CN117668351A - Recommendation method, training method and device of model, electronic equipment and storage medium - Google Patents
Recommendation method, training method and device of model, electronic equipment and storage medium Download PDFInfo
- Publication number
- CN117668351A CN117668351A CN202311334820.XA CN202311334820A CN117668351A CN 117668351 A CN117668351 A CN 117668351A CN 202311334820 A CN202311334820 A CN 202311334820A CN 117668351 A CN117668351 A CN 117668351A
- Authority
- CN
- China
- Prior art keywords
- sample
- resource
- interest
- interaction
- target
- 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 114
- 238000012549 training Methods 0.000 title claims abstract description 46
- 230000003993 interaction Effects 0.000 claims abstract description 521
- 230000004913 activation Effects 0.000 claims abstract description 198
- 238000001994 activation Methods 0.000 claims description 196
- 238000013507 mapping Methods 0.000 claims description 58
- 230000002452 interceptive effect Effects 0.000 claims description 41
- 238000013136 deep learning model Methods 0.000 claims description 37
- 238000012512 characterization method Methods 0.000 claims description 24
- 238000004590 computer program Methods 0.000 claims description 14
- 230000003213 activating effect Effects 0.000 claims description 8
- 238000012163 sequencing technique Methods 0.000 claims description 7
- 238000013473 artificial intelligence Methods 0.000 abstract description 3
- 238000013135 deep learning Methods 0.000 abstract description 2
- 238000012545 processing Methods 0.000 description 36
- 239000013598 vector Substances 0.000 description 25
- 238000010586 diagram Methods 0.000 description 19
- 238000013528 artificial neural network Methods 0.000 description 17
- 230000008569 process Effects 0.000 description 12
- 238000004891 communication Methods 0.000 description 11
- 230000006399 behavior Effects 0.000 description 10
- 238000004364 calculation method Methods 0.000 description 9
- 230000004044 response Effects 0.000 description 9
- 230000006870 function Effects 0.000 description 7
- 230000000694 effects Effects 0.000 description 5
- 230000002829 reductive effect Effects 0.000 description 4
- 238000004422 calculation algorithm Methods 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 230000003287 optical effect Effects 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000005265 energy consumption Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000000670 limiting effect Effects 0.000 description 2
- 238000005295 random walk Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000012216 screening Methods 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 230000002411 adverse Effects 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 238000013475 authorization Methods 0.000 description 1
- 230000001413 cellular effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 239000004973 liquid crystal related substance Substances 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 239000013307 optical fiber Substances 0.000 description 1
- 230000036961 partial effect Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 230000001953 sensory effect Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The disclosure provides a recommendation method, a training device, training equipment and a storage medium of a model, relates to the technical field of artificial intelligence, and particularly relates to the technical field of deep learning and the technical field of data statistics. The specific implementation scheme is as follows: determining a historical interaction resource sequence of interest activation of the target object; determining object characteristics of the target object about the target interest based on the historical interaction resource sequence of interest activation and the object attribute information of the target object; determining a plurality of candidate resource features relating to the target interest; a resource to be recommended is determined from the candidate resources based on the candidate resource characteristics and the object characteristics.
Description
Technical Field
The disclosure relates to the technical field of artificial intelligence, in particular to the technical field of deep learning and the technical field of data statistics, and specifically relates to a recommendation method, a training method and device of a deep learning model, electronic equipment, a storage medium and a program product.
Background
With the rapid development of information technology and network technology, information expansion and redundancy bring confusion to information selection for social activities and entertainment activities of people. Finding the required information from vast and vast resources presents a significant challenge. The personalized information service of the Internet can provide different personalized information service strategies for different users. Automated information recommendation is performed based on different characteristics of the user and requirements. However, in the recommendation process, satisfaction of the recommendation result meeting the personalized requirements of the user needs to be improved.
Disclosure of Invention
The disclosure provides a recommendation method, a training method and device of a deep learning model, electronic equipment, a storage medium and a program product.
According to an aspect of the present disclosure, there is provided a recommendation method including:
determining an interest activated historical interaction resource sequence of a target object, wherein the interest activated historical interaction resource sequence comprises a plurality of historical interaction resources interacted by the target object in a historical period, and the plurality of historical interaction resources belong to the target interest resource;
determining object characteristics of the target object with respect to the target interest based on the historical interaction resource sequence of interest activation and the object attribute information of the target object;
determining a plurality of candidate resource features related to the target interest, wherein the plurality of candidate resource features are in one-to-one correspondence with a plurality of candidate resources, the candidate resource features are used for characterizing resource attribute features of the candidate resources corresponding to the candidate resource features, and the plurality of candidate resources belong to the resource of the target interest; and
and determining resources to be recommended from the plurality of candidate resources based on the plurality of candidate resource characteristics and the object characteristics.
According to another aspect of the present disclosure, there is provided a training method of a deep learning model, including:
determining an interest activated sample historical interaction resource sequence of a sample target object, wherein the interest activated sample historical interaction resource sequence comprises a plurality of sample historical interaction resources interacted by the sample target object in a sample historical period, and the plurality of sample historical interaction resources belong to sample target interest resources;
determining sample object characteristics of the sample target object with respect to the sample target interest based on the sample history interaction resource sequence activated by the interest and sample object attribute information of the sample target object;
determining a target sample historical interaction resource from the sample historical interaction resource sequence of interest activation of the sample target object;
obtaining sample resource characteristics based on sample resource attribute information of the target sample history interaction resource; and
based on the sample object characteristics and the sample resource characteristics, parameters of the deep learning model are adjusted, and training is completed.
According to another aspect of the present disclosure, there is provided a recommendation device including:
the resource sequence determining module is used for determining a historical interaction resource sequence of interest activation of the target object, wherein the historical interaction resource sequence of interest activation comprises a plurality of historical interaction resources interacted by the target object in a historical period, and the plurality of historical interaction resources belong to resources of the target interest;
The object feature determining module is used for determining object features of the target object about the target interest based on the historical interaction resource sequence of interest activation and the object attribute information of the target object;
a candidate resource feature determining module, configured to determine a plurality of candidate resource features related to the target interest, where the plurality of candidate resource features are in one-to-one correspondence with a plurality of candidate resources, and the candidate resource features are used to characterize resource attribute features of the candidate resources corresponding to the candidate resource features, and the plurality of candidate resources all belong to the resource of the target interest;
and the resource to be recommended determining module is used for determining the resource to be recommended from the plurality of candidate resources based on the plurality of candidate resource characteristics and the object characteristics.
According to another aspect of the present disclosure, there is provided a training apparatus of a deep learning model, including:
the sample resource sequence determining module is used for determining a sample historical interaction resource sequence of interest activation of a sample target object, wherein the sample historical interaction resource sequence of interest activation comprises a plurality of sample historical interaction resources interacted by the sample target object in a sample historical period, and the plurality of sample historical interaction resources belong to resources of sample target interest;
A sample object feature determining module, configured to determine a sample object feature of the sample object with respect to the sample object interest based on the sample historical interaction resource sequence activated by the interest and sample object attribute information of the sample object;
the sample interaction resource determining module is used for determining target sample history interaction resources from the sample history interaction resource sequences activated by the interests of the sample target objects;
the sample resource feature determining module is used for obtaining sample resource features based on the sample resource attribute information of the target sample history interaction resources;
and the parameter adjustment module is used for adjusting parameters of the deep learning model based on the sample object characteristics and the sample resource characteristics to complete training.
According to another aspect of the present disclosure, there is provided an electronic device including: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform a method as disclosed herein.
According to another aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing computer instructions for causing a computer as described above to perform a method as disclosed herein.
According to another aspect of the present disclosure, there is provided a computer program product comprising a computer program which, when executed by a processor, implements a method as disclosed herein.
It should be understood that the description in this section is not intended to identify key or critical features of the embodiments of the disclosure, nor is it intended to be used to limit the scope of the disclosure. Other features of the present disclosure will become apparent from the following specification.
Drawings
The drawings are for a better understanding of the present solution and are not to be construed as limiting the present disclosure. Wherein:
FIG. 1 schematically illustrates a schematic diagram of a recommendation method according to a related example;
FIG. 2 schematically illustrates an exemplary system architecture in which recommendation methods and apparatus may be applied, according to embodiments of the present disclosure;
FIG. 3 schematically illustrates a flow chart of a recommendation method according to an embodiment of the present disclosure;
FIG. 4 schematically illustrates a schematic diagram of generating an offline database according to an embodiment of the disclosure;
FIG. 5 schematically illustrates a schematic diagram of generating a feature of interest in accordance with an embodiment of the present disclosure;
FIG. 6 schematically illustrates a schematic diagram of a training method of the neural network of FIG. 6, in accordance with an embodiment of the present disclosure;
FIG. 7 schematically illustrates a schematic diagram of a recommendation method according to another embodiment of the present disclosure;
FIG. 8 schematically illustrates a schematic diagram of a recommendation model, according to an embodiment of the present disclosure;
FIG. 9 schematically illustrates a flow chart of a training method of a deep learning model according to an embodiment of the disclosure;
FIG. 10 schematically illustrates a block diagram of a recommendation device, according to an embodiment of the present disclosure;
FIG. 11 schematically illustrates a block diagram of a training apparatus of a deep learning model in accordance with an embodiment of the present disclosure; and
fig. 12 schematically illustrates a block diagram of an electronic device adapted to implement a recommendation method, according to an embodiment of the disclosure.
Detailed Description
Exemplary embodiments of the present disclosure are described below in conjunction with the accompanying drawings, which include various details of the embodiments of the present disclosure to facilitate understanding, and should be considered as merely exemplary. Accordingly, one of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present disclosure. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.
According to a related example, in the recommendation field, a deep neural network (Deep Neural Networks, DNN) may be employed as a recommendation model to recall resources to be recommended. For example, relevant resource attribute information of the advertisement resource, such as the clicked advertisement resource, and the plurality of candidate advertisement resources, which are interacted by the target object in the history period, is input into the recommendation model, and the result is output. Based on the results, a resource to be recommended is determined from the plurality of candidate advertising resources.
In the process of implementing the embodiment of the disclosure, the recommendation method is discovered, focusing on the propagation relationship among a plurality of historical interaction resources in the historical interaction resource sequence of the target object, and discovering the similarity among the resources, and does not have the capability of expressing the long-period interest of the target object. Therefore, the recommendation method does not have the ability to recall the resources to be recommended, which are attributed to long period interests, with high precision.
According to another related example, in the recommendation field, a double tower model may also be employed as a recommendation model to recall resources to be recommended. For example, the features of User (target object) and the features of Item (resource) are respectively extracted through a double-depth neural network, the object features and the resource attribute features are obtained, and the attention or interest degree of the target object to the resource is calculated through cosine similarity.
Fig. 1 schematically shows a schematic diagram of a recommendation method according to a related example.
As shown in fig. 1, the recommendation model may include a dual tower model including an Embedding layer (Embedding layer) M111, an object processing module M112, a resource processing module M122, and a similarity calculation layer M131. The historical interaction resource sequence 111 of the target object may be input to the embedding layer M111 to obtain a historical interaction resource feature sequence of the target object, and the historical interaction resource feature sequence of the target object may be input to the object processing module M112 to obtain an object feature. The resource attribute information 121 of the resource is input to the embedded layer M111 to obtain a resource coding feature, and the resource coding feature is input to the resource processing module M122 to obtain a resource attribute feature. The object features and the resource attribute features are input to the similarity calculation layer M131, and the similarity 131 between the object features and the resource attribute features is determined. It is determined whether to regard the resource as a resource to be recommended based on the similarity 131.
In the process of implementing the embodiment of the disclosure, the recommendation method is found, and features of a historical interaction resource sequence of a target object can be introduced, and the historical interest features of the target object are established through a cyclic neural network (Recurrent Neural Network, RNN), an Attention mechanism (Attention), and the like. However, on the one hand, when the historical interaction resource sequence of the target object is longer, the interests of the related different interest categories are more and more, the weight distinction degree of the interest feature of the target object output by the recommendation model on each interest category is lower and lower, and further, the preference on the whole interest space is more and more fuzzy, so that the recall precision is reduced on the resources related to each interest. On the other hand, due to the difference of interest frequency, the recommendation method still leads to the high-frequency interest to be dominant in the weight of the interest feature, so that the resources belonging to the interest with long period cannot be effectively recalled.
In the recommendation field, positive feedback behavior may refer to the interaction behavior of a target object occurring within a historical period of time. Such as advertisement clicks, form fills, order payments, credit applications, etc. For different recommended scenes, the period of continuous positive feedback behaviors generated by the target object is different, for example, in daily commodity scenes, the positive feedback duty ratio is high; on the contrary, in the scenes of financial credit, payment training and the like, the positive feedback duty ratio is low.
By using the traditional recommendation method, the long-period interests of the target object cannot be found effectively in the recall stage, so that recall results are concentrated in resources belonging to high-frequency interests, and resources related to the long-period interests cannot be recalled effectively. In addition, recall accuracy is low even if a small number of resources that are of long period interest are recalled. Finally, resources of industries related to long-period interests cannot be positioned efficiently, popularization effects are poor, and user experience is poor.
The disclosure provides a recommendation method, a training method and device of a deep learning model, electronic equipment, a storage medium and a program product.
According to an embodiment of the present disclosure, there is provided a recommendation method including: and determining a historical interaction resource sequence of interest activation of the target object, wherein the historical interaction resource sequence of interest activation comprises a plurality of historical interaction resources interacted by the target object in a historical period, and the plurality of historical interaction resources belong to the resources of the target interest. Object characteristics of the target object with respect to the target interests are determined based on the historical interaction resource sequences of interest activation and the object attribute information of the target object. A plurality of candidate resource features related to the target interest are determined, wherein the plurality of candidate resource features are in one-to-one correspondence with the plurality of candidate resources, the candidate resource features are used for characterizing resource attribute features of the candidate resources corresponding to the candidate resource features, and the plurality of candidate resources belong to the resources of the target interest. A resource to be recommended is determined from the plurality of candidate resources based on the plurality of candidate resource features and the object feature.
According to the embodiment of the disclosure, a resource recall algorithm based on interest activated historical interaction resource serialization modeling is provided, the historical interaction resource sequence of a target object, which occurs in a historical period, such as a full life cycle, is subjected to ultralong span serialization modeling, and the historical interaction resource sequence is processed by interest activation, so that the interest activated historical interaction resource sequence simultaneously takes into consideration high-frequency and low-frequency interests, and high-precision recall of the full quantity of candidate resources is realized. The resource recommendation precision can be remarkably improved, the user experience is improved, and the recommendation efficiency is improved.
In the technical scheme of the disclosure, the related processes of collecting, storing, using, processing, transmitting, providing, disclosing, applying and the like of the personal information of the user all conform to the regulations of related laws and regulations, necessary security measures are adopted, and the public order harmony is not violated.
In the technical scheme of the disclosure, the authorization or consent of the user is obtained before the personal information of the user is obtained or acquired.
Fig. 2 schematically illustrates an exemplary system architecture to which recommendation methods and apparatus may be applied, according to embodiments of the present disclosure.
It should be noted that fig. 2 is only an example of a system architecture to which embodiments of the present disclosure may be applied to assist those skilled in the art in understanding the technical content of the present disclosure, but does not mean that embodiments of the present disclosure may not be used in other devices, systems, environments, or scenarios. For example, in another embodiment, an exemplary system architecture to which the recommendation method and apparatus may be applied may include a terminal device, but the terminal device may implement the recommendation method and apparatus provided by the embodiments of the present disclosure without interacting with a server.
As shown in fig. 2, the system architecture 200 according to this embodiment may include terminal devices 201, 202, 203, a network 204, and a server 205. The network 204 is the medium used to provide communication links between the terminal devices 201, 202, 203 and the server 205. The network 204 may include various connection types, such as wired and/or wireless communication links, and the like.
The user may interact with the server 205 via the network 204 using the terminal devices 201, 202, 203 to receive or send messages or the like. Various communication client applications may be installed on the terminal devices 201, 202, 203, such as a knowledge reading class application, a web browser application, a search class application, an instant messaging tool, a mailbox client and/or social platform software, etc. (as examples only).
The terminal devices 201, 202, 203 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smartphones, tablets, laptop and desktop computers, and the like.
The server 205 may be a server providing various services, such as a background management server (by way of example only) providing support for content browsed by a user using the terminal devices 201, 202, 203. The background management server may analyze and process the received data such as the user request, and feed back the processing result (e.g., the web page, information, or data obtained or generated according to the user request) to the terminal device.
It should be noted that, the recommendation method provided by the embodiments of the present disclosure may be generally performed by the terminal device 201, 202, or 203. Accordingly, the recommendation device provided in the embodiments of the present disclosure may also be provided in the terminal device 201, 202, or 203.
Alternatively, the recommendation methods provided by the embodiments of the present disclosure may also be generally performed by the server 205. Accordingly, the recommendation device provided in the embodiments of the present disclosure may be generally disposed in the server 205. The recommendation method provided by the embodiments of the present disclosure may also be performed by a server or a server cluster that is different from the server 205 and is capable of communicating with the terminal devices 201, 202, 203 and/or the server 205. Accordingly, the recommendation apparatus provided by the embodiments of the present disclosure may also be provided in a server or a server cluster that is different from the server 205 and is capable of communicating with the terminal devices 201, 202, 203 and/or the server 205.
For example, when a user browses videos online, the terminal device 201, 202, 203 may acquire a login request or a browsing request of the user, and send the login request or the browsing request to the server 205, and the server 205 determines a historical interaction resource sequence activated by interest of the user in response to the login request or the browsing request. Based on the historical interaction resource sequence of interest activation and the object attribute information of the user, object characteristics of the user about the target interest are determined. A plurality of candidate resource features are determined that pertain to the target interest. A resource to be recommended is determined from the plurality of candidate resources based on the plurality of candidate resource features and the object feature. And sends the resources to be recommended to the terminal devices 201, 202 and 203, so as to achieve the purpose of recommending the resources to be recommended to the user. Or the server cluster capable of communicating with the terminal devices 201, 202, 203 and/or the server 205 responds to the login request or the browsing request of the user, analyzes the user, and finally achieves the purpose of recommending the resource to be recommended to the user.
It should be understood that the number of terminal devices, networks and servers in fig. 1 is merely illustrative. There may be any number of terminal devices, networks, and servers, as desired for implementation.
It should be noted that the sequence numbers of the respective operations in the following methods are merely representative of the operations for the purpose of description, and should not be construed as representing the order of execution of the respective operations. The method need not be performed in the exact order shown unless explicitly stated.
Fig. 3 schematically shows a flow chart of a recommendation method according to an embodiment of the present disclosure.
As shown in FIG. 3, the method 300 includes operations S310-S340.
In operation S310, a historical interaction resource sequence of interest activation of the target object is determined.
In operation S320, object characteristics of the target object with respect to the target interests are determined based on the historical interaction resource sequences of interest activation and the object attribute information of the target object.
In operation S330, a plurality of candidate resource features regarding the target interest are determined.
In operation S340, a resource to be recommended is determined from the plurality of candidate resources based on the plurality of candidate resource characteristics and the object characteristics.
According to an embodiment of the present disclosure, the interest activated historical interaction resource sequence includes a plurality of historical interaction resources that the target object interacted with in the historical period, the plurality of historical interaction resources all belonging to the resource of the target interest.
According to embodiments of the present disclosure, the target object may refer to a user who obtains network resources using a terminal device.
According to embodiments of the present disclosure, resources may be used to characterize content presented to a target object, including, for example, but not limited to: short video, long video, news, advertisement, etc.
According to embodiments of the present disclosure, interacted historical interaction resources may refer to resources where a target object has interacted with during a historical period. The interactive behavior may refer to browsing, collecting, clicking, purchasing, filling in forms or trust applications, etc. The plurality of historical interaction resources can be ordered according to the interaction time sequence to obtain a historical interaction resource sequence.
According to an embodiment of the present disclosure, the historical interaction resource sequence of interest activation may be obtained by interest activation of the historical interaction resource sequence of the target object. For example, based on interest categories of interests, at least one target historical interaction resource that belongs to the same interest is determined from a sequence of historical interaction resources. And obtaining a historical interaction resource sequence of interest activation based on at least one target historical interaction resource.
For example, the historical interaction resource sequence of the target object may include { Ad } 1 、Ad 2 、Ad 3 、Ad 4 、......、Ad n }，Ad n Representing the nth advertising asset that the target object interacted with during the history period. Wherein, the history interaction resource Ad 1 、Ad 2 、Ad 3 Advertisement resources belonging to the interest category of the electronic product are the same as the target interests. Historical interaction resource Ad 4 And Ad (Ad) 5 Advertisement resources belonging to the interest class of living things as well as target interests. Based on the historical interaction resource sequence of the target object, obtaining a historical interaction resource sequence { Ad ] with target interest activated for the interest of the electronic product interest category 1 、Ad 2 、Ad 3 Historical interaction resource sequence { Ad } for interest activation of interest in a target interest class for living things 4 、Ad 5 }。
According to an embodiment of the present disclosure, a plurality of candidate resource features are in one-to-one correspondence with a plurality of candidate resources, the candidate resource features are used to characterize resource attribute features of the candidate resources corresponding to the candidate resource features, and the plurality of candidate resources all belong to the resources of the target interest.
According to an embodiment of the present disclosure, object features are derived based on a sequence of historical interaction resources activated by an interest, the characterized features comprising resource features of the historical interaction resources belonging to the target interest. Therefore, the target object can be used as reference data such as object characteristics for screening the resources to be recommended in a long history period such as a full life cycle period, is a characteristic distinguished in interests, other interest characteristics are shielded, doping of characteristics irrelevant to the target interests is avoided, the difference of interest frequency is ignored, and interest preference in an interest space is highlighted. In addition, the plurality of candidate resources also belong to the resources of the target interest. Therefore, the recommended resources and the reference data belong to the same interest space, and the aim of screening the sub-interests is fulfilled. The candidate resource characteristics and the object characteristics are compared, and the resource to be recommended is determined from the candidate resources, so that the resource to be recommended is the resource belonging to the target interest, is close to the object characteristics of the target object, meets the personalized requirements of the target object, improves the personalized experience of the target object, and improves recall accuracy.
In accordance with an embodiment of the present disclosure, before performing operation S310 as shown in fig. 3, determining a history of interaction resource sequences of interest activation of a target object, the recommendation method may further include: based on the historical interaction resource sequence, a historical interaction resource sequence for interest activation is generated.
According to an embodiment of the present disclosure, generating a historical interaction resource sequence for interest activation based on the historical interaction resource sequence may include: and acquiring a historical interaction resource sequence of the target object in the historical period. And according to the mapping relation between the interests and the resources, performing interest activation on the historical interaction resource sequence to obtain the historical interaction resource sequence activated by the interests.
According to embodiments of the present disclosure, the mapping relationship between interests and resources may be used to characterize whether resources are attributed to interests. For example, advertisement resource A for a computer and advertisement resource B for a mouse are of interest belonging to the interest class of electronic products. Then there is a mapping relationship between advertisement resource a and advertisement resource B, respectively, and the interests of the electronic product interest categories.
According to an embodiment of the present disclosure, the mapping relationship between the interests and the resources may be generated by statistics performed in advance according to an open source data set, for example, a mapping relationship list between the interests and the resources is generated. The mapping relationship between the target interests and the historical interaction resources may be determined based on the mapping relationship list. But is not limited thereto. And the mapping relation between the interests and the resources can be determined according to the similarity between the interests and the historical interaction resource characteristics.
According to the embodiment of the disclosure, according to the mapping relation between interests and resources, the historical interaction resource sequences are subjected to interest activation to obtain the historical interaction resource sequences activated by the interests, so that the historical interaction resource sequences activated by the interests can be ensured to ignore the interest frequency and also ignore the occurrence time period of the historical interaction behaviors, and the historical interaction resource sequences activated by the interests are compatible with the historical interaction resources belonging to low-frequency long-period interests and high-frequency short-period interests, so that the historical interaction resource sequences activated by the interests contain abundant information and are high in accuracy.
In accordance with an embodiment of the present disclosure, before operation S310 shown in fig. 3, the recommendation method may further include the operations of: and acquiring a historical interaction resource sequence of the target object from the starting moment to the current moment.
According to an embodiment of the present disclosure, the start time is used to characterize a historical registration time of the target object. For example, the target object downloads the application program by using the terminal device, and logs in the application program at the moment. But is not limited thereto. And the time of completing registration can also refer to filling in related information. The time period from the starting time to the current time is taken as a history time period, and the history time period can be understood as a full life cycle time period of the target object.
According to an embodiment of the disclosure, a historical interaction resource sequence of a target object in a historical period is obtained, and a historical interaction resource sequence of interest activation is determined based on the historical interaction resource sequence in the historical period. The historical interaction resource sequences capable of guaranteeing interest activation are sequences after interest differentiation is carried out on the historical interaction resources, and meanwhile, the historical interaction resource sequences capable of achieving interest activation are compatible with the historical interaction resources belonging to low-frequency long-period interests and high-frequency short-period interests. Therefore, the influence of interest frequency and the occurrence period of historical interaction behavior on recall operation is avoided, and the problem that low-frequency interests cannot be effectively represented while mutual interference of multiple interests is avoided.
According to the embodiment of the disclosure, according to the mapping relation between interests and resources, the historical interaction resource sequence is subjected to interest activation, and the historical interaction resource sequence subjected to interest activation is obtained, which can comprise the following operations.
For example, according to the mapping relation between interests and resources, according to preset activation weights, performing interest activation on each historical interaction resource in the historical interaction resource sequence with respect to the target interests to obtain the historical interaction resource sequence with interest activation.
According to an embodiment of the present disclosure, an activation weight w is preset i The determination of (2) may be performed by the following formula (1).
Wherein f (a) i ) Representing the ith historical interaction resource a i Interest in f (a) cand ) Representing target interest, at the ith historical interaction resource a i In the case that the interest of (1) belongs to the target interest, presetting an activation weight w i 1, and vice versa is 0.
According to an embodiment of the present disclosure, according to a preset activation weight, performing interest activation on a target interest on each historical interaction resource in a historical interaction resource sequence to obtain a historical interaction resource sequence with interest activation may include: and multiplying the preset activation weight by each historical interaction resource in the historical interaction resource sequence to obtain the historical interaction resource sequence activated by the interest.
According to an embodiment of the present disclosure, the historical interaction resource sequence of interest activation may be obtained by the following formula (2).
Wherein,and representing a historical interaction resource sequence of interest activation, wherein Au represents the historical interaction resource sequence.
According to the embodiment of the disclosure, the preset activation weight is set, so that the historical interaction resources related to the target interests can be highlighted by using the preset activation weight, and recall accuracy is further improved. In addition, the preset activation weight is set to be a value of 0 or 1, so that the historical interaction resource sequence of interest activation only comprises at least one historical interaction resource belonging to the same interest, the preset activation weight is equal to 0, and the fact that no historical interaction resource of the target object is activated in the interest is indicated, and the recall meaning is not achieved. Therefore, the purpose of interest analysis is achieved, interference among a plurality of historical interaction resources of multiple interests in the historical interaction resource sequence of interest activation is avoided, and the problem that low-frequency interests cannot be effectively represented is avoided.
According to embodiments of the present disclosure, a historical interaction resource sequence of a target object may be determined based on identification information of the target object in a request in response to the request of the target object. And performing interest activation on the historical interaction resource sequence based on the mapping relation between the interest and the resource to obtain the historical interaction resource sequence activated by the interest.
According to the embodiment of the disclosure, in the process of executing the recommendation method provided by the embodiment of the disclosure by the terminal equipment, interest activation is performed on the historical interaction resource sequence of the target object in real time, the calculated amount is large, the calculation is time-consuming, and the response requirement of the target object cannot be met.
According to an alternative embodiment of the present disclosure, for operation S310 as shown in fig. 3, determining a historical interaction resource sequence for interest activation of a target object may include: a historical interaction resource sequence for interest activation of the target object is determined from an offline database based on the identification information of the target object and the interest information of the target interest.
According to embodiments of the present disclosure, the interest information of the target interest may include an interest tag or interest identification for characterizing the interest category, as long as it is information of unique attribute information for characterizing the target interest.
According to an embodiment of the present disclosure, the identification information of the target object may be information for uniquely identifying the target object with other objects. For example, identification information of the target object.
According to embodiments of the present disclosure, an offline database may be generated using a plurality of historical interaction resource sequences of interest activation. Generating a vocabulary and storing the vocabulary into an offline database for online query based on object identification, interest information and a historical interaction resource sequence of interest activation of the object.
Fig. 4 schematically illustrates a schematic diagram of generating an offline database according to an embodiment of the disclosure.
As shown in FIG. 4, for Interest1, the historical interaction resource sequence A-S-400 of the object A400 is Interest activated to obtain the Interest activated historical interaction resource sequence of the object A400, interest1-A. And performing Interest activation on the historical interaction resource sequence B-S-400 of the object B400 to obtain an Interest activated historical interaction resource sequence Intrest 1-B of the object B400. And performing Interest activation on the historical interaction resource sequence N-S-400 of the object N400 to obtain an Interest activated historical interaction resource sequence Intrest 1-N of the object N.
And aiming at the Interest of the Interest tag Interest2, performing Interest activation on the historical interaction resource sequence A-S-400 of the object A400 to obtain the historical interaction resource sequence Interest2-A of the Interest activation of the object A400. And performing Interest activation on the historical interaction resource sequence B-S-400 of the object B400 to obtain an Interest activated historical interaction resource sequence Intrest 2-B of the object B. And performing Interest activation on the historical interaction resource sequence N-S-400 of the object N400 to obtain an Interest activated historical interaction resource sequence Intrest 2-N of the object N400.
The offline database DB400 may be generated based on the historical interaction resource sequence of Interest activation of the object a400, the historical interaction resource sequence of Interest activation of the object B400, the historical interaction resource sequence of Interest1-B of Interest activation of the object N400, the historical interaction resource sequence of Interest1-N of Interest activation of the object a400, the historical interaction resource sequence of Interest2-a of Interest activation of the object a400, the historical interaction resource sequence of Interest2-B of Interest activation of the object B400, and the historical interaction resource sequence of Interest2-N of Interest activation of the object N400.
According to the embodiment of the disclosure, an offline database can be stored in a terminal device, and in response to a request of a target object, a historical interaction resource sequence of interest activation of the target object is directly acquired from the offline database based on the identification information of the target object and the interest information of the target interest in the request. Therefore, the processing efficiency is improved, the hardware energy consumption of the terminal equipment is reduced, the hardware processing performance of the terminal equipment is improved, and the response processing efficiency is improved.
According to an exemplary embodiment of the present disclosure, before performing operation S310 as shown in fig. 3, the recommendation method may further include the operations of: and determining the mapping relation between the interests and the resources.
According to embodiments of the present disclosure, a mapping relationship between interests and resources may be determined based on the interest features and the historical interaction resource features.
According to embodiments of the present disclosure, historical interaction resource characteristics are used to characterize resource attribute characteristics of historical interaction resources. The interest feature is used to characterize a feature related to the interest information.
According to an embodiment of the present disclosure, determining a mapping relationship between interests and resources based on the interest features and the historical interaction resource features may include: and determining the similarity between the interest characteristic and the historical interaction resource characteristic. And determining the mapping relation between the interests and the resources based on the similarity.
According to an embodiment of the present disclosure, determining a mapping relationship between interests and resources based on similarity may include: and under the condition that the similarity is greater than or equal to a preset similarity threshold, determining that the mapping relation between the interests and the resources belongs to the interests for the historical interaction resources. Otherwise, under the condition that the similarity is smaller than a preset similarity threshold, determining that the mapping relation between the interests and the resources is that the historical interaction resources are not attributed to the interests.
According to the embodiment of the disclosure, the mapping relation between the interests and the resources is determined by utilizing the interest characteristics and the historical interaction resource characteristics, so that the judgment result is quantized, the accuracy of determining the mapping relation between the interests and the resources is improved, the processing range is improved, and the problem that the historical interaction resources cannot be judged because the historical interaction resources are not in a preset mapping relation list between the interests and the resources is avoided.
In accordance with an embodiment of the present disclosure, before operation S310 shown in fig. 3, the recommendation method may further include: an interest feature of interest is generated.
According to the embodiment of the disclosure, the interest information can be subjected to feature extraction to obtain the interest feature. For example, for the interest category a, interest information of the interest category a is extracted, and the interest feature of the interest category a is obtained. But is not limited thereto. Interest features of interest may also be generated, which may include the following operations.
For example, for each of a plurality of object interaction resources, an object interaction resource characteristic is obtained based on resource attribute information of the object interaction resource. The object interaction resource is used for representing the history interaction resource interacted by the object. And clustering the plurality of object interaction resource characteristics to obtain at least one object interaction resource characteristic cluster. At least one object interaction resource feature cluster corresponds one-to-one with at least one interest. And obtaining interest features of the interests corresponding to the object interaction resource feature clusters based on each object interaction resource feature cluster.
According to embodiments of the present disclosure, the object interaction resource may be a historical interaction resource of the object in the open source data. Object interaction resources for each of a plurality of objects may be collected from open source data. The larger the data volume is, the more abundant the object interaction resource feature clusters are obtained by clustering.
According to an embodiment of the present disclosure, obtaining the object interaction resource feature based on the resource attribute information of the object interaction resource may include: and extracting the characteristics of the resource attribute information of the object interaction resource to obtain the characteristics of the object interaction resource. For example, the characteristics of the resource attribute information of the object interaction resource may be extracted by using the graph neural network (Graph Neural Network), but the present invention is not limited thereto, and any network may be used as long as it can extract the characteristics.
According to the embodiment of the disclosure, the plurality of object interaction resource features can be clustered through a clustering algorithm to obtain at least one object interaction resource feature cluster. The size of each object interaction resource feature cluster can be at 10 5 Magnitude. The larger the data volume is, the more accurate the clustering effect of the object interaction resource feature clusters is. Each object interaction resource feature cluster represents an interest.
According to an embodiment of the present disclosure, obtaining, based on each object interaction resource feature cluster, an interest feature of interest corresponding to the object interaction resource feature cluster may include: and calculating the average value of the object interaction resource feature clusters to obtain the interest features of the corresponding interests.
According to the embodiment of the disclosure, the object interaction resource characteristics of a plurality of objects are clustered by carrying out vector characterization on the resource attribute information of the object interaction resource of the objects, so that the object interaction resource characteristic cluster after big data statistics is obtained, the interest characteristic of interest is represented by the vector average value of the object interaction resource characteristic cluster, and the interest of the abstract concept can be specified and simultaneously digitized, so that the accuracy and the application range of the interest characteristic are improved.
Fig. 5 schematically illustrates a schematic diagram of generating a feature of interest according to an embodiment of the present disclosure.
As shown in fig. 5, a collection of object interaction resources 510, e.g., ad1, ad2, ad3, adn, is collected. The resource attribute information of each of the plurality of object interaction resources in the object interaction resource set 510 is input to the neural network M510, so as to obtain an object interaction resource feature set 520. The object interaction resource feature sets 520 are clustered to obtain a plurality of object interaction resource feature subsets, and each object interaction resource feature subset is used as an object interaction resource feature cluster to obtain a plurality of object interaction resource feature clusters 530. Each object interaction resource feature cluster is assigned to an interest. For example, ad1, ad4, ad..adj's object interaction resource feature cluster is attributed to Interest1.Ad2, ad3, ad.i. the object interaction resource feature cluster of Adi belongs to interest. For each interest, the clusters of object interaction resource features are averaged to obtain an interest feature 540 for that interest. For example, an Interest feature 1 of Interest1, an Interest feature k of Interest rest, is obtained.
Fig. 6 schematically illustrates a schematic diagram of a training method of the neural network of fig. according to an embodiment of the present disclosure.
As shown in fig. 6, a sequence of object interaction resources 610 for each of a plurality of objects may be collected, e.g., object interaction resources for which a historical click behavior has occurred for a historical period of time for the object may be combined to obtain a sequence of object interaction resources.
As shown in FIG. 6, an interaction resource map 620 may be generated based on a plurality of object interaction resource sequences 610. The interaction resource map 620 includes a plurality of nodes, each node representing an object interaction resource, and an edge between two nodes P-a and P-B representing an object interaction resource a and an object interaction resource B having an adjacency relationship in the sequence of object interaction resources. The edges are directed edges, and the arrow of the edge between the nodes P-A and P-B points from the node P-A to the node P-B, representing that the historical click behavior of the object is: click on historical interaction resource a first, then click on historical interaction resource B second, and vice versa. The side band has a weight indicating how often the click behavior occurs. There may be two directional edges between two nodes at the same time, but the weights may be different.
As shown in fig. 6, random walk sampling is performed on a frequency basis based on directed edges in the interaction resource map 620, resulting in a plurality of interaction resource sequences as a training sample set 630.
As shown in fig. 6, the graph neural network M610 is trained using the training sample set 630, resulting in a graph neural network for extracting object interaction resource features of the object interaction resource.
According to the embodiment of the disclosure, based on the directed edges in the interaction resource map, random walk sampling is performed according to frequency to obtain a plurality of interaction resource sequences, the property that all the plurality of interaction resources in the interaction resource sequences belong to the same interest is utilized, the plurality of interaction resource sequences are used as training sample sets to train the graph neural network, the graph neural network can learn the characteristics of the interaction resources of the same interest, and further the trained graph neural network is utilized to extract that the object interaction resource characteristics of the object interaction resources are vector representations on the interest, so that the accuracy of the graph neural network in extracting the interaction resource characteristics is improved.
According to an embodiment of the present disclosure, for operation S320 as shown in fig. 3, determining object characteristics of a target object with respect to a target interest based on a history of interaction resource sequences of interest activation and object attribute information of the target object may include the following operations.
For example, the historical interaction resource sequence of interest activation is subjected to serialization characterization, so that the feature of the interaction resource of interest activation is obtained. And activating the interactive resource features based on the object attribute information and the interests of the target object to obtain object features of the target object about the target interests.
According to an embodiment of the present disclosure, performing a sequential characterization on a historical interaction resource sequence of interest activation to obtain an interest activation interaction resource feature may include: and extracting the serialization features of the historical interaction resource sequences activated by the interests to obtain the features of the interaction resources activated by the interests.
In accordance with embodiments of the present disclosure, serialized characterization may refer to serialized feature extraction. Specifically, the cross characteristics between the interest activated historical interaction resources and other interest activated historical interaction resources except the interest activated historical interaction resources in the interest activated historical interaction resource sequence can be determined by indicating the interest activated historical interaction resources in each interest activated historical interaction resource in the interest activated historical interaction resource sequence, so that the global characterization vector of the interest activated historical interaction resources is obtained, and the overall depiction of the target object is enhanced.
According to the embodiment of the disclosure, the embedded layer (Embedding layer) can be utilized to carry out vectorization processing on the historical interaction resource sequence of interest activation, so as to obtain the historical interaction resource vector sequence of interest activation. The historical interaction resource vector sequence of interest activation can be characterized in a serialization manner by using a multi-layer stacked coder-decoder (transducer) to obtain the interaction resource characteristics of interest activation. But is not limited thereto. The historical interaction resource vector sequence of interest activation can be further characterized in a serialization manner by using a multi-layer stacked decoder (transducer-decoder) to obtain the interaction resource characteristics of interest activation.
According to the embodiment of the disclosure, the historic interaction resource vector sequence of interest activation is subjected to serialization characterization by using the multi-layer stacked decoder, so that the adverse sequence influence can be avoided, and the characteristics of the interaction resource of interest activation are accurate and effective.
According to an embodiment of the present disclosure, object attribute information (profile) of the target object may include object representation information of the target object. For example, identity information such as the sex, age, etc. of the target object or other attribute information.
According to the embodiment of the disclosure, the embedded layer can be utilized to carry out vectorization processing on the object attribute information of the target object, so as to obtain the object attribute information vector of the target object.
According to the embodiment of the disclosure, the object attribute information vector and the interest activation interaction resource feature of the target object can be input into the object processing layer, so that the object feature of the target object about the target interest can be obtained.
According to an embodiment of the present disclosure, the object processing layer may include a DNN network, but is not limited thereto, as long as the network is capable of extracting an object attribute information vector of a target object and features of interest activation interaction resource features, and obtaining object features of the target object with respect to the target interest.
According to the embodiment of the disclosure, the historical interaction resource sequence of interest activation is subjected to sequential characterization to obtain the feature of the interest activation interaction resource, so that the feature of the interest activation interaction resource can be accurate and effective. In addition, the interactive resource features are activated based on the object attribute information vector and the interests of the target object, so that the object features of the target object about the target interests are obtained, feature information contained in the object features can be rich and accurate, and recall accuracy of resources to be recommended, which are obtained based on the object features, is improved.
According to an embodiment of the present disclosure, for operation S330 as shown in fig. 3, determining a plurality of candidate resource features regarding the target interest may include: a plurality of candidate resources corresponding to the target interests is determined. A plurality of candidate resource features relating to the target interest are determined based on the respective resource attribute information of the plurality of candidate resources. But is not limited thereto. A plurality of resource characteristics may also be determined based on the resource attribute information for each of the plurality of resources. A plurality of candidate resources corresponding to the target interest and candidate resource characteristics of each of the plurality of candidate resources are determined from the plurality of resources based on a similarity between the resource characteristics and the interest characteristics.
According to the embodiment of the disclosure, the resource attribute information of each of a plurality of resources can be input into the advertisement processing module, so that a plurality of resource characteristics are obtained.
According to an embodiment of the present disclosure, the interest of a resource is determined based on the resource characteristics, and the following formula (3) may be employed.
Wherein, in the resource and resource cluster i And under the condition that the historical interaction resources are the same, taking the interest i of the historical interaction resources as the interest f (a) of the resources. Computing resource characteristics W in the event that the resource is not in multiple resource cluster groups a e a Interest feature W with ith interest i e i And obtaining a plurality of similarities, wherein the interest with the largest similarity is used as the interest f (a) of the resource. The resource clusters may represent resource clusters corresponding to object interaction resource feature clusters.
According to embodiments of the present disclosure, a plurality of candidate resources corresponding to a target interest may be determined from a plurality of resources based on the interest of the resources in response to a request of the target object. The resource may refer to content pre-stored in a database for presentation to the subject.
According to the embodiment of the disclosure, in the process of executing the recommendation method provided by the embodiment of the disclosure by the terminal equipment, feature extraction is performed on the candidate resources in real time, and a plurality of candidate resource features related to the target interest are determined.
According to an alternative embodiment of the present disclosure, for operation S330 as shown in fig. 3, determining a plurality of candidate resource features regarding the target interest may include: a target resource index corresponding to the target interest is determined from the plurality of resource indices. A plurality of candidate resource features that match the interest feature of the target interest are determined based on the target resource index.
According to the embodiment of the disclosure, the mapping relation between the candidate resource characteristics and the interests can be obtained in advance in an offline state. And generating a resource index based on the mapping relation between the candidate resource characteristics and the interests. Illustratively, an ANN index (Approximate Nearest Neighbor, approximate neighbor) may be generated. And the ANN index is utilized to facilitate the searching of the feature vector.
According to an embodiment of the present disclosure, it is found in practical implementations that a plurality of candidate resource features corresponding to a plurality of interests of different interest categories are each different from each other in feature vector representation. Accordingly, a resource index corresponding to an interest may be generated based on the interest category of the interest to increase index efficiency, thereby increasing efficiency and accuracy in determining a plurality of candidate resource features for the target interest.
In accordance with an embodiment of the present disclosure, before performing operation S330 as shown in fig. 3, determining a plurality of candidate resource features regarding a target interest, the recommendation method may further include the following operations.
For example, for each of a plurality of interests, a set of resource characteristics is determined based on respective resource attribute information for a plurality of resources. A plurality of resource indexes is generated based on the plurality of resource feature sets. The plurality of resource indexes are in one-to-one correspondence with the plurality of interests.
According to embodiments of the present disclosure, the resource may refer to content pre-stored in the database for presentation or recommendation to the subject. The resource attribute information of the resource can be input to the embedded layer to obtain a resource attribute information vector. And inputting the resource attribute information vector into a resource processing layer to obtain the resource characteristics. The resource handling layer may include a deep neural network.
According to the embodiment of the disclosure, the resource index can be generated in advance and rapidly by using an offline processing mode, so that the response to the request of the object in the online process is facilitated, the user experience is improved, the hardware requirement of the terminal equipment is reduced, and the hardware operation performance is improved.
According to an embodiment of the present disclosure, for operation S340 as shown in fig. 3, determining a resource to be recommended from among a plurality of candidate resources based on a plurality of candidate resource characteristics and object characteristics may include: and determining the similarity between the object feature and each candidate resource feature to obtain a plurality of similarities. Based on the plurality of similarities, a resource to be recommended is determined from the plurality of candidate resources.
According to the embodiment of the present disclosure, the similarity between the object feature and each candidate resource feature may be determined using a Cosine similarity (Cosine) calculation method, but is not limited thereto, and a euclidean distance calculation method may be used as long as it is a calculation method that can be used to evaluate the similarity between the two features.
According to an embodiment of the present disclosure, determining a resource to be recommended from a plurality of candidate resources based on a plurality of similarities may include: and determining K candidate resources with highest similarity as resources to be recommended. K may be 1 or an integer greater than 1.
According to the embodiment of the disclosure, the resource to be recommended is determined from the plurality of candidate resources based on the similarity, so that the recall precision is improved, the calculation method is simplified, and the processing efficiency is improved.
According to embodiments of the present disclosure, the resources to be recommended may include a plurality of. The interests to which the plurality of resources to be recommended belong are not the same, but are not limited thereto, and the interests to which the plurality of resources to be recommended belong may be the same.
According to an embodiment of the present disclosure, after performing operation S340 as shown in fig. 3, the recommendation method may further include the operations of: and sequencing the plurality of resources to be recommended according to the similarity to obtain a sequencing result. And recommending a plurality of resources to be recommended according to the sequencing result.
According to embodiments of the present disclosure, the similarity may be used to characterize the degree of interest or interest of the target object in the recommended resource. According to the similarity, the plurality of resources to be recommended are ranked and recommended, so that the priority display sequence of the resources to be recommended is closer to the expected or personalized requirements of the target object.
According to the embodiment of the disclosure, the recommendation method can be adapted to a display interface of the terminal equipment, and a plurality of resources to be recommended are displayed in a form of a drop-down list. And sequencing the plurality of resources to be recommended according to the similarity, and displaying the resources. The speed of the target object to preferentially acquire the resource expected or interested by the target object can be improved, and the processing efficiency is further improved.
Fig. 7 schematically illustrates a schematic diagram of a recommendation method according to another embodiment of the present disclosure.
As shown in fig. 7, in an online process, a terminal device obtains a plurality of interest-activated historical interaction resource sequences 730 of a target object from an offline database 720 obtained through an offline (offine) process based on an object identification of the target object in response to a request, such as a login request, of the target object 710. The plurality of interest-activated historical interaction resource sequences 730 are in one-to-one correspondence with the plurality of target interests. Interest categories of the plurality of target interests are different.
As shown in fig. 7, a plurality of target resource indexes corresponding to the plurality of target interests one-to-one may be determined from the plurality of resource indexes 740 based on the plurality of target interests.
As shown in fig. 7, a plurality of candidate resource features 750 corresponding to the target interest are determined from each target resource index.
As shown in fig. 7, object characteristics 770 of the target object with respect to the target interest are determined based on the historical interaction resource sequences 730 of interest activation of the target interest and object attribute information 760 of the target object.
As shown in FIG. 7, a resource to be recommended for a target interest is determined from a plurality of candidate resources based on a plurality of candidate resource features 750 and object features 770 for the target interest.
According to the embodiment of the disclosure, the plurality of resources to be recommended, which are in one-to-one correspondence with the plurality of target interests, can be ranked, and the plurality of resources to be recommended are displayed according to the ranking order.
According to the embodiment of the disclosure, by using the recommendation method provided by the embodiment of the disclosure, the interest distribution of the whole life cycle of the target object can be effectively captured, the recommendation effect on the resources is improved, the interest activation of the historical interaction resource sequences is respectively carried out aiming at the interests of different interest categories, the diversity of recall results is improved, and the browsing duration and the forwarding number of the target object on the resources are improved. In addition, partial data is pre-generated in advance in an offline processing mode, so that the running efficiency of the online processing process of the terminal equipment can be ensured, and the user experience is improved in the processing efficiency and the response speed.
FIG. 8 schematically illustrates a schematic diagram of a recommendation model, according to an embodiment of the present disclosure.
As shown in fig. 8, a recommendation method may be performed using a recommendation model, resulting in resources to be recommended. The recommendation model may include: an interest activation layer M810, an embedding layer M820, a serialization characterization layer M830, an object processing layer M840, a resource processing layer M850, and a similarity calculation layer M860.
As shown in FIG. 8, a historical interaction resource sequence 810 of a target object and a target interest of a candidate resource 840 may be input to the interest activation layer M810, resulting in a historical interaction resource sequence 820 of interest activation. The historical interaction resource sequence 820 of interest activation is input to the embedded layer M820, and the historical interaction resource vector sequence of interest activation is obtained.
As shown in FIG. 8, a sequence of historical interaction resource vectors for interest activation may be input to the serialization characterization layer M830 to obtain the interaction resource features for interest activation.
As shown in fig. 8, the object attribute information 830 of the target object may be input to the embedding layer M820, and used for vectorization processing, to obtain an object attribute information vector of the target object.
As shown in fig. 8, the object attribute information vector and the interest activation interaction resource feature of the target object may be input to the object processing layer M840, resulting in an object feature of the target object about the target interest.
As shown in fig. 8, resource attribute information of the candidate resource 840 may be input to the embedding layer M820, resulting in a resource attribute information vector.
As shown in fig. 8, the resource attribute information vector may be input into the resource processing layer M850, resulting in a resource feature.
As shown in fig. 8, the object feature and each candidate resource feature may be input to the similarity calculation layer M860, to obtain a similarity. Based on the similarity, it is determined whether the candidate resource is a resource to be recommended.
According to embodiments of the present disclosure, a deep learning model may be trained using the training method of the deep learning model as shown in fig. 9 to obtain a recommended model as shown in fig. 8. The recommendation model, for example, as shown in fig. 8, performs the recommendation method provided by the embodiments of the present disclosure.
Fig. 9 schematically illustrates a flowchart of a training method of a deep learning model according to an embodiment of the present disclosure.
As shown in fig. 9, the method includes operations S910 to S950.
In operation S910, a sample history interaction resource sequence of interest activation of the sample target object is determined.
In operation S920, sample object characteristics of the sample target object with respect to the sample target interest are determined based on the sample history interaction resource sequence of interest activation and sample object attribute information of the sample target object.
In operation S930, a target sample history interaction resource is determined from the sample history interaction resource sequence of interest activation of the sample target object.
In operation S940, sample resource characteristics are obtained based on sample resource attribute information of the target sample history interaction resource.
In operation S950, parameters of the deep learning model are adjusted based on the sample object features and the sample resource features, completing training.
According to an embodiment of the present disclosure, the sample history interaction resource sequence for interest activation includes a plurality of sample history interaction resources that have been interacted by the sample target object within a sample history period, each of the plurality of sample history interaction resources belonging to a resource of sample target interest.
According to an embodiment of the present disclosure, a technical term such as a sample history interaction resource sequence of interest activation involved in a training method of a deep learning model as shown in fig. 9 is different from a technical term such as a history interaction resource sequence of interest activation involved in a recommendation method as shown in fig. 3 only in that samples are added, and the main purpose is to distinguish between differences in technical solutions, but the definitions are similar, and will not be repeated here.
According to an embodiment of the present disclosure, adjusting parameters of a deep learning model based on sample object features and sample resource features may include: and inputting the sample object characteristics and the sample resource characteristics into a loss function to obtain a loss value. Parameters of the deep learning model are adjusted based on the loss value until the loss value converges or reaches a predetermined parameter adjustment turn threshold. And taking the training-stopped deep learning model as a target deep learning model to finish training.
According to an embodiment of the present disclosure, the loss function may be a cross entropy loss function, but is not limited thereto as long as the loss function is capable of adjusting parameters of the deep learning model with the sample object feature and the sample resource feature as references.
According to embodiments of the present disclosure, sample history interaction resources of a sample target object are activated according to interest such that a sample history interaction resource sequence of interest activation is compatible with a sample short period interest (e.g., high frequency interest) and a sample long period interest (e.g., low frequency interest) of the sample target object. Thus enabling the sample target object to be a feature that is differentiated in interest over a sample history period, such as a full lifecycle period. And further, the reference object sample object characteristics for adjusting the parameters of the deep learning model are standard reference objects which ignore interest frequency differences and highlight interest preference on a sample interest space. Therefore, the training accuracy of the model can be improved, and meanwhile, the training efficiency of the model is improved.
In accordance with an embodiment of the present disclosure, before performing operation S910 as shown in fig. 9, determining a sample history interaction resource sequence of interest activation of a sample target object, the training method of the deep learning model may further include: based on the sample historical interaction resource sequence, a sample historical interaction resource sequence for interest activation is generated.
According to an embodiment of the present disclosure, generating a sample historical interaction resource sequence for interest activation based on the sample historical interaction resource sequence may include: a sample history interaction resource sequence of the sample target object in a sample history period is obtained. And according to the mapping relation between the sample interests and the resources, carrying out interest activation on the sample historical interaction resource sequence to obtain the sample historical interaction resource sequence with the interest activated.
According to the embodiment of the disclosure, according to the mapping relation between the sample interests and the resources, the sample history interaction resource sequences are subjected to interest activation to obtain the sample history interaction resource sequences subjected to interest activation, so that the sample history interaction resource sequences subjected to interest activation can be ensured to ignore interest frequencies, the occurrence time period of sample history interaction behaviors is ignored, and the sample long-period interests and the sample short-period interests can be compatible, so that the sample history interaction resource sequences subjected to interest activation contain abundant information and have high accuracy.
According to the embodiment of the disclosure, according to the mapping relation between the sample interest and the resource, the sample history interaction resource sequence is subjected to interest activation, and the sample history interaction resource sequence subjected to interest activation is obtained, which can comprise the following operations.
For example, according to the mapping relation between the sample interests and the resources, according to the preset activation weight of the samples, performing interest activation on each sample history interaction resource in the sample history interaction resource sequence with respect to the sample target interests to obtain an interest activated sample history interaction resource sequence.
According to the embodiment of the disclosure, the sample preset activation weight is set, so that sample history interaction resources related to sample target interests can be highlighted by using the sample preset activation weight, and recall accuracy is improved. In addition, the sample preset activation weight is set to be a value of 0 or 1, so that the sample history interaction resource sequence of interest activation only comprises at least one sample history interaction resource belonging to the same sample interest, and the sample preset activation weight is equal to 0, which indicates that no sample history interaction resource of a sample target object is activated on the sample interest and has no recall meaning. Therefore, the purpose of sample interest analysis is achieved, interference among a plurality of sample history interaction resources of multiple sample interests in a sample history interaction resource sequence with the activated interests is avoided, and the problem that low-frequency interests cannot be effectively represented is avoided.
According to an embodiment of the present disclosure, for operation S910 as shown in fig. 9, determining a sample history interaction resource sequence of interest activation of a sample target object may include: based on the sample identification information of the sample target object and the sample interest information of the sample target interest, a sample historical interaction resource sequence of interest activation of the sample target object is determined from a sample offline database. The sample offline database is generated using a plurality of sample historical interaction resource sequences of interest activations.
According to an embodiment of the present disclosure, a sample historical interaction resource sequence of interest activation of a sample target object is obtained directly from a sample offline database. Therefore, the processing efficiency is improved, the hardware energy consumption is reduced, the hardware processing performance is improved, and the training efficiency is improved.
In accordance with an embodiment of the present disclosure, before performing operation S910 as shown in fig. 9, determining a sample history interaction resource sequence of interest activation of a sample target object, the training method of the deep learning model may further include the following operations.
For example, a mapping relationship between sample interests and resources is determined based on sample interest features and sample historical interaction resource features. The sample historical interaction resource features are used for characterizing sample resource attribute features of the sample historical interaction resource, and the sample interest features are used for characterizing features related to sample interest information.
According to the embodiment of the disclosure, the mapping relation between the sample interest and the resource is determined by utilizing the sample interest characteristic and the sample history interaction resource characteristic, so that the judgment result is quantized, the accuracy of determining the mapping relation between the sample interest and the resource is improved, the processing range is improved, and the problem that the sample history interaction resource cannot be judged because the sample history interaction resource does not exist in a preset mapping relation list between the sample interest and the resource is avoided.
In accordance with an embodiment of the present disclosure, before operation S910 shown in fig. 9, the training method of the deep learning model may further include: sample interest features of the sample interest are generated.
For example, for each sample object interaction resource of the plurality of sample object interaction resources, sample object interaction resource characteristics are derived based on sample resource attribute information of the sample object interaction resource. The sample object interaction resource is used for representing sample history interaction resources interacted by the sample object.
Clustering the interactive resource features of the plurality of sample objects to obtain at least one sample object interactive resource feature cluster. The at least one sample object interaction resource feature cluster corresponds one-to-one with the at least one sample interest.
And obtaining sample interest features of sample interests corresponding to the sample object interaction resource feature clusters based on each sample object interaction resource feature cluster.
According to the embodiment of the disclosure, the sample object interaction resource attribute information of the sample object interaction resources of the sample objects is subjected to vector characterization, and on the basis, the sample object interaction resource characteristics of a plurality of sample objects are clustered to obtain the sample object interaction resource characteristic clusters after big data statistics, and the vector average value of the sample object interaction resource characteristic clusters is used for characterizing the interest characteristic of interest, so that the interest of the abstract concept can be specified and simultaneously digitized, and the accuracy and the application range of the sample interest characteristic are improved.
According to an embodiment of the present disclosure, for operation S920 as shown in fig. 9, determining a sample object feature of a sample target object for a sample target interest based on a sample history interaction resource sequence of interest activation and sample object attribute information of the sample target object may include:
and carrying out serialization characterization on the sample history interaction resource sequence of interest activation to obtain sample interest activation interaction resource characteristics.
And activating the interactive resource feature based on the sample object attribute information and the sample interest of the sample object to obtain the sample object feature of the sample object about the sample object interest.
According to the embodiment of the disclosure, the sample history interaction resource sequence of interest activation is subjected to sequential characterization, so that the sample interest activation interaction resource characteristics are obtained, and the sample interest activation interaction resource characteristics can be accurate and effective. In addition, the interactive resource features are activated based on the sample object attribute information vector and the sample interest of the sample object, so that the sample object features of the sample object about the sample object interest are obtained, feature information contained in the sample object features can be rich and accurate, and training precision and efficiency of training based on the sample object features are improved.
Fig. 10 schematically illustrates a block diagram of a recommendation device according to an embodiment of the disclosure.
As shown in fig. 10, the recommendation apparatus 1000 of this embodiment includes a resource sequence determination module 1010, an object feature determination module 1020, a candidate resource feature determination module 1030, and a resource to be recommended determination module 1040.
A resource sequence determination module 1010 is configured to determine a historical interaction resource sequence for interest activation of the target object. The historical interaction resource sequence activated by the interest comprises a plurality of historical interaction resources interacted by the target object in the historical period, and the plurality of historical interaction resources belong to the resources of the target interest.
The object feature determination module 1020 is configured to determine object features of the target object with respect to the target interest based on the historical interaction resource sequence of interest activation and the object attribute information of the target object.
A candidate resource feature determination module 1030 is configured to determine a plurality of candidate resource features for the target interest. The plurality of candidate resource features are in one-to-one correspondence with the plurality of candidate resources, the candidate resource features are used for representing resource attribute features of the candidate resources corresponding to the candidate resource features, and the plurality of candidate resources belong to the resources of the target interest.
The resource to be recommended determining module 1040 is configured to determine a resource to be recommended from the plurality of candidate resources based on the plurality of candidate resource features and the object feature.
According to an embodiment of the present disclosure, the resource sequence determination module 1010 includes: and a resource sequence determination unit.
And the resource sequence determining unit is used for determining a historical interaction resource sequence of interest activation of the target object from the offline database based on the identification information of the target object and the interest information of the target interest. An offline database is generated using a plurality of historical interaction resource sequences of interest activation.
According to an embodiment of the present disclosure, the recommendation device further includes: a resource sequence acquisition module and an interest activation module.
And the resource sequence acquisition module is used for acquiring the historical interaction resource sequence of the target object in the historical period.
And the interest activation module is used for carrying out interest activation on the historical interaction resource sequence according to the mapping relation between the interest and the resource to obtain the historical interaction resource sequence activated by the interest.
According to an embodiment of the present disclosure, an interest activation module includes: an interest activation unit.
And the interest activation unit is used for carrying out interest activation on each historical interaction resource in the historical interaction resource sequence according to the mapping relation between the interest and the resource and the preset activation weight to obtain the historical interaction resource sequence of interest activation.
According to an embodiment of the present disclosure, the recommendation device further includes: and a mapping relation determining module.
And the mapping relation determining module is used for determining the mapping relation between the interests and the resources based on the interest characteristics and the historical interaction resource characteristics. The historical interaction resource features are used for characterizing resource attribute features of the historical interaction resources, and the interest features are used for characterizing features related to interest information.
According to an embodiment of the present disclosure, the recommendation device further includes: the system comprises an interaction resource feature determining module, an interaction resource feature clustering module and an interest feature determining module.
And the interactive resource feature determining module is used for obtaining the object interactive resource feature based on the resource attribute information of the object interactive resource aiming at each object interactive resource in the plurality of object interactive resources. The object interaction resource is used for representing the history interaction resource interacted by the object.
And the interactive resource feature clustering module is used for clustering the interactive resource features of the plurality of objects to obtain at least one interactive resource feature cluster of the objects. At least one object interaction resource feature cluster corresponds one-to-one with at least one interest.
And the interest feature determining module is used for obtaining interest features corresponding to the object interaction resource feature clusters based on each object interaction resource feature cluster.
According to an embodiment of the present disclosure, the object feature determination module 1020 includes: and the interactive resource sequence characterization unit and the object feature determination unit.
And the interactive resource sequence characterization unit is used for carrying out serialization characterization on the historical interactive resource sequence of interest activation to obtain the interactive resource characteristics of interest activation.
And the object feature determining unit is used for activating the interactive resource feature based on the object attribute information and the interest of the target object to obtain the object feature of the target object about the target interest.
According to an embodiment of the present disclosure, candidate resource feature determination module 1030 includes: a target resource index determination unit and a candidate resource feature determination unit.
And the target resource index determining unit is used for determining a target resource index corresponding to the target interest from the plurality of resource indexes.
And the candidate resource feature determining unit is used for determining a plurality of candidate resource features matched with the interest features of the target interest based on the target resource index.
According to an embodiment of the present disclosure, the recommendation device further includes: the resource feature set determining module and the resource index generating module.
And the resource feature set determining module is used for determining a resource feature set based on the respective resource attribute information of the plurality of resources for each interest in the plurality of interests.
And the resource index generation module is used for generating a plurality of resource indexes based on the plurality of resource feature sets. The plurality of resource indexes are in one-to-one correspondence with the plurality of interests.
According to an embodiment of the present disclosure, the resource to be recommended determination module 1040 includes: and the similarity determining unit and the resource to be recommended determining unit.
And the similarity determining unit is used for determining the similarity between the object feature and each candidate resource feature to obtain a plurality of similarities.
And the resource to be recommended determining unit is used for determining the resource to be recommended from the plurality of candidate resources based on the plurality of similarities.
According to an embodiment of the present disclosure, the resources to be recommended include a plurality of.
According to an embodiment of the present disclosure, the recommendation device further includes: a ranking module and a recommending module.
And the sorting module is used for sorting the plurality of resources to be recommended according to the similarity to obtain a sorting result.
And the recommending module is used for recommending a plurality of resources to be recommended according to the sequencing result.
According to an embodiment of the present disclosure, the recommendation device further includes: and a resource sequence acquisition module.
The resource sequence acquisition module is used for acquiring a historical interaction resource sequence of the target object from the starting moment to the current moment. The start time is used to characterize the historical registration time of the target object.
Fig. 11 schematically illustrates a block diagram of a training apparatus of a deep learning model according to an embodiment of the present disclosure.
As shown in fig. 11, the training apparatus 1100 for deep learning model includes: sample resource sequence determination module 1110, sample object feature determination module 1120, sample interaction resource determination module 1130, sample resource feature determination module 1140, and parameter adjustment module 1150.
A sample resource sequence determination module 1110 for determining a sample historical interaction resource sequence of interest activation of the sample target object. The sample history interaction resource sequence of interest activation comprises a plurality of sample history interaction resources interacted by the sample target object in a sample history period, and the plurality of sample history interaction resources belong to resources of sample target interest.
A sample object feature determination module 1120 for determining sample object features of the sample target object with respect to the sample target interest based on the sample historical interaction resource sequence of interest activation and sample object attribute information of the sample target object.
A sample interaction resource determination module 1130 for determining a target sample history interaction resource from a sequence of sample history interaction resources of interest activation of a sample target object.
The sample resource feature determining module 1140 is configured to obtain sample resource features based on sample resource attribute information of the target sample history interaction resource.
The parameter adjustment module 11 is configured to adjust parameters of the deep learning model based on the sample object features and the sample resource features, and complete training.
According to an embodiment of the present disclosure, the sample resource sequence determination module includes: and a sample resource sequence determination unit.
And the sample resource sequence determining unit is used for determining a sample historical interaction resource sequence activated by the interest of the sample target object from the sample offline database based on the sample identification information of the sample target object and the sample interest information of the sample target interest. The sample offline database is generated using a plurality of sample historical interaction resource sequences of interest activations.
According to an embodiment of the present disclosure, the training apparatus of the deep learning model further includes: a sample resource sequence acquisition module and a sample interest activation module.
And the sample resource sequence acquisition module is used for acquiring a sample history interaction resource sequence of the sample target object in the sample history period.
And the sample interest activation module is used for carrying out interest activation on the sample history interaction resource sequence according to the mapping relation between the sample interest and the resource to obtain the sample history interaction resource sequence with the interest activated.
According to an embodiment of the present disclosure, a sample interest activation module includes: sample interest activation unit.
The sample interest activation unit is used for carrying out interest activation on each sample history interaction resource in the sample history interaction resource sequence according to the mapping relation between the sample interest and the resource and the preset activation weight of the sample, so as to obtain the sample history interaction resource sequence of interest activation.
According to an embodiment of the present disclosure, the training apparatus of the deep learning model further includes: and a sample mapping relation determining module.
And the sample mapping relation determining module is used for determining the mapping relation between the sample interests and the resources based on the sample interest characteristics and the sample history interaction resource characteristics. The sample historical interaction resource features are used for characterizing sample resource attribute features of the sample historical interaction resource, and the sample interest features are used for characterizing features related to sample interest information.
According to an embodiment of the present disclosure, the training apparatus of the deep learning model further includes: and a sample interaction resource characteristic determining module.
The sample interaction resource feature determining module is used for obtaining sample object interaction resource features based on sample resource attribute information of the sample object interaction resources for each sample object interaction resource in the plurality of sample object interaction resources. The sample object interaction resource is used for representing sample history interaction resources interacted by the sample object.
And the sample interaction resource feature clustering module is used for clustering the interaction resource features of the plurality of sample objects to obtain at least one sample object interaction resource feature cluster. The at least one sample object interaction resource feature cluster corresponds one-to-one with the at least one sample interest.
And the sample interest feature determining module is used for obtaining sample interest features of sample interests corresponding to the sample object interaction resource feature clusters based on each sample object interaction resource feature cluster.
According to an embodiment of the present disclosure, the sample object feature determination module includes: and the sample interaction resource sequence characterization unit and the sample object characteristic determination unit.
And the sample interaction resource sequence characterization unit is used for carrying out serialization characterization on the sample history interaction resource sequence of interest activation to obtain sample interest activation interaction resource characteristics.
And the sample object feature determining unit is used for activating the interactive resource feature based on the sample object attribute information and the sample interest of the sample object to obtain the sample object feature of the sample object about the sample object interest.
According to embodiments of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.
According to an embodiment of the present disclosure, an electronic device includes: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform a method as in an embodiment of the present disclosure.
According to an embodiment of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform a method as in an embodiment of the present disclosure.
According to an embodiment of the present disclosure, a computer program product comprising a computer program which, when executed by a processor, implements a method as an embodiment of the present disclosure.
Fig. 12 shows a schematic block diagram of an example electronic device 1200 that can be used to implement embodiments of the present disclosure. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular telephones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be exemplary only, and are not meant to limit implementations of the disclosure described and/or claimed herein.
As shown in fig. 12, the apparatus 1200 includes a computing unit 1201, which may perform various appropriate actions and processes according to a computer program stored in a Read Only Memory (ROM) 1202 or a computer program loaded from a storage unit 1208 into a Random Access Memory (RAM) 1203. In the RAM 1203, various programs and data required for the operation of the device 1200 may also be stored. The computing unit 1201, the ROM 1202, and the RAM 1203 are connected to each other via a bus 1204. An input/output (I/O) interface 1205 is also connected to the bus 1204.
Various components in device 1200 are connected to an input/output (I/O) interface 1205, including: an input unit 1206 such as a keyboard, mouse, etc.; an output unit 1207 such as various types of displays, speakers, and the like; a storage unit 1208 such as a magnetic disk, an optical disk, or the like; and a communication unit 1209, such as a network card, modem, wireless communication transceiver, etc. The communication unit 1209 allows the device 1200 to exchange information/data with other devices via a computer network, such as the internet, and/or various telecommunications networks.
The computing unit 1201 may be a variety of general and/or special purpose processing components having processing and computing capabilities. Some examples of computing unit 1201 include, but are not limited to, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), various specialized Artificial Intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital Signal Processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 1201 performs the respective methods and processes described above, such as the recommendation method and the training method of the deep learning model. For example, in some embodiments, the recommendation method and training method of the deep learning model may be implemented as a computer software program tangibly embodied on a machine-readable medium, such as the storage unit 1208. In some embodiments, part or all of the computer program may be loaded and/or installed onto device 1200 via ROM 1202 and/or communication unit 1209. When the computer program is loaded into the RAM 1203 and executed by the computing unit 1201, one or more steps of the recommendation method and the training method of the deep learning model described above may be performed. Alternatively, in other embodiments, the computing unit 1201 may be configured to perform the recommendation method and the training method of the deep learning model in any other suitable way (e.g., by means of firmware).
Various implementations of the systems and techniques described here above may be implemented in digital electronic circuitry, integrated circuit systems, field Programmable Gate Arrays (FPGAs), application Specific Integrated Circuits (ASICs), application Specific Standard Products (ASSPs), systems On Chip (SOCs), complex Programmable Logic Devices (CPLDs), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs, the one or more computer programs may be executed and/or interpreted on a programmable system including at least one programmable processor, which may be a special purpose or general-purpose programmable processor, that may receive data and instructions from, and transmit data and instructions to, a storage system, at least one input device, and at least one output device.
Program code for carrying out methods of the present disclosure may be written in any combination of one or more programming languages. These program code may be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus such that the program code, when executed by the processor or controller, causes the functions/operations specified in the flowchart and/or block diagram to be implemented. The program code may execute entirely on the machine, partly on the machine, as a stand-alone software package, partly on the machine and partly on a remote machine or entirely on the remote machine or server.
In the context of this disclosure, a machine-readable medium may be a tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. The machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.
To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and pointing device (e.g., a mouse or trackball) by which a user can provide input to the computer. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic input, speech input, or tactile input.
The systems and techniques described here can be implemented in a computing system that includes a background component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such background, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), wide Area Networks (WANs), and the internet.
The computer system may include a client and a server. The client and server are typically remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server may be a cloud server, a server of a distributed system, or a server incorporating a blockchain.
It should be appreciated that various forms of the flows shown above may be used to reorder, add, or delete steps. For example, the steps recited in the present disclosure may be performed in parallel or sequentially or in a different order, provided that the desired results of the technical solutions of the present disclosure are achieved, and are not limited herein.
The above detailed description should not be taken as limiting the scope of the present disclosure. It will be apparent to those skilled in the art that various modifications, combinations, sub-combinations and alternatives are possible, depending on design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present disclosure are intended to be included within the scope of the present disclosure.
Claims (41)
1. A recommendation method, comprising:
determining an interest activated historical interaction resource sequence of a target object, wherein the interest activated historical interaction resource sequence comprises a plurality of historical interaction resources interacted by the target object in a historical period, and the plurality of historical interaction resources belong to resources of the target interest;
determining object features of the target object with respect to the target interest based on the historical interaction resource sequence of interest activation and object attribute information of the target object;
Determining a plurality of candidate resource features related to the target interest, wherein the plurality of candidate resource features are in one-to-one correspondence with a plurality of candidate resources, the candidate resource features are used for characterizing resource attribute features of the candidate resources corresponding to the candidate resource features, and the plurality of candidate resources belong to the resource of the target interest; and
a resource to be recommended is determined from the plurality of candidate resources based on the plurality of candidate resource features and the object feature.
2. The method of claim 1, wherein the determining a historical interaction resource sequence of interest activation for a target object comprises:
and determining a historical interaction resource sequence of interest activation of the target object from an offline database based on the identification information of the target object and the interest information of the target interest, wherein the offline database is generated by utilizing a plurality of historical interaction resource sequences of interest activation.
3. The method of claim 1 or 2, further comprising:
acquiring a historical interaction resource sequence of the target object in the historical period; and
and according to the mapping relation between the interests and the resources, performing interest activation on the historical interaction resource sequence to obtain the historical interaction resource sequence activated by the interests.
4. The method of claim 3, wherein the performing interest activation on the historical interaction resource sequence according to the mapping relationship between interest and resource to obtain the interest activated historical interaction resource sequence comprises:
according to the mapping relation between interests and resources, according to preset activation weights, performing interest activation on each historical interaction resource in the historical interaction resource sequence with respect to the target interests to obtain the historical interaction resource sequence with the interest activation.
5. The method of claim 3 or 4, further comprising:
and determining a mapping relation between the interests and the resources based on the interest characteristics and the historical interaction resource characteristics, wherein the historical interaction resource characteristics are used for representing resource attribute characteristics of the historical interaction resources, and the interest characteristics are used for representing characteristics related to interest information.
6. The method of claim 5, further comprising:
aiming at each object interaction resource in a plurality of object interaction resources, obtaining object interaction resource characteristics based on resource attribute information of the object interaction resources, wherein the object interaction resources are used for representing history interaction resources interacted by objects;
Clustering the object interaction resource characteristics to obtain at least one object interaction resource characteristic cluster, wherein the at least one object interaction resource characteristic cluster corresponds to at least one interest one by one; and
and obtaining interest features of interests corresponding to the object interaction resource feature clusters based on each object interaction resource feature cluster.
7. The method of any of claims 1 to 6, wherein the determining object features of the target object with respect to the target interest based on the historical interaction resource sequence of interest activation and object attribute information of the target object comprises:
carrying out serialization characterization on the historical interaction resource sequence of interest activation to obtain the feature of the interest activation interaction resource; and
and activating interactive resource features based on object attribute information of the target object and the interests to obtain the object features of the target object about the target interests.
8. The method of any of claims 1 to 7, wherein the determining a plurality of candidate resource features related to the target interest comprises:
determining a target resource index corresponding to the target interest from a plurality of resource indexes; and
Based on the target resource index, the plurality of candidate resource features that match the interest feature of the target interest are determined.
9. The method of claim 8, further comprising:
determining a resource feature set for each of the plurality of interests based on respective resource attribute information of the plurality of resources; and
and generating a plurality of resource indexes based on a plurality of the resource feature sets, wherein the plurality of resource indexes are in one-to-one correspondence with the plurality of interests.
10. The method of any of claims 1-9, wherein the determining a resource to be recommended from the plurality of candidate resources based on the plurality of candidate resource characteristics and the object characteristic comprises:
determining the similarity between the object feature and each candidate resource feature to obtain a plurality of similarities; and
and determining the resources to be recommended from the plurality of candidate resources based on the plurality of similarities.
11. The method of claim 10, wherein the resource to be recommended comprises a plurality of;
the method further comprises the steps of:
sorting the plurality of resources to be recommended according to the similarity to obtain a sorting result; and
And recommending a plurality of resources to be recommended according to the sequencing result.
12. The method of any one of claims 1 to 11, further comprising:
and acquiring a historical interaction resource sequence of the target object from the starting moment to the current moment, wherein the starting moment is used for representing the historical registration moment of the target object.
13. A training method of a deep learning model, comprising:
determining an interest activated sample historical interaction resource sequence of a sample target object, wherein the interest activated sample historical interaction resource sequence comprises a plurality of sample historical interaction resources interacted by the sample target object in a sample historical period, and the plurality of sample historical interaction resources belong to sample target interest resources;
determining sample object features of the sample target object with respect to the sample target interest based on the sample historical interaction resource sequence of interest activation and sample object attribute information of the sample target object;
determining a target sample historical interaction resource from the sample historical interaction resource sequence of interest activation of the sample target object;
obtaining sample resource characteristics based on sample resource attribute information of the target sample history interaction resource; and
And adjusting parameters of the deep learning model based on the sample object characteristics and the sample resource characteristics to complete training.
14. The method of claim 13, wherein the determining the sample historical interaction resource sequence of interest activations for the sample target object comprises:
and determining a sample historical interaction resource sequence activated by the interest of the sample target object from a sample offline database based on the sample identification information of the sample target object and the sample interest information of the sample target interest, wherein the sample offline database is generated by using a plurality of sample historical interaction resource sequences activated by the interest.
15. The method of claim 13 or 14, further comprising:
acquiring a sample history interaction resource sequence of the sample target object in the sample history period; and
and according to the mapping relation between the sample interest and the resource, performing interest activation on the sample historical interaction resource sequence to obtain the sample historical interaction resource sequence activated by the interest.
16. The method of claim 15, wherein the performing interest activation on the sample historical interaction resource sequence according to the mapping relationship between the sample interest and the resource to obtain the sample historical interaction resource sequence with the interest activated comprises:
And according to the mapping relation between the sample interests and the resources, presetting an activation weight according to a sample, and performing interest activation on each sample history interaction resource in the sample history interaction resource sequence with respect to the sample target interests to obtain a sample history interaction resource sequence activated by the interests.
17. The method of claim 15 or 16, further comprising:
and determining a mapping relation between the sample interests and the resources based on sample interest features and sample history interaction resource features, wherein the sample history interaction resource features are used for representing sample resource attribute features of the sample history interaction resources, and the sample interest features are used for representing features related to sample interest information.
18. The method of claim 17, further comprising:
for each sample object interaction resource in a plurality of sample object interaction resources, obtaining sample object interaction resource characteristics based on sample resource attribute information of the sample object interaction resource, wherein the sample object interaction resource is used for representing sample history interaction resources interacted by sample objects;
clustering the interaction resource characteristics of the plurality of sample objects to obtain at least one sample object interaction resource characteristic cluster, wherein the at least one sample object interaction resource characteristic cluster corresponds to at least one sample interest one by one;
And obtaining sample interest features of sample interests corresponding to the sample object interaction resource feature clusters based on each sample object interaction resource feature cluster.
19. The method of any of claims 13 to 18, wherein the determining sample object features of the sample target object for the sample target object interest based on the sample historical interaction resource sequence of interest activation and sample object attribute information of the sample target object comprises:
carrying out serialization characterization on the sample history interaction resource sequence activated by the interest to obtain sample interest activated interaction resource characteristics; and
and activating an interactive resource feature based on sample object attribute information of the sample target object and the sample interest to obtain the sample object feature of the sample target object about the sample target interest.
20. A recommendation device, comprising:
the resource sequence determining module is used for determining a historical interaction resource sequence of interest activation of a target object, wherein the historical interaction resource sequence of interest activation comprises a plurality of historical interaction resources interacted by the target object in a historical period, and the plurality of historical interaction resources belong to resources of the target interest;
An object feature determining module, configured to determine an object feature of the target object about the target interest based on the historical interaction resource sequence of interest activation and object attribute information of the target object;
a candidate resource feature determining module, configured to determine a plurality of candidate resource features related to the target interest, where the plurality of candidate resource features are in one-to-one correspondence with a plurality of candidate resources, and the candidate resource features are used to characterize resource attribute features of the candidate resources corresponding to the candidate resource features, and the plurality of candidate resources all belong to the resource of the target interest;
and the resource to be recommended determining module is used for determining resources to be recommended from the plurality of candidate resources based on the plurality of candidate resource characteristics and the object characteristics.
21. The apparatus of claim 20, wherein the resource sequence determination module comprises:
and the resource sequence determining unit is used for determining a historical interaction resource sequence activated by the interest of the target object from an offline database based on the identification information of the target object and the interest information of the target interest, wherein the offline database is generated by utilizing a plurality of historical interaction resource sequences activated by the interest.
22. The apparatus of claim 20 or 21, further comprising:
the resource sequence acquisition module is used for acquiring a historical interaction resource sequence of the target object in the historical period;
and the interest activation module is used for carrying out interest activation on the historical interaction resource sequence according to the mapping relation between the interest and the resource to obtain the historical interaction resource sequence activated by the interest.
23. The apparatus of claim 22, wherein the interest activation module comprises:
and the interest activation unit is used for carrying out interest activation on each historical interaction resource in the historical interaction resource sequence according to the mapping relation between the interest and the resource and the preset activation weight to obtain the historical interaction resource sequence activated by the interest.
24. The apparatus of claim 22 or 23, further comprising:
the mapping relation determining module is used for determining the mapping relation between the interests and the resources based on the interest characteristics and the historical interaction resource characteristics, wherein the historical interaction resource characteristics are used for representing resource attribute characteristics of the historical interaction resources, and the interest characteristics are used for representing characteristics related to interest information.
25. The apparatus of claim 24, further comprising:
the interactive resource feature determining module is used for obtaining object interactive resource features according to the resource attribute information of each object interactive resource in the plurality of object interactive resources, wherein the object interactive resources are used for representing history interactive resources interacted by the object;
the interactive resource feature clustering module is used for clustering the object interactive resource features to obtain at least one object interactive resource feature cluster, wherein the at least one object interactive resource feature cluster corresponds to at least one interest one by one;
and the interest feature determining module is used for obtaining interest features of the interests corresponding to the object interaction resource feature clusters based on each object interaction resource feature cluster.
26. The apparatus of any of claims 20 to 25, wherein the object feature determination module comprises:
the interactive resource sequence characterization unit is used for carrying out serialization characterization on the historical interactive resource sequence activated by the interest to obtain the interactive resource characteristics activated by the interest;
and the object feature determining unit is used for activating interactive resource features based on the object attribute information of the target object and the interests to obtain the object features of the target object about the target interests.
27. The apparatus of any of claims 20 to 26, wherein the candidate resource feature determination module comprises:
a target resource index determining unit configured to determine a target resource index corresponding to the target interest from a plurality of resource indexes;
and the candidate resource feature determining unit is used for determining the plurality of candidate resource features matched with the interest feature of the target interest based on the target resource index.
28. The apparatus of claim 27, further comprising:
the resource feature set determining module is used for determining a resource feature set based on the respective resource attribute information of the plurality of resources aiming at each interest in the plurality of interests;
and the resource index generation module is used for generating a plurality of resource indexes based on a plurality of the resource feature sets, wherein the plurality of resource indexes are in one-to-one correspondence with the plurality of interests.
29. The apparatus of any of claims 20 to 28, wherein the resource to be recommended determination module comprises:
a similarity determining unit, configured to determine a similarity between the object feature and each candidate resource feature, so as to obtain a plurality of similarities;
and the resource to be recommended determining unit is used for determining the resource to be recommended from the plurality of candidate resources based on the plurality of similarities.
30. The apparatus of claim 29, wherein the resource to be recommended comprises a plurality of;
the apparatus further comprises:
the sorting module is used for sorting the plurality of resources to be recommended according to the similarity to obtain a sorting result;
and the recommending module is used for recommending a plurality of resources to be recommended according to the sequencing result.
31. The apparatus of any one of claims 20 to 30, further comprising:
the resource sequence acquisition module is used for acquiring a historical interaction resource sequence of the target object from the starting moment to the current moment, wherein the starting moment is used for representing the historical registration moment of the target object.
32. A training device for a deep learning model, comprising:
the sample resource sequence determining module is used for determining a sample historical interaction resource sequence of interest activation of a sample target object, wherein the sample historical interaction resource sequence of interest activation comprises a plurality of sample historical interaction resources interacted by the sample target object in a sample historical period, and the plurality of sample historical interaction resources belong to resources of sample target interest;
a sample object feature determination module for determining sample object features of the sample target object with respect to the sample target interest based on the interest-activated sample historical interaction resource sequence and sample object attribute information of the sample target object;
A sample interaction resource determining module, configured to determine a target sample history interaction resource from the sample history interaction resource sequence activated by the interest of the sample target object;
the sample resource feature determining module is used for obtaining sample resource features based on sample resource attribute information of the target sample history interaction resources;
and the parameter adjustment module is used for adjusting parameters of the deep learning model based on the sample object characteristics and the sample resource characteristics to complete training.
33. The apparatus of claim 32, wherein the sample resource sequence determination module comprises:
and the sample resource sequence determining unit is used for determining a sample historical interaction resource sequence activated by the interest of the sample target object from a sample offline database based on the sample identification information of the sample target object and the sample interest information of the sample target interest, wherein the sample offline database is generated by using a plurality of sample historical interaction resource sequences activated by the interest.
34. The apparatus of claim 32 or 33, further comprising:
a sample resource sequence acquisition module, configured to acquire a sample history interaction resource sequence of the sample target object in the sample history period;
And the sample interest activation module is used for carrying out interest activation on the sample historical interaction resource sequence according to the mapping relation between the sample interest and the resource to obtain the sample historical interaction resource sequence activated by the interest.
35. The apparatus of claim 34, wherein the sample interest activation module comprises:
and the sample interest activation unit is used for carrying out interest activation on the sample target interest on each sample history interaction resource in the sample history interaction resource sequence according to the mapping relation between the sample interest and the resource and the preset activation weight of the sample, so as to obtain the sample history interaction resource sequence activated by the interest.
36. The apparatus of claim 34 or 35, further comprising:
the sample mapping relation determining module is used for determining a mapping relation between the sample interests and the resources based on sample interest features and sample history interaction resource features, wherein the sample history interaction resource features are used for representing sample resource attribute features of the sample history interaction resources, and the sample interest features are used for representing features related to sample interest information.
37. The apparatus of claim 36, further comprising:
The sample interaction resource feature determining module is used for obtaining sample object interaction resource features according to sample resource attribute information of each sample object interaction resource in the plurality of sample object interaction resources, wherein the sample object interaction resources are used for representing sample history interaction resources interacted by sample objects;
the sample interaction resource feature clustering module is used for clustering the interaction resource features of the plurality of sample objects to obtain at least one sample object interaction resource feature cluster, wherein the at least one sample object interaction resource feature cluster corresponds to at least one sample interest one by one;
and the sample interest feature determining module is used for obtaining sample interest features of sample interests corresponding to the sample object interaction resource feature clusters based on each sample object interaction resource feature cluster.
38. The apparatus of any of claims 32 to 37, wherein the sample object feature determination module comprises:
the sample interactive resource sequence characterization unit is used for carrying out serialization characterization on the sample historical interactive resource sequence activated by the interest to obtain sample interest activated interactive resource characteristics;
And the sample object feature determining unit is used for activating interactive resource features based on sample object attribute information of the sample target object and the sample interest to obtain the sample object features of the sample target object about the sample target interest.
39. An electronic device, comprising:
at least one processor; and
a memory communicatively coupled to the at least one processor; wherein,
the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 19.
40. A non-transitory computer readable storage medium storing computer instructions for causing the computer to perform the method of any one of claims 1 to 19.
41. A computer program product comprising a computer program which, when executed by a processor, implements the method according to any one of claims 1 to 19.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311334820.XA CN117668351A (en) | 2023-10-16 | 2023-10-16 | Recommendation method, training method and device of model, electronic equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311334820.XA CN117668351A (en) | 2023-10-16 | 2023-10-16 | Recommendation method, training method and device of model, electronic equipment and storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117668351A true CN117668351A (en) | 2024-03-08 |
Family
ID=90070203
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311334820.XA Pending CN117668351A (en) | 2023-10-16 | 2023-10-16 | Recommendation method, training method and device of model, electronic equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117668351A (en) |
-
2023
- 2023-10-16 CN CN202311334820.XA patent/CN117668351A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109460514B (en) | Method and device for pushing information | |
| WO2022041979A1 (en) | Information recommendation model training method and related device | |
| CN110909182B (en) | Multimedia resource searching method, device, computer equipment and storage medium | |
| US11875241B2 (en) | Aspect pre-selection using machine learning | |
| CN114036398B (en) | Content recommendation and ranking model training method, device, equipment and storage medium | |
| CN113360711B (en) | Model training and executing method, device, equipment and medium for video understanding task | |
| CN110598084A (en) | Object sorting method, commodity sorting device and electronic equipment | |
| CN114692778B (en) | Multi-mode sample set generation method, training method and device for intelligent inspection | |
| CN112000763A (en) | Method, device, equipment and medium for determining competition relationship of interest points | |
| CN110992127B (en) | Article recommendation method and device | |
| CN113592593A (en) | Training and application method, device, equipment and storage medium of sequence recommendation model | |
| US20210357553A1 (en) | Apparatus and method for option data object performance prediction and modeling | |
| CN114037059A (en) | Pre-training model, model generation method, data processing method and data processing device | |
| WO2020211616A1 (en) | Method and device for processing user interaction information | |
| CN116228301A (en) | Method, device, equipment and medium for determining target user | |
| CN115618079A (en) | Session recommendation method, device, electronic equipment and storage medium | |
| CN112966108B (en) | Method, apparatus, device and storage medium for detecting data and training classification model | |
| CN117668351A (en) | Recommendation method, training method and device of model, electronic equipment and storage medium | |
| CN113888216A (en) | Advertisement information pushing method and device, electronic equipment and storage medium | |
| CN112905885A (en) | Method, apparatus, device, medium, and program product for recommending resources to a user | |
| CN111860870A (en) | Training method, device, equipment and medium for interactive behavior determination model | |
| CN113704596A (en) | Method and apparatus for generating a set of recall information | |
| CN117076962B (en) | Data analysis method, device and equipment applied to artificial intelligence field | |
| CN117131197B (en) | Method, device, equipment and storage medium for processing demand category of bidding document | |
| CN116342183A (en) | Information recommendation method, training device, electronic equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |