parameter updating amount for current updating according to the model parameter updating amount for previous updating, the disturbance parameters of the target disturbance model and the feedback score; and updating the first object partition model updated in the previous time according to the update quantity of the model parameters used for the current update to obtain the first object partition model updated in the current time.
In one embodiment, the apparatus further includes a second updating module, configured to obtain a second object partition model updated last time; acquiring a plurality of groups of disturbance parameters used for current updating and test objects corresponding to each group of disturbance parameters; for each group of disturbance parameters, carrying out disturbance processing on a second object partition model updated in the previous time according to the disturbance parameters to obtain a disturbance model, outputting second fusion parameters corresponding to each preset index for the test object according to object attribute characteristics, behavior statistical characteristics and position characteristics of the test object through the disturbance model, and after carrying out content pushing on the test object based on the second fusion parameters, carrying out statistics on feedback scores of the disturbance parameters according to feedback data of the test object; determining a target disturbance model from the disturbance models according to the feedback scores of the disturbance models; determining a model parameter updating amount for current updating according to the model parameter updating amount for previous updating, the disturbance parameters and the feedback parameters of the target disturbance model; and updating the second object partition model updated in the previous time according to the updating quantity of the model parameters used for the current time to obtain the second object partition model updated in the current time.
The respective modules in the content pushing apparatus described above may be implemented in whole or in part by software, hardware, and a combination thereof. The above modules may be embedded in hardware or may be independent of a processor in the computer device, or may be stored in software in a memory in the computer device, so that the processor may call and execute operations corresponding to the above modules.
In one embodiment, a computer device is provided, which may be a server, and the internal structure of which may be as shown in fig. 12. The computer device includes a processor, a memory, an Input/Output interface (I/O) and a communication interface. The processor, the memory and the input/output interface are connected through a system bus, and the communication interface is connected to the system bus through the input/output interface. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input/output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to implement a content pushing method.
It will be appreciated by those skilled in the art that the structure shown in FIG. 12 is merely a block diagram of some of the structures associated with the present inventive arrangements and is not limiting of the computer device to which the present inventive arrangements may be applied, and that a particular computer device may include more or fewer components than shown, or may combine some of the components, or have a different arrangement of components.
In an embodiment, there is also provided a computer device comprising a memory and a processor, the memory having stored therein a computer program, the processor implementing the steps of the method embodiments described above when the computer program is executed.
In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored which, when executed by a processor, carries out the steps of the method embodiments described above.
In an embodiment, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the steps of the method embodiments described above.
It should be noted that, the user information (including but not limited to user equipment information, user personal information, etc.) and the data (including but not limited to data for analysis, stored data, presented data, etc.) related to the present application are information and data authorized by the user or sufficiently authorized by each party, and the collection, use and processing of the related data need to comply with the related laws and regulations and standards of the related country and region.
Those skilled in the art will appreciate that implementing all or part of the above described methods may be accomplished by way of a computer program stored on a non-transitory computer readable storage medium, which when executed, may comprise the steps of the embodiments of the methods described above. Any reference to memory, database, or other medium used in embodiments provided herein may include at least one of non-volatile and volatile memory. The nonvolatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, high density embedded nonvolatile Memory, resistive random access Memory (ReRAM), magnetic random access Memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric Memory (Ferroelectric Random Access Memory, FRAM), phase change Memory (Phase Change Memory, PCM), graphene Memory, and the like. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, and the like. By way of illustration, and not limitation, RAM can be in the form of a variety of forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), and the like. The databases referred to in the embodiments provided herein may include at least one of a relational database and a non-relational database. The non-relational database may include, but is not limited to, a blockchain-based distributed database, and the like. The processor referred to in the embodiments provided in the present application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, or the like, but is not limited thereto.
The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.
The foregoing examples illustrate only a few embodiments of the application and are described in detail herein without thereby limiting the scope of the application. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the application, which are all within the scope of the application. Accordingly, the scope of the application should be assessed as that of the appended claims.
Claims (18)
1. A content pushing method, the method comprising:
acquiring candidate contents for pushing the content stream of the target object; each candidate content has a corresponding content type;
for each content type, screening out primary screening content from candidate content belonging to the content type based on the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content type;
For a plurality of exposure positions for content stream pushing, starting from the first exposure position in the plurality of exposure positions, determining the position characteristics of the current exposure position; the position features comprise preamble content statistical features of the current exposure position;
based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content types, the exposure content of the current exposure position is screened out from the preliminary screening content;
and pushing a content stream to the target object according to the exposure content of each exposure position.
2. The method of claim 1, wherein for each of the content types, the screening the primary content from candidate content belonging to the content type based on the object attribute feature of the target object and the behavior statistics corresponding to the content type, respectively, comprises:
aiming at a target object, obtaining the estimated score of each candidate content corresponding to each preset index;
and for each content type, respectively based on the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content type, merging the estimated scores of the corresponding preset indexes of the candidate content belonging to the content type, and screening the primary screening content from the candidate content belonging to the content type according to the merging result.
3. The method according to claim 2, wherein the obtaining, for the target object, the estimated score of each candidate content corresponding to each preset index includes:
determining a preset index for evaluating the candidate content;
and for any candidate content, estimating the estimated score of the candidate content corresponding to the preset index for the target object according to the object attribute characteristics of the target object and the content characteristics of the candidate content by a trained estimated model corresponding to the preset index.
4. The method according to claim 2, wherein the filtering the preliminary screening content from the candidate content belonging to the content type according to the fusion result after fusing the pre-estimated scores of the corresponding preset indexes of the candidate content belonging to the content type based on the object attribute feature of the target object and the behavior statistical feature corresponding to the content type includes:
acquiring object attribute characteristics of the target object and behavior statistical characteristics corresponding to the content type;
determining a first fusion parameter corresponding to the content type based on the object attribute features and the behavior statistical features;
according to the first fusion parameters corresponding to the content types, fusing the estimated scores of the preset indexes corresponding to the candidate content belonging to the content types to obtain the primary screening score of the candidate content;
And screening the primary screening content from the candidate content belonging to the content type according to the primary screening number.
5. The method of claim 4, wherein the determining a first fusion parameter corresponding to the content type based on the object attribute feature and the behavior statistics feature comprises:
inputting the object attribute characteristics and the behavior statistical characteristics corresponding to the content types into a first trained object division model;
outputting a first fusion parameter corresponding to the content type aiming at the target object through the first trained object division model, wherein the first fusion parameter comprises fusion parameters respectively corresponding to preset indexes, and the first fusion parameter is used for fusing the estimated score of each preset index corresponding to the candidate content estimated aiming at the target object.
6. The method of claim 1, wherein the screening the exposure content of the current exposure position from the preliminary screening content based on the location feature, the object attribute feature of the target object, and the behavior statistics feature corresponding to the content type comprises:
determining the above content features of the current exposure position;
Determining a re-estimation score of the primary screening content corresponding to each preset index based on the content characteristics of the primary screening content and the content characteristics of the above content;
and for each content type, based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content type, the overestimation numbers of the pre-screening content belonging to the content type corresponding to each preset index are fused, and then the exposure content of the current exposure position is screened from the pre-screening content according to the fusion result.
7. The method of claim 6, wherein determining a re-estimation score of the preliminary screening content for each of the predetermined indicators based on the above content characteristics and the content characteristics of the preliminary screening content comprises:
determining a preset index for evaluating the primary screening content;
and for any primary screening content, estimating the re-estimation number of the primary screening content corresponding to the preset index according to the content characteristics of the primary screening content and the content characteristics of the above content through a trained re-estimation model corresponding to the preset index, and aiming at the target object.
8. The method according to claim 6, wherein the step of merging the overestimated contents belonging to the content type corresponding to the overestimated contents of each preset index based on the position feature, the object attribute feature of the target object, and the behavior statistical feature corresponding to the content type, and then screening the exposed contents of the current exposure position from the undersifted contents according to the merging result comprises:
Determining a second fusion parameter corresponding to the content type at the current exposure position based on the position characteristic, the object attribute characteristic of the target object and the behavior statistical characteristic corresponding to the content type;
according to a second fusion parameter corresponding to the content type at the current exposure position, fusing the overestimated scores of the pre-screening content belonging to the content type corresponding to each preset index to obtain a fusion score corresponding to the pre-screening content belonging to the content type at the current exposure position;
and screening the exposure content of the current exposure position from the preliminary screening content according to the fusion score.
9. The method of claim 8, wherein the determining a second fusion parameter for the content type at the current exposure position based on the location feature, the object attribute feature of the target object, and the behavior statistics feature for the content type comprises:
inputting the position feature, the object attribute feature of the target object and the behavior statistical feature corresponding to the content type into a second trained object division model;
outputting, by the second trained object partition model, a second fusion parameter corresponding to the content type at the current exposure position for the target object, where the second fusion parameter includes fusion parameters corresponding to preset indexes respectively, and the second fusion parameter is used to fuse a re-estimation score of each preset index corresponding to the preliminary screening content re-estimated for the target object.
10. The method of claim 8, wherein the method further comprises:
determining a score adjustment parameter corresponding to the content type at the current exposure position based on the position characteristic, the object attribute characteristic of the target object and the behavior statistical characteristic corresponding to the content type;
and screening the exposure content of the current exposure position from the preliminary screening content according to the fusion score, wherein the method comprises the following steps:
adjusting the fusion score of the primary screening content belonging to the content type at the current exposure position according to the score adjustment parameter corresponding to the content type at the current exposure position, so as to obtain an adjustment score corresponding to the primary screening content belonging to the content type at the current exposure position;
and screening the exposure content of the current exposure position from the preliminary screening content according to the adjustment score of each preliminary screening content corresponding to the current exposure position.
11. The method of claim 10, wherein the determining the score adjustment parameter for the content type at the current exposure position based on the location feature, the object attribute feature of the target object, and the behavior statistics feature for the content type comprises:
Inputting the position feature, the object attribute feature of the target object and the behavior statistical feature corresponding to the content type into a second trained object division model;
and outputting a score adjustment parameter corresponding to the content type at the current exposure position aiming at the target object through the second trained object division model, wherein the score adjustment parameter is used for adjusting the fusion score corresponding to the primary screening content belonging to the content type at the current exposure position.
12. The method of claim 8, wherein said screening exposure content for a current exposure position from said preliminary screening content by said fusion score comprises:
according to the fusion score, the exposure content of which the current exposure position is matched with each preset screening condition is screened out from the primary screening content;
the pushing a content stream to the target object according to the exposure content of each exposure position comprises the following steps:
according to the exposure content of each exposure position corresponding to the same preset screening condition, obtaining a content stream formed by the exposure content corresponding to the same screening condition;
determining a target content stream from the content streams corresponding to the preset screening conditions;
And pushing the target content stream to the target object.
13. The method of claim 5, wherein the updating of the first trained object partition model comprises:
acquiring a first object division model updated in the previous time;
acquiring a plurality of groups of disturbance parameters used for current updating and test objects corresponding to each group of disturbance parameters;
for each group of disturbance parameters, carrying out disturbance processing on the first object division model updated in the previous time according to the disturbance parameters to obtain a disturbance model, outputting first fusion parameters corresponding to each preset index for the test object according to object attribute characteristics and behavior statistical characteristics of the test object through the disturbance model, and carrying out content pushing on the test object based on the first fusion parameters, and then carrying out statistics on feedback scores of the disturbance model according to feedback data of the test object;
determining a target disturbance model from each disturbance model according to the feedback score of each disturbance model;
determining a model parameter updating amount for updating at present according to the model parameter updating amount for updating at last time, the disturbance parameters of the target disturbance model and the feedback score;
And updating the first object division model updated in the previous time according to the update quantity of the model parameters used for the current time to obtain the first object division model updated in the current time.
14. The method of claim 9, wherein the step of updating the second trained object partition model comprises:
acquiring a second object division model updated in the previous time;
acquiring a plurality of groups of disturbance parameters used for current updating and test objects corresponding to each group of disturbance parameters;
for each group of disturbance parameters, carrying out disturbance processing on the second object division model updated in the previous time according to the disturbance parameters to obtain a disturbance model, outputting second fusion parameters corresponding to each preset index of the test object according to object attribute characteristics, behavior statistical characteristics and position characteristics of the test object through the disturbance model, and carrying out content pushing on the test object based on the second fusion parameters, and then carrying out statistics on feedback scores of the disturbance parameters according to feedback data of the test object;
determining a target disturbance model from each disturbance model according to the feedback score of each disturbance model;
determining a model parameter updating amount for updating at present according to the model parameter updating amount for updating at last time, the disturbance parameter and the feedback parameter of the target disturbance model;
And updating the second object division model updated in the previous time according to the update quantity of the model parameters used for the current time to obtain the second object division model updated in the current time.
15. A content pushing apparatus, the apparatus comprising:
the acquisition module is used for acquiring candidate contents for pushing the content stream of the target object; each candidate content has a corresponding content type;
the first screening module is used for screening the primary screening content from candidate content belonging to the content type based on the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content type;
a determining module, configured to determine, for a plurality of exposure positions for content stream pushing, a position feature of a current exposure position from a first exposure position of the plurality of exposure positions; the position features comprise preamble content statistical features of the current exposure position;
the second screening module is used for screening exposure content of the current exposure position from the preliminary screening content based on the position characteristics, the object attribute characteristics of the target object and the behavior statistical characteristics corresponding to the content type;
And the pushing module is used for pushing the content stream to the target object according to the exposure content of each exposure position.
16. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any one of claims 1 to 14 when the computer program is executed.
17. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any of claims 1 to 14.
18. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any one of claims 1 to 14.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310100967.6A CN116980471A (en) | 2023-01-19 | 2023-01-19 | Content pushing method, device, apparatus, storage medium and computer program product |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310100967.6A CN116980471A (en) | 2023-01-19 | 2023-01-19 | Content pushing method, device, apparatus, storage medium and computer program product |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116980471A true CN116980471A (en) | 2023-10-31 |
Family
ID=88477283
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310100967.6A Pending CN116980471A (en) | 2023-01-19 | 2023-01-19 | Content pushing method, device, apparatus, storage medium and computer program product |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116980471A (en) |
-
2023
- 2023-01-19 CN CN202310100967.6A patent/CN116980471A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2020207196A1 (en) | Method and apparatus for generating user tag, storage medium and computer device | |
| US9495645B2 (en) | Method and system of iteratively autotuning prediction parameters in a media content recommender | |
| CN112052387B (en) | Content recommendation method, device and computer readable storage medium | |
| CN111127142A (en) | Article recommendation method based on generalized neural attention | |
| CN108230007A (en) | A kind of recognition methods of user view, device, electronic equipment and storage medium | |
| CN113536097B (en) | Recommendation method and device based on automatic feature grouping | |
| CN114265979A (en) | Method for determining fusion parameters, information recommendation method and model training method | |
| CN112749330B (en) | Information pushing method, device, computer equipment and storage medium | |
| CN113761359B (en) | Data packet recommendation method, device, electronic equipment and storage medium | |
| CN110781377B (en) | Article recommendation method and device | |
| US20230069999A1 (en) | Method and apparatus for updating recommendation model, computer device and storage medium | |
| CN111611488A (en) | Information recommendation method and device based on artificial intelligence and electronic equipment | |
| US20240193402A1 (en) | Method and apparatus for determining representation information, device, and storage medium | |
| CN112817563B (en) | Target attribute configuration information determining method, computer device, and storage medium | |
| CN114417058A (en) | Video material screening method and device, computer equipment and storage medium | |
| Sun | [Retracted] Variational Fuzzy Neural Network Algorithm for Music Intelligence Marketing Strategy Optimization | |
| CN113836406A (en) | Information flow recommendation method and device | |
| US20230334314A1 (en) | Content recommendation method and apparatus, device, storage medium, and program product | |
| CN113535939A (en) | Text processing method and device, electronic equipment and computer readable storage medium | |
| CN116541592A (en) | Vector generation method, information recommendation method, device, equipment and medium | |
| CN116980471A (en) | Content pushing method, device, apparatus, storage medium and computer program product | |
| CN115809339A (en) | Cross-domain recommendation method, system, device and storage medium | |
| CN114510627A (en) | Object pushing method and device, electronic equipment and storage medium | |
| CN112785328A (en) | Content pushing method and device and computer storage medium | |
| CN112000888B (en) | Information pushing method, device, server and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication |