CN112883265A - Information recommendation method and device, server and computer readable storage medium - Google Patents
Information recommendation method and device, server and computer readable storage medium Download PDFInfo
- Publication number
- CN112883265A CN112883265A CN202110185529.5A CN202110185529A CN112883265A CN 112883265 A CN112883265 A CN 112883265A CN 202110185529 A CN202110185529 A CN 202110185529A CN 112883265 A CN112883265 A CN 112883265A
- Authority
- CN
- China
- Prior art keywords
- information
- training
- pieces
- sample information
- piece
- 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.)
- Withdrawn
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The disclosure provides an information recommendation method, an information recommendation device, a server and a storage medium, and belongs to the technical field of internet. The method comprises the following steps: inputting the target user characteristics, the information characteristics of each candidate information and the cross characteristics into a rough ranking model, outputting the rough ranking score of each candidate information, and training the rough ranking model based on the training result of the fine ranking model to obtain the target user characteristics, the information characteristics of each candidate information and the cross characteristics; screening out a plurality of pieces of first information from the plurality of pieces of candidate information based on the rough ranking score of each piece of candidate information; screening out a plurality of pieces of second information from the plurality of pieces of first information based on the fine ranking model; and recommending the plurality of pieces of second information to the target user. The rough ranking model is obtained based on training results of the fine ranking model, so that a ranking mechanism of the fine ranking model can be learned, and when the rough ranking model is applied online, candidate information is ranked based on the same ranking mechanism, so that information meeting user requirements can be screened out, and accuracy of recommended information is improved.
Description
Technical Field
The present disclosure relates to the field of internet technologies, and in particular, to an information recommendation method, an information recommendation apparatus, a server, and a computer-readable storage medium.
Background
In the technical field of internet, information recommendation comprises three stages, namely recall, rough typesetting and fine typesetting. The recalling refers to screening out information which is possibly interested by the user from the mass information; the rough arrangement refers to the rough arrangement model trained by adopting fewer features, and the recalled information is sequenced, so that the recalled information is narrowed to a reasonable range; the fine ranking is to sort the screened information by adopting a fine ranking model trained by more characteristics, so that the screened information is reduced to an expected number. In the face of mass information on the Internet, the coverage and accuracy of the recommended information are ensured by performing rough arrangement and fine arrangement on the recalled information.
In the related technology, the recalled information is ranked based on a rough model, a first quantity of pieces of information are screened out from the information, and the rough model is obtained based on a plurality of training sample information and a first optimization target training; sequencing the first quantity of information based on a fine ranking model, and screening out second quantity of information, wherein the fine ranking model is obtained based on a plurality of training sample information and a second optimization target training; and recommending the second quantity of information to the user.
However, the coarse-ranking model and the fine-ranking model are obtained based on different optimization target training, so that the ranking mechanisms of the fine-ranking model and the coarse-ranking model are different, and the coarse-ranking model and the fine-ranking model cannot be ranked according to a uniform standard, so that the recommended information is inaccurate.
Disclosure of Invention
The embodiment of the disclosure provides an information recommendation method, an information recommendation device, a server and a computer readable storage medium, which can improve the accuracy of recommended information. The technical scheme is as follows:
in a first aspect, an information recommendation method is provided, where the method includes:
acquiring a plurality of pieces of candidate information;
extracting target user characteristics of a target user, and extracting information characteristics of each piece of candidate information and cross characteristics of each piece of candidate information and the target user;
inputting the target user characteristics, the information characteristics of each candidate information and the cross characteristics into a rough ranking model, and outputting a rough ranking score of each candidate information, wherein the rough ranking model is obtained by training based on a training result of a fine ranking model;
screening out a plurality of pieces of first information from the plurality of pieces of candidate information based on the rough ranking score of each piece of candidate information;
screening out a plurality of pieces of second information from the plurality of pieces of first information based on the fine ranking model;
recommending the second information to the target user.
In another embodiment of the present disclosure, the training process of the coarse row model is as follows:
based on the fine ranking model, acquiring a plurality of pieces of coarse ranking sample information from a plurality of pieces of training sample information used for training the fine ranking model;
extracting training information characteristics of each piece of coarse-line sample information, user characteristics of a user clicking each piece of coarse-line sample information and training cross characteristics between each user and each piece of coarse-line sample information;
training an initial rough-layout model based on the training information characteristics of each piece of rough-layout sample information, the user characteristics of the user clicking each piece of rough-layout sample information and the training cross characteristics between each user and each piece of rough-layout sample information to obtain the rough-layout model.
In another embodiment of the present disclosure, the obtaining, based on the fine line model, a plurality of pieces of rough line sample information from a plurality of pieces of training sample information used for training the fine line model includes:
acquiring training sample information with a preset number of ranks in front based on the fine ranking scores of the plurality of pieces of training sample information by the fine ranking model;
determining the training sample information with the front preset digit as a positive sample, and determining the rest training sample information as a negative sample;
and composing the positive sample and the negative sample into the plurality of pieces of coarse-sized sample information.
In another embodiment of the present disclosure, the obtaining, based on the fine line model, a plurality of pieces of rough line sample information from a plurality of pieces of training sample information used for training the fine line model includes:
acquiring training sample information with a preset number of ranks in front based on the fine ranking scores of the plurality of pieces of training sample information by the fine ranking model;
and determining the training sample information with the preset digit and at least one piece of training sample information which is from the same query request and has the rank behind the preset digit as the plurality of pieces of coarse training sample information.
In another embodiment of the present disclosure, the obtaining, based on the fine line model, a plurality of pieces of rough line sample information from a plurality of pieces of training sample information used for training the fine line model includes:
estimating the click rate and the conversion rate of each piece of training sample information based on the rough model;
determining the ranking score of each piece of training sample information according to the click rate and the conversion rate of each piece of training sample information;
and acquiring a plurality of pieces of rough training sample information from the plurality of pieces of training sample information based on the ranking score of each piece of training sample information.
In a second aspect, an information recommendation apparatus is provided, the apparatus comprising:
the first acquisition module is used for acquiring a plurality of pieces of candidate information;
the first extraction module is used for extracting the target user characteristics of the target user;
the second extraction module is also used for extracting the information characteristic of each piece of candidate information and the cross characteristic of each piece of candidate information and the target user;
the input and output module is used for inputting the target user characteristics, the information characteristics of each candidate message and the cross characteristics into a rough ranking model, and outputting a rough ranking score of each candidate message, wherein the rough ranking model is obtained by training based on a training result of a fine ranking model;
the first screening module is used for screening out a plurality of pieces of first information from the plurality of pieces of candidate information based on the rough ranking score of each piece of candidate information;
the second screening module is further used for screening out a plurality of pieces of second information from the plurality of pieces of first information based on the fine ranking model;
and the recommending module is used for recommending the second information to the target user.
In another embodiment of the present disclosure, the apparatus for the course row model includes:
the second obtaining module is used for obtaining a plurality of pieces of rough-arranged sample information from a plurality of pieces of training sample information used for training the fine-arranged model based on the fine-arranged model;
the third extraction module is used for extracting the training information characteristics of each piece of coarse-line sample information, the user characteristics of the user clicking each piece of coarse-line sample information and the training cross characteristics between each user and each piece of coarse-line sample information;
and the training module is used for training the initial coarse layout model based on the training information characteristics of each piece of coarse layout sample information, the user characteristics of the user clicking each piece of coarse layout sample information and the training cross characteristics between each user and each piece of coarse layout sample information to obtain the coarse layout model.
In another embodiment of the present disclosure, the second obtaining module is further configured to obtain training sample information with a preset number of ranks in front based on the ranking scores of the plurality of pieces of training sample information by the ranking model; determining the training sample information with the front preset digit as a positive sample, and determining the rest training sample information as a negative sample; and composing the positive sample and the negative sample into the plurality of pieces of coarse-sized sample information.
In another embodiment of the present disclosure, the second obtaining module is further configured to obtain training sample information with a preset number of ranks in front based on the ranking scores of the plurality of pieces of training sample information by the ranking model; and determining the training sample information with the preset digit and at least one piece of training sample information which is from the same query request and has the rank behind the preset digit as the plurality of pieces of coarse training sample information.
In another embodiment of the present disclosure, the second obtaining module is further configured to estimate a click rate and a conversion rate of each piece of training sample information based on the rough model; determining the ranking score of each piece of training sample information according to the click rate and the conversion rate of each piece of training sample information; and acquiring a plurality of pieces of rough training sample information from the plurality of pieces of training sample information based on the ranking score of each piece of training sample information.
In a third aspect, a server is provided, which includes a processor and a memory, where at least one program code is stored in the memory, and the at least one program code is loaded and executed by the processor to implement the information recommendation method according to the first aspect.
In a fourth aspect, a computer-readable storage medium is provided, in which at least one program code is stored, and the at least one program code is loaded and executed by a processor to implement the information recommendation method according to the first aspect.
The technical scheme provided by the embodiment of the disclosure has the following beneficial effects:
the rough ranking model is obtained by training based on the training result of the fine ranking model, so that the ranking mechanism of the fine ranking model can be learned, and when the method is applied on line, the candidate information is ranked based on the same ranking mechanism, so that the information meeting the user requirements can be screened out. Meanwhile, when information screening is carried out based on the rough ranking model and the fine ranking model, not only the information characteristics of the candidate information and the user characteristics of the target user are considered, but also the cross characteristics between the candidate information and the target user are considered, and the target user and the candidate information are comprehensively depicted, so that the accuracy of the recommended information is greatly improved when information recommendation is carried out.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present disclosure, and it is obvious for those skilled in the art to obtain other drawings based on the drawings without creative efforts.
Fig. 1 is a schematic diagram of an implementation environment related to an information recommendation method provided by an embodiment of the present disclosure;
fig. 2 is an overall architecture diagram of an information recommendation method provided by an embodiment of the present disclosure;
FIG. 3 is an architecture diagram of a coarse row model provided by an embodiment of the present disclosure;
fig. 4 is a flowchart of an information recommendation method provided by an embodiment of the present disclosure;
fig. 5 is a flowchart of an information recommendation method provided by an embodiment of the present disclosure;
FIG. 6 is a flowchart of a training method of a steak model provided by an embodiment of the present disclosure;
fig. 7 is a schematic structural diagram of an information recommendation device provided in an embodiment of the present disclosure;
fig. 8 shows a block diagram of a server according to an exemplary embodiment of the present disclosure.
Detailed Description
To make the objects, technical solutions and advantages of the present disclosure more apparent, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings.
It is to be understood that the terms "each," "a plurality," and "any" and the like, as used in the embodiments of the present disclosure, are intended to encompass two or more, each referring to each of the corresponding plurality, and any referring to any one of the corresponding plurality. For example, the plurality of words includes 10 words, and each word refers to each of the 10 words, and any word refers to any one of the 10 words.
Referring to fig. 1, an implementation environment related to an information recommendation method provided by an embodiment of the present disclosure is shown, where the implementation environment includes: a terminal 101 and a server 102.
The terminal 101 is installed with at least one application, such as a shopping application, a news application, and a search application, and based on the installed application, the terminal 101 can initiate an information query request to the server and can also receive information pushed by the server. The terminal 101 may be a smart phone, a tablet computer, a notebook computer, a desktop computer, or other devices, and the product type of the terminal 101 is not specifically limited in the embodiments of the present disclosure.
The server 102 is a background server of an application installed in the terminal 101, and the server 102 may be an independent physical server, or a server cluster or distributed system formed by a plurality of physical servers. The server has strong computing power, can train a fine model based on a plurality of pieces of training sample information, determines a plurality of pieces of coarse sample information for training a coarse model from the plurality of pieces of training sample information based on a training result of the fine model, and trains the coarse model according to the plurality of pieces of coarse sample information; the server can also acquire information to be recommended to the user from the recalled candidate information based on the trained rough ranking model and the trained fine ranking model, and recommend the information to the user.
The terminal 101 and the server 102 may be directly or indirectly connected through wired or wireless communication, which is not limited in the embodiment of the present disclosure.
Fig. 2 is an architecture diagram of an information recommendation process provided by an embodiment of the present disclosure, and referring to fig. 2, the information recommendation process includes an offline service process and an online service process.
The offline service process is mainly used for training the fine model and the rough model. Specifically, training sample information is obtained, a refined model is trained on the basis of the obtained training sample information, the training sample information is ranked on the basis of the trained refined model, and then, on the basis of a ranking result, rough-arranged sample information is obtained from the training sample information used for training the rough-arranged model, and then, the rough-arranged model is trained on the basis of the rough-arranged sample information.
And in the online service process, information needing to be recommended to the user is obtained from the recalled candidate information mainly based on the trained rough ranking model and the refined ranking model.
In the Field of information recommendation, models commonly used for the rough Model include LR (Logistic Regression), DSSM (Deep Semantic Model), FFM (Field-aware Factorization Machine), and the like. For the architecture diagram of the coarse layout model, taking fig. 3 as an example for explanation, referring to fig. 3, the coarse layout model adopts Wide & Deep framework, Deep part adopts double tower architecture, which is the user side and the information side respectively, and Wide part includes the cross features of the user and the information. The following describes the rough model in terms of two aspects of effect and performance, respectively.
The effect is that: the user side of the rough model takes into account the basic characteristics of the user, the contextual characteristics of the request (real-time scene information), and the historical behavior sequence information of the user. The information contains information on each dimension of the user, and the interest of the user can be well described by combining the behavior sequence information and the attention model; the information side mainly describes various information of the information. And calculating the cosine similarity of the vector of the user side and the vector of the information side to obtain the fraction of the double-tower part. Meanwhile, considering that the double-tower model does not consider the intersection between the user and the information, the drawing is mainly performed from a single side, in order to make up for the defect, a Wide part is added in the embodiment of the disclosure, and the Wide part adopts the intersection characteristics of two sides to further describe the user and the information so as to assist the model in improving the effect.
Performance aspects: the deep learning model has huge parameter quantity and long time consumption in the calculation process, in order to shorten the calculation time, the vector of the information side in the double-tower part can be calculated off line and updated and loaded into resources by taking days as a unit, and only the off-line calculation result needs to be inquired when the double-tower part is used on line. The vectors of the user side in the double-tower part need to be calculated on line in real time, and the Wide part needs to be calculated on line in real time. For example, when a query request of a user is received, calculation is performed once on a user side sending the request, although an information side contains a large amount of candidate information, the information does not need online real-time calculation and only needs to be loaded from resources, and a Wide part adopts an LR model, which has a higher calculation speed, so that the calculation time of the rough model is greatly shortened.
In summary, the coarse row model provided by the embodiment of the present disclosure can be used to greatly improve the performance and effect of the model.
Based on the implementation environment shown in fig. 1, an embodiment of the present disclosure provides an information recommendation method, and referring to fig. 4, a flow of the method provided by the embodiment of the present disclosure includes:
401. a plurality of pieces of candidate information are acquired.
402. And extracting the target user characteristics of the target user, and extracting the information characteristics of each piece of candidate information and the cross characteristics of each piece of candidate information and the target user.
403. And inputting the target user characteristics, the information characteristics of each candidate information and the cross characteristics into a rough ranking model, and outputting a rough ranking score of each candidate information.
And the coarse ranking model is obtained by training based on the training result of the fine ranking model.
404. And screening out a plurality of pieces of first information from the plurality of pieces of candidate information based on the rough ranking score of each piece of candidate information.
405. And screening out a plurality of pieces of second information from the plurality of pieces of first information based on the fine ranking model.
406. And recommending the plurality of pieces of second information to the target user.
According to the method provided by the embodiment of the disclosure, because the rough ranking model is obtained by training based on the training result of the fine ranking model, the ranking mechanism of the fine ranking model can be learned, and when the method is applied online, the candidate information is ranked based on the same ranking mechanism, so that the information meeting the user requirements can be screened out. Meanwhile, when information screening is carried out based on the rough ranking model and the fine ranking model, not only the information characteristics of the candidate information and the user characteristics of the target user are considered, but also the cross characteristics between the candidate information and the target user are considered, and the target user and the candidate information are comprehensively depicted, so that the accuracy of the recommended information is greatly improved when information recommendation is carried out.
In another embodiment of the present disclosure, the training process of the coarse-row model is:
based on the fine ranking model, acquiring a plurality of pieces of coarse ranking sample information from a plurality of pieces of training sample information for training the fine ranking model;
extracting training information characteristics of each piece of coarse-line sample information, user characteristics of a user clicking each piece of coarse-line sample information and training cross characteristics between each user and each piece of coarse-line sample information;
and training the initial rough-layout model based on the training information characteristics of each piece of rough-layout sample information, the user characteristics of the user clicking each piece of rough-layout sample information and the training cross characteristics between each user and each piece of rough-layout sample information to obtain a rough-layout model.
In another embodiment of the present disclosure, obtaining, based on the fine rule model, a plurality of pieces of coarse rule sample information from a plurality of pieces of training sample information used for training the fine rule model includes:
acquiring training sample information with a preset number of ranks in front based on the fine ranking scores of the fine ranking model to the pieces of training sample information;
determining training sample information with a preset number of bits in the front as a positive sample, and determining the rest training sample information as a negative sample;
and composing the positive sample and the negative sample into a plurality of pieces of coarse-row sample information.
In another embodiment of the present disclosure, obtaining, based on the fine rule model, a plurality of pieces of coarse rule sample information from a plurality of pieces of training sample information used for training the fine rule model includes:
acquiring training sample information with a preset number of ranks in front based on the fine ranking scores of the fine ranking model to the pieces of training sample information;
and determining the training sample information with the preset digit in the front and at least one piece of training sample information which is from the same query request and has the rank behind the preset digit as a plurality of pieces of coarse training sample information.
In another embodiment of the present disclosure, obtaining, based on the fine rule model, a plurality of pieces of coarse rule sample information from a plurality of pieces of training sample information used for training the fine rule model includes:
estimating the click rate and the conversion rate of each piece of training sample information based on the rough model;
determining the ranking score of each piece of training sample information according to the click rate and the conversion rate of each piece of training sample information;
and acquiring a plurality of pieces of coarse training sample information from the plurality of pieces of training sample information based on the ranking score of each piece of training sample information.
All the above optional technical solutions may be combined arbitrarily to form the optional embodiments of the present disclosure, and are not described herein again.
Based on the implementation environment shown in fig. 1, an embodiment of the present disclosure provides an information recommendation method, taking a server to execute the embodiment of the present disclosure as an example, referring to fig. 5, a flow of the method provided by the embodiment of the present disclosure includes:
501. the server acquires a plurality of pieces of candidate information.
When receiving an information query request sent by a target user, the server recalls information which is possibly interested by the target user from mass information on the Internet by adopting an information recall algorithm based on the information query request, wherein the recalled information is candidate information. The information recall algorithm includes a base algorithm, a modified algorithm, a ranki2i algorithm, and the like, and the embodiment of the present disclosure does not specifically limit the information recall algorithm.
502. And the server extracts the target user characteristics of the target user, and extracts the information characteristics of each piece of candidate information and the cross characteristics of each piece of candidate information and the target user.
The target user characteristics comprise behavior characteristics of the target user and attribute characteristics of the target user. The behavior characteristics comprise at least one of click behavior characteristics, collection behavior characteristics, viewing behavior characteristics and the like. The attribute characteristics comprise at least one of age, gender, native place, attribution and the like.
The information characteristics are determined according to the content of the candidate information, for example, the information characteristics of the gourmet candidate information include merchant address, price, sales volume, evaluation and the like.
The cross feature does not belong to the target user feature and the information feature and is a feature associated between the candidate information and the target user. For example, in a scenario of food recommendation in a take-away application, the cross-feature may be a distance between the merchant and the target user, and the like.
503. And the server inputs the target user characteristics, the information characteristics of each piece of candidate information and the cross characteristics into the rough ranking model and outputs the rough ranking score of each piece of candidate information.
And based on the acquired target user characteristics, the information characteristics of each candidate information and the cross characteristics of each candidate information and the target user, the server inputs the target user characteristics, the information characteristics of each candidate information and the cross characteristics of each candidate information and the target user into a rough ranking model, and outputs a rough ranking score of each candidate information. The rough model is obtained by training based on the training result of the fine model, so that the rough model learns the sorting mechanism of the fine model, information meeting requirements can be screened out for a target user based on the same sorting mechanism, and the accuracy of recommended information is greatly improved. For the training process of the rough model, see the following steps 601-603, which will not be described in detail here.
504. And based on the rough ranking score of each piece of candidate information, the server screens out a plurality of pieces of first information from the plurality of pieces of candidate information.
The server sorts the plurality of pieces of candidate information according to the rough ranking score of each piece of candidate information and the sequence of the rough ranking scores from high to low, and then screens out a plurality of pieces of first information from the plurality of pieces of candidate information based on the sorting result.
505. The server inputs the first information into the fine ranking model and outputs fine ranking scores of the first information.
Based on the acquired first information, the server inputs the first information into the refined model and outputs refined scores of the first information.
506. And based on the fine ranking scores of the first information, the server screens out the second information from the first information.
The server sorts the plurality of pieces of first information according to the fine ranking scores of the plurality of pieces of first information and the sequence from high to low of the fine ranking scores, and then screens out the plurality of pieces of second information from the plurality of pieces of first information based on the sorting result.
507. The server recommends the plurality of pieces of second information to the target user.
According to the method provided by the embodiment of the disclosure, because the rough ranking model is obtained by training based on the training result of the fine ranking model, the ranking mechanism of the fine ranking model can be learned, and when the method is applied online, the candidate information is ranked based on the same ranking mechanism, so that the information meeting the user requirements can be screened out. Meanwhile, when information screening is carried out based on the rough ranking model and the fine ranking model, not only the information characteristics of the candidate information and the user characteristics of the target user are considered, but also the cross characteristics between the candidate information and the target user are considered, and the target user and the candidate information are comprehensively depicted, so that the accuracy of the recommended information is greatly improved when information recommendation is carried out.
The embodiment of the present disclosure provides a training method for a coarse training model, and referring to fig. 6, a flow of the method provided by the embodiment of the present disclosure includes:
601. based on the fine ranking model, the server obtains a plurality of pieces of coarse ranking sample information from a plurality of pieces of training sample information used for training the fine ranking model.
Since the coarse model in the embodiment of the present disclosure is obtained based on the fine model training, the fine model also needs to be trained before the step is performed. Specifically, the training process of the refined model is as follows:
in the first step, the server obtains a plurality of pieces of training sample information.
In this step, the server obtains a plurality of pieces of training sample information from the internet according to the real exposure data, click data and conversion data on the internet, and each piece of training sample information is labeled with a corresponding click rate, conversion rate and the like.
And secondly, training the initial refined model by the server based on a plurality of pieces of training sample information to obtain a trained refined model.
The initial fine ranking model may be a DNN (Deep Neural Networks) model, and the like. The initial fine ranking model comprises a click rate estimation module, a conversion rate estimation module, a multi-target mechanism module and the like, when the fine ranking model is trained, each module in the fine ranking model can be trained respectively according to each piece of training sample information, wherein the multi-target mechanism module can be set according to the performance of the fine ranking model to be trained, for example, the fine ranking model needs to screen out 10 pieces of information from 200 pieces of information, and then a target mechanism in the multi-target mechanism module screens out 10 pieces of information from 200 pieces of information. Specifically, the server can train the click rate estimation module in the refined model according to the click rate marked in each piece of training sample information; the server can train the conversion rate module in the refined model according to the conversion rate marked in each piece of training sample information.
Based on the trained refined model, the server can obtain a plurality of pieces of rough-arranged sample information from a plurality of pieces of training sample information. The server generally has two modes when acquiring the coarse-layout sample information, wherein one mode is to acquire the coarse-layout sample information by adopting a sorting result of the fine-layout model, and the mode can learn the final sorting of a multi-target mechanism contained in the fine-layout model; the other mode is that the click rate and the conversion rate of each piece of training sample information are estimated according to the refined model, and the mode can learn how to score each piece of training sample information. Based on the two manners, when the server obtains the rough-layout sample information, the method includes, but is not limited to, the following steps:
in a possible implementation manner, based on the ordering concept of the first manner, the server inputs a plurality of pieces of training sample information into the fine ordering model, outputs the fine ordering score of each piece of training sample information, orders the plurality of pieces of training sample information according to the sequence of the fine ordering scores from high to low, then acquires the training sample information with the front preset number of orders, determines the training sample information with the front preset number of orders as a positive sample, determines the rest training sample information as a negative sample, and further forms the positive sample and the negative sample into a plurality of pieces of coarse ordering sample information. The sample obtaining mode trains the rough model according to the positive sample and the negative sample, so that the rough model learns how to obtain the preset number of pieces of positive sample information from a plurality of pieces of rough sample information, and since all candidate information participates in the training process of the rough model, the loss function of the mode is a global loss function. In the field of internet technology, this approach is called poitwise. The preset number of bits may be determined according to a learning mechanism of the coarse rule model, and if the learning mechanism of the coarse rule model is how to obtain 20 pieces of information from 100 pieces of information, the preset number of bits may be 20 bits.
In another possible implementation manner, based on the ordering concept of the first manner, the server inputs a plurality of pieces of training sample information into the fine-ordering model, outputs the fine-ordering score of each piece of training sample information, orders the plurality of pieces of training sample information according to the sequence from high to low of the fine-ordering score, then acquires the training sample information with the front preset number of orders, and determines the training sample information with the front preset number of orders and at least one piece of training sample information with each piece of training sample information from the same query request and with the order behind the preset number of orders as a plurality of pieces of coarse-ordering training sample information. The sample acquisition mode trains the rough model based on the positive case and the negative case under the same request at the same time, so that the rough model can learn how to reasonably sequence information under the same request. The loss function of this method is generally the interval between the positive case and the negative case, and the larger the difference between the positive case and the negative case is, the more the rough model can distinguish which information is good and which information is not good under the same request. This mode is called pairwise.
In another possible implementation manner, based on the ordering idea of the second manner, the server estimates the click rate and the conversion rate of each piece of training sample information based on the rough-arranged model, determines the ordering score of each piece of training sample information according to the click rate and the conversion rate of each piece of training sample information, and then acquires a plurality of pieces of rough-arranged training sample information from the plurality of pieces of training sample information based on the ordering score of each piece of training sample information.
602. The server extracts the training information characteristics of each piece of coarse-scale sample information, the user characteristics of the user clicking each piece of coarse-scale sample information and the training cross characteristics between each user and each piece of coarse-scale sample information.
For each piece of rough-arranged sample information, the server extracts the training information characteristics of each piece of training sample information, extracts the behavior characteristics and the attribute characteristics of the user clicking each piece of rough-arranged sample information, obtains the user characteristics of the user clicking each piece of rough-arranged sample information, and then extracts the training cross characteristics between each piece of rough-arranged sample information and the user clicking each piece of rough-arranged sample information.
603. And training the initial rough-layout model by the server based on the training information characteristics of each piece of rough-layout sample information, the user characteristics of the user clicking each piece of rough-layout sample information and the training cross characteristics between each user and each piece of rough-layout sample information to obtain a rough-layout model.
The initial coarse-line model may be LR, DSSM, FFM, or the like, and the server does not specifically limit the initial coarse-line model. The server sets model parameters for the initial coarse-layout model in advance, inputs training information characteristics of each piece of coarse-layout sample information, user characteristics of a user clicking each piece of coarse-layout sample information and training cross characteristics between each user and each piece of coarse-layout sample information into the initial coarse-layout model, and outputs a predicted coarse-layout score of each piece of coarse-layout sample information. And then, the server inputs the predicted coarse ranking score and the marked coarse ranking score of each piece of training sample information into a pre-constructed loss function to obtain a function value of the loss function, and if the function value of the loss function is larger than a preset value, model parameters of the initial coarse ranking model are adjusted until the function value of the loss function is smaller than the preset value. And the server determines the rough model corresponding to the model parameter which enables the function value of the loss function to be smaller than the preset value as the trained rough model. The preset value can be obtained according to the training precision of the rough model.
According to the method provided by the embodiment of the disclosure, because the rough ranking model is obtained by training based on the training result of the fine ranking model, the ranking mechanism of the fine ranking model can be learned, and when the method is applied online, the candidate information is ranked based on the same ranking mechanism, so that the information meeting the user requirements can be screened out. Meanwhile, when information screening is carried out based on the rough ranking model and the fine ranking model, not only the information characteristics of the candidate information and the user characteristics of the target user are considered, but also the cross characteristics between the candidate information and the target user are considered, and the target user and the candidate information are comprehensively depicted, so that the accuracy of the trained model is improved.
Referring to fig. 7, an embodiment of the present disclosure provides an information recommendation apparatus, including:
a first obtaining module 701, configured to obtain multiple pieces of candidate information;
a first extraction module 702, configured to extract a target user feature of a target user;
the second extraction module 703 is further configured to extract an information feature of each piece of candidate information and a cross feature of each piece of candidate information and the target user;
the input and output module 704 is used for inputting the target user characteristics, the information characteristics of each candidate information and the cross characteristics into the rough ranking model, outputting the rough ranking score of each candidate information, and obtaining the rough ranking model based on the training result of the fine ranking model;
the first screening module 705 is configured to screen out a plurality of pieces of first information from the plurality of pieces of candidate information based on the rough ranking score of each piece of candidate information;
the second screening module 706 is further configured to screen out a plurality of pieces of second information from the plurality of pieces of first information based on the fine ranking model;
and a recommending module 707 for recommending the plurality of pieces of second information to the target user.
In another embodiment of the present disclosure, a training apparatus for a steal model includes:
the second acquisition module is used for acquiring a plurality of pieces of rough-arranged sample information from a plurality of pieces of training sample information used for training the fine-arranged model based on the fine-arranged model;
the third extraction module is used for extracting the training information characteristics of each piece of coarse-line sample information, the user characteristics of the user clicking each piece of coarse-line sample information and the training cross characteristics between each user and each piece of coarse-line sample information;
and the training module is used for training the initial coarse layout model based on the training information characteristics of each piece of coarse layout sample information, the user characteristics of the user clicking each piece of coarse layout sample information and the training cross characteristics between each user and each piece of coarse layout sample information to obtain the coarse layout model.
In another embodiment of the present disclosure, the second obtaining module is further configured to obtain training sample information with a preset number of ranks in front based on the ranking scores of the plurality of pieces of training sample information by the ranking model; determining training sample information with a preset number of bits in the front as a positive sample, and determining the rest training sample information as a negative sample; and composing the positive sample and the negative sample into a plurality of pieces of coarse-row sample information.
In another embodiment of the present disclosure, the second obtaining module is further configured to obtain training sample information with a preset number of ranks in front based on the ranking scores of the plurality of pieces of training sample information by the ranking model; and determining the training sample information with the preset digit in the front and at least one piece of training sample information which is from the same query request and has the rank behind the preset digit as a plurality of pieces of coarse training sample information.
In another embodiment of the disclosure, the second obtaining module is further configured to estimate a click rate and a conversion rate of each piece of training sample information based on the rough model; determining the ranking score of each piece of training sample information according to the click rate and the conversion rate of each piece of training sample information; and acquiring a plurality of pieces of coarse training sample information from the plurality of pieces of training sample information based on the ranking score of each piece of training sample information.
In summary, according to the device provided by the embodiment of the present disclosure, since the rough ranking model is obtained by training based on the training result of the fine ranking model, the ranking mechanism of the fine ranking model can be learned, and when the device is applied online, the candidate information is ranked based on the same ranking mechanism, so that the information meeting the user requirements can be screened out. Meanwhile, when information screening is carried out based on the rough ranking model and the fine ranking model, not only the information characteristics of the candidate information and the user characteristics of the target user are considered, but also the cross characteristics between the candidate information and the target user are considered, and the target user and the candidate information are comprehensively depicted, so that the accuracy of the recommended information is greatly improved when information recommendation is carried out.
FIG. 8 illustrates a server for information recommendation, according to an example embodiment. Referring to FIG. 8, server 800 includes a processing component 822, which further includes one or more processors and memory resources, represented by memory 832, for storing instructions, such as applications, that are executable by processing component 822. The application programs stored in memory 832 may include one or more modules that each correspond to a set of instructions. Further, the processing component 822 is configured to execute instructions to perform the functions performed by the server in the information recommendation method described above.
The server 800 may also include a power component 826 configured to perform power management of the server 800, a wired or wireless network interface 850 configured to connect the server 800 to a network, and an input/output (I/O) interface 858. The Server 800 may operate based on an operating system, such as Windows Server, stored in the memory 832TM，Mac OS XTM，UnixTM,LinuxTM，FreeBSDTMOr the like.
The embodiment of the present disclosure provides a computer-readable storage medium, in which at least one program code is stored, and the at least one program code is loaded and executed by a processor to implement the information recommendation method shown in fig. 4 or fig. 5. The computer readable storage medium may be non-transitory. For example, the computer-readable storage medium may be a Read-Only Memory (ROM), a Random Access Memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, and the like.
It will be understood by those skilled in the art that all or part of the steps for implementing the above embodiments may be implemented by hardware, or may be implemented by a program instructing relevant hardware, where the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.
The above description is intended to be exemplary only and not to limit the present disclosure, and any modification, equivalent replacement, or improvement made without departing from the spirit and scope of the present disclosure is to be considered as the same as the present disclosure.
Claims (12)
1. An information recommendation method, characterized in that the method comprises:
acquiring a plurality of pieces of candidate information;
extracting target user characteristics of a target user, and extracting information characteristics of each piece of candidate information and cross characteristics of each piece of candidate information and the target user;
inputting the target user characteristics, the information characteristics of each candidate information and the cross characteristics into a rough ranking model, and outputting a rough ranking score of each candidate information, wherein the rough ranking model is obtained by training based on a training result of a fine ranking model;
screening out a plurality of pieces of first information from the plurality of pieces of candidate information based on the rough ranking score of each piece of candidate information;
screening out a plurality of pieces of second information from the plurality of pieces of first information based on the fine ranking model;
recommending the second information to the target user.
2. The method of claim 1, wherein the training process of the coarse row model is:
based on the fine ranking model, acquiring a plurality of pieces of coarse ranking sample information from a plurality of pieces of training sample information used for training the fine ranking model;
extracting training information characteristics of each piece of coarse-line sample information, user characteristics of a user clicking each piece of coarse-line sample information and training cross characteristics between each user and each piece of coarse-line sample information;
training an initial rough-layout model based on the training information characteristics of each piece of rough-layout sample information, the user characteristics of the user clicking each piece of rough-layout sample information and the training cross characteristics between each user and each piece of rough-layout sample information to obtain the rough-layout model.
3. The method according to claim 2, wherein the obtaining, based on the fine-ranking model, a plurality of pieces of coarse-ranking sample information from a plurality of pieces of training sample information used for training the fine-ranking model comprises:
acquiring training sample information with a preset number of ranks in front based on the fine ranking scores of the plurality of pieces of training sample information by the fine ranking model;
determining the training sample information with the front preset digit as a positive sample, and determining the rest training sample information as a negative sample;
and composing the positive sample and the negative sample into the plurality of pieces of coarse-sized sample information.
4. The method according to claim 2, wherein the obtaining, based on the fine-ranking model, a plurality of pieces of coarse-ranking sample information from a plurality of pieces of training sample information used for training the fine-ranking model comprises:
acquiring training sample information with a preset number of ranks in front based on the fine ranking scores of the plurality of pieces of training sample information by the fine ranking model;
and determining the training sample information with the preset digit and at least one piece of training sample information which is from the same query request and has the rank behind the preset digit as the plurality of pieces of coarse training sample information.
5. The method according to claim 2, wherein the obtaining, based on the fine-ranking model, a plurality of pieces of coarse-ranking sample information from a plurality of pieces of training sample information used for training the fine-ranking model comprises:
estimating the click rate and the conversion rate of each piece of training sample information based on the rough model;
determining the ranking score of each piece of training sample information according to the click rate and the conversion rate of each piece of training sample information;
and acquiring a plurality of pieces of rough training sample information from the plurality of pieces of training sample information based on the ranking score of each piece of training sample information.
6. An information recommendation apparatus, characterized in that the apparatus comprises:
the first acquisition module is used for acquiring a plurality of pieces of candidate information;
the first extraction module is used for extracting the target user characteristics of the target user;
the second extraction module is also used for extracting the information characteristic of each piece of candidate information and the cross characteristic of each piece of candidate information and the target user;
the input and output module is used for inputting the target user characteristics, the information characteristics of each candidate message and the cross characteristics into a rough ranking model, and outputting a rough ranking score of each candidate message, wherein the rough ranking model is obtained by training based on a training result of a fine ranking model;
the first screening module is used for screening out a plurality of pieces of first information from the plurality of pieces of candidate information based on the rough ranking score of each piece of candidate information;
the second screening module is further used for screening out a plurality of pieces of second information from the plurality of pieces of first information based on the fine ranking model;
and the recommending module is used for recommending the second information to the target user.
7. The apparatus of claim 6, wherein the training means for the coarse row model comprises:
the second obtaining module is used for obtaining a plurality of pieces of rough-arranged sample information from a plurality of pieces of training sample information used for training the fine-arranged model based on the fine-arranged model;
the third extraction module is used for extracting the training information characteristics of each piece of coarse-line sample information, the user characteristics of the user clicking each piece of coarse-line sample information and the training cross characteristics between each user and each piece of coarse-line sample information;
and the training module is used for training the initial coarse layout model based on the training information characteristics of each piece of coarse layout sample information, the user characteristics of the user clicking each piece of coarse layout sample information and the training cross characteristics between each user and each piece of coarse layout sample information to obtain the coarse layout model.
8. The apparatus according to claim 7, wherein the second obtaining module is further configured to obtain training sample information ranked in a previous preset number of bits based on the ranking scores of the plurality of pieces of training sample information by the ranking model; determining the training sample information with the front preset digit as a positive sample, and determining the rest training sample information as a negative sample; and composing the positive sample and the negative sample into the plurality of pieces of coarse-sized sample information.
9. The apparatus according to claim 7, wherein the second obtaining module is further configured to obtain training sample information ranked in a previous preset number of bits based on the ranking scores of the plurality of pieces of training sample information by the ranking model; and determining the training sample information with the preset digit and at least one piece of training sample information which is from the same query request and has the rank behind the preset digit as the plurality of pieces of coarse training sample information.
10. The apparatus according to claim 7, wherein the second obtaining module is further configured to estimate click through rate and conversion rate of each piece of training sample information based on the rough model; determining the ranking score of each piece of training sample information according to the click rate and the conversion rate of each piece of training sample information; and acquiring a plurality of pieces of rough training sample information from the plurality of pieces of training sample information based on the ranking score of each piece of training sample information.
11. A server, characterized in that the server comprises a processor and a memory, wherein at least one program code is stored in the memory, and the at least one program code is loaded and executed by the processor to implement the information recommendation method according to any one of claims 1 to 5.
12. A computer-readable storage medium, wherein at least one program code is stored in the storage medium, and the at least one program code is loaded and executed by a processor to implement the information recommendation method according to any one of claims 1 to 5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110185529.5A CN112883265A (en) | 2021-02-10 | 2021-02-10 | Information recommendation method and device, server and computer readable storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110185529.5A CN112883265A (en) | 2021-02-10 | 2021-02-10 | Information recommendation method and device, server and computer readable storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112883265A true CN112883265A (en) | 2021-06-01 |
Family
ID=76056491
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110185529.5A Withdrawn CN112883265A (en) | 2021-02-10 | 2021-02-10 | Information recommendation method and device, server and computer readable storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112883265A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113221017A (en) * | 2021-07-08 | 2021-08-06 | 智者四海（北京）技术有限公司 | Rough arrangement method and device and storage medium |
| CN115129975A (en) * | 2022-05-13 | 2022-09-30 | 腾讯科技（深圳）有限公司 | Recommendation model training method, recommendation device, recommendation equipment and storage medium |
| CN116383458A (en) * | 2023-06-02 | 2023-07-04 | 支付宝(杭州)信息技术有限公司 | Information pushing method and device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106339383A (en) * | 2015-07-07 | 2017-01-18 | 阿里巴巴集团控股有限公司 | Method and system for sorting search |
| US20190205701A1 (en) * | 2017-12-29 | 2019-07-04 | Guangdong Oppo Mobile Telecommunications Corp., Ltd. | Method for Training Model and Information Recommendation System |
| CN111143686A (en) * | 2019-12-30 | 2020-05-12 | 北京百度网讯科技有限公司 | Resource recommendation method and device |
| CN111369271A (en) * | 2018-12-25 | 2020-07-03 | 北京达佳互联信息技术有限公司 | Advertisement sorting method and device, electronic equipment and storage medium |
-
2021
- 2021-02-10 CN CN202110185529.5A patent/CN112883265A/en not_active Withdrawn
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106339383A (en) * | 2015-07-07 | 2017-01-18 | 阿里巴巴集团控股有限公司 | Method and system for sorting search |
| US20190205701A1 (en) * | 2017-12-29 | 2019-07-04 | Guangdong Oppo Mobile Telecommunications Corp., Ltd. | Method for Training Model and Information Recommendation System |
| CN111369271A (en) * | 2018-12-25 | 2020-07-03 | 北京达佳互联信息技术有限公司 | Advertisement sorting method and device, electronic equipment and storage medium |
| CN111143686A (en) * | 2019-12-30 | 2020-05-12 | 北京百度网讯科技有限公司 | Resource recommendation method and device |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113221017A (en) * | 2021-07-08 | 2021-08-06 | 智者四海（北京）技术有限公司 | Rough arrangement method and device and storage medium |
| CN113221017B (en) * | 2021-07-08 | 2021-10-29 | 智者四海（北京）技术有限公司 | Rough arrangement method and device and storage medium |
| CN115129975A (en) * | 2022-05-13 | 2022-09-30 | 腾讯科技（深圳）有限公司 | Recommendation model training method, recommendation device, recommendation equipment and storage medium |
| CN115129975B (en) * | 2022-05-13 | 2024-01-23 | 腾讯科技（深圳）有限公司 | Recommendation model training method, recommendation device, recommendation equipment and storage medium |
| CN116383458A (en) * | 2023-06-02 | 2023-07-04 | 支付宝(杭州)信息技术有限公司 | Information pushing method and device |
| CN116383458B (en) * | 2023-06-02 | 2023-08-11 | 支付宝(杭州)信息技术有限公司 | Information pushing method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110321422B (en) | Method for training model on line, pushing method, device and equipment | |
| CN109902708B (en) | Recommendation model training method and related device | |
| CN108921221B (en) | User feature generation method, device, equipment and storage medium | |
| CN109086439B (en) | Information recommendation method and device | |
| CN110162693B (en) | Information recommendation method and server | |
| CN110704739B (en) | Resource recommendation method and device and computer storage medium | |
| EP4181026A1 (en) | Recommendation model training method and apparatus, recommendation method and apparatus, and computer-readable medium | |
| CN112883265A (en) | Information recommendation method and device, server and computer readable storage medium | |
| CN111143686B (en) | Resource recommendation method and device | |
| CN109063104B (en) | Recommendation information refreshing method and device, storage medium and terminal equipment | |
| CN110413867B (en) | Method and system for content recommendation | |
| CN110825966A (en) | Information recommendation method and device, recommendation server and storage medium | |
| CN111506820A (en) | Recommendation model, method, device, equipment and storage medium | |
| CN112395487A (en) | Information recommendation method and device, computer-readable storage medium and electronic equipment | |
| CN111782937A (en) | Information sorting method and device, electronic equipment and computer readable medium | |
| CN112364184A (en) | Method, device, server and storage medium for ordering multimedia data | |
| CN114428910A (en) | Resource recommendation method and device, electronic equipment, product and medium | |
| CN115222433A (en) | Information recommendation method and device and storage medium | |
| CN112115354A (en) | Information processing method, information processing apparatus, server, and storage medium | |
| CN116956996A (en) | Data processing method, device, medium and equipment | |
| CN114491093B (en) | Multimedia resource recommendation and object representation network generation method and device | |
| CN112905885B (en) | Method, apparatus, device, medium and program product for recommending resources to user | |
| CN115391589A (en) | Training method and device for content recall model, electronic equipment and storage medium | |
| CN115238676A (en) | Method and device for identifying hot spots of bidding demands, storage medium and electronic equipment | |
| CN114357242A (en) | Training evaluation method and device based on recall model, 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 | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20210601 |
|
| WW01 | Invention patent application withdrawn after publication |