CN-121998730-A - Sequencing model training, commodity sequencing method and commodity recommendation system

CN121998730ACN 121998730 ACN121998730 ACN 121998730ACN-121998730-A

Abstract

The embodiment of the application discloses a sequencing model training method, a commodity sequencing method and a commodity recommendation system, which comprise the steps of constructing an autoregressive-based generated user behavior model, training the generated user behavior model based on a user history behavior sequence, wherein the generated user behavior model comprises an embedded layer and a coding layer, the embedded layer and the coding layer are used for predicting the next behavior target of a user according to the user behavior sequence, after the generated user behavior model is trained, transferring parameters of the embedded layer in the generated user behavior model to a sequencing model, and fixing the parameters of the embedded layer in the sequencing model and participating in training in the process of training the sequencing model. By the embodiment of the application, the problem of overfitting of the sequencing model can be relieved.

Inventors

WU BINGCHAO
WANG CHUNQI
Pang Taotian
SHEN LEI
WANG BING
ZENG XIAOYI

Assignees

杭州阿里巴巴海外互联网产业有限公司
阿里巴巴新加坡控股有限公司

Dates

Publication Date: 20260508
Application Date: 20251205

Claims (13)

1. A training method of a sorting model is characterized in that the sorting model is used for calculating and sorting matching degree between candidate commodities and user interests in commodity recommendation scenes, an embedded layer for preliminarily mapping input data into vectors and a coding layer for modeling relations between different commodities in a user behavior sequence are included in a network structure of the sorting model, and the method comprises the following steps: Constructing an autoregressive-based generated user behavior model, and training the generated user behavior model based on a user history behavior sequence, wherein the generated user behavior model comprises the embedded layer and the coding layer and is used for predicting the next behavior target of a user according to the user behavior sequence; After the training of the generated user behavior model is completed, migrating parameters of the embedded layer in the generated user behavior model to a sequencing model; And in the process of training the sequencing model, fixing the parameters of the embedded layer in the sequencing model, and taking the parameters of the coding layer into training.
2. The method of claim 1, wherein the step of determining the position of the substrate comprises, The generated user behavior model and the sequencing model adopt a periodic training mode, and when each period is trained, the complete model parameters obtained by training the generated user behavior model in the previous period are migrated into the generated user behavior model to serve as the initialization weight of the current period; after the generated user behavior model finishes the training of the current period, the parameters of the embedded layer are migrated to the sorting model, and after the parameters of the coding layer obtained by the last training period of the sorting model are migrated to the sorting model, the sorting model is trained for the current period.
3. The method of claim 1, wherein the step of determining the position of the substrate comprises, The ordering model multiplexes the network structure of the coding layer in the generated user behavior model.
4. A method according to claim 1 or 3, characterized in that, The coding layer in the generated user behavior model comprises a self-attention layer, wherein the self-attention layer avoids the generated user behavior model from knowing information of the next commodity and the commodity behind the next commodity in the user behavior sequence in the prediction process by adding a mask to commodity information behind the target position in the user behavior sequence.
5. A method according to claim 1 or 3, characterized in that, The coding layer in the generated user behavior model adopts a multi-head attention mechanism and/or a plurality of stacked block structures so as to realize multi-order interaction between different commodities in the process of modeling the relationship between different commodities in the user behavior sequence.
6. A method for ordering goods in a process of recommending goods, comprising: When commodity recommendation is required to be carried out on a target user, acquiring a user behavior sequence of the target user, and determining a candidate commodity set; Processing the user behavior sequence and commodity information in the candidate commodity set by using a sequencing model to calculate matching degree and sequencing of a plurality of candidate commodities and user interests of the target user respectively, wherein the sequencing model is trained by using the method of any one of claims 1 to 5.
7. The method as recited in claim 6, further comprising: in the process of processing by using the ranking model, the user history sequence modeling part and the candidate commodity modeling part are calculated separately so as to calculate the user behavior sequence part only once and buffer the calculation result for subsequent calculation.
8. The method as recited in claim 7, further comprising: And filling the length of the user behavior sequence into a fixed value, and constructing an operator by utilizing a bottom library, wherein the internal matrix operation of the computing processing unit is optimized by a parallel computing platform and/or a template library for realizing high-performance matrix multiplication and related computation, and before computation, a memory address for caching tensors of the computing result of the user behavior sequence part is obtained, so that the computing processing unit writes the computing result into the memory address.
9. The method of claim 8, wherein the step of determining the position of the first electrode is performed, The constructed operators support half-precision floating point numbers and brain floating point numbers so as to select different precision according to hardware support conditions and/or model requirements.
10. A commodity recommendation system is characterized by comprising a service end and a client end, wherein, The server side is used for acquiring a user behavior sequence of a target user when commodity recommendation is required to be carried out on the target user, processing commodity information in the user behavior sequence and the candidate commodity set by using a sequencing model after determining a candidate commodity set so as to calculate matching degree between a plurality of candidate commodities and user interests of the target user respectively and arranging the commodity information, wherein the sequencing model is trained by using the method of any one of claims 1 to 5; And the client is used for displaying the recommended commodity information in the target page according to the candidate commodities and the sequencing result thereof determined by the server.
11. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the steps of the method according to any one of claims 1 to 9.
12. An electronic device, comprising: one or more processors, and A memory associated with the one or more processors for storing program instructions that, when read for execution by the one or more processors, perform the steps of the method of any of claims 1 to 9.
13. A computer program product comprising computer program/computer-executable instructions which, when executed by a processor in an electronic device, implement the steps of the method of any one of claims 1 to 9.

Description

Sequencing model training, commodity sequencing method and commodity recommendation system Technical Field The application relates to the technical field of information processing, in particular to a sequencing model training, a commodity sequencing method and a commodity recommendation system. Background In the merchandise information service system, merchandise recommendation is an important function. Through commodity recommendation system, can help the user to discover the commodity that they are interested in fast, reduce search cost. From the aspect of the platform, the commodity which is most likely to be clicked or purchased can be displayed in a limited page space, so that the high-efficiency utilization of the flow is realized, and the indexes such as GMV (Gross Merchandise Volume, commodity transaction total) and the like of the platform are facilitated to be improved. From the technical implementation point of view, the commodity recommendation system is a fusion of various algorithms and technologies, and the core flow of the commodity recommendation system can be summarized into a multi-layer funnel model such as recall-sequencing-rearrangement. The recall layer is mainly used for primarily screening hundreds to thousands of candidate commodities possibly related to users from millions or even tens of millions of commodity libraries, the sorting layer is used for accurately scoring hundreds of candidate commodities generated by the recall layer, predicting the click rate, conversion rate and the like of each commodity by the users and sorting the commodities according to the scores, and the rearrangement layer is used for fine-tuning a sorted list before finally displaying the recommended results to the users so as to meet diversified requirements of business rules, user experience and the like. The ranking layer is usually implemented by using a complex ranking algorithm model, for example, by automatically learning complex interactions between features through a deep learning model, capturing the immediate interests of the user through a user history behavior sequence, and then ranking the candidate commodities according to the scoring result by combining information such as the features of the candidate commodities (including commodity category, brand, price, sales volume, score, time to shelf, etc.), the context features (including time (weekday/weekend, holiday), season, weather, etc.), and the scoring result is used to represent the probability of clicking/purchasing the candidate commodities by the user. That is, the commodity recommendation result finally displayed to the user is closely related to the scoring of the ranking model and the ranking result, and obviously, the quality of the ranking model directly affects the quality of the recommendation result, and how to further optimize the ranking model to reduce the service resource waste is also an important subject of attention of the person skilled in the art all the time. Disclosure of Invention The application provides a sequencing model training, commodity sequencing method and commodity recommendation system, which can relieve the problem of overfitting of a sequencing model caused by mismatching of a learnable parameter scale and an optimization target, further improve the recommendation quality and reduce the waste of service resources. The application provides the following scheme: A training method of a sorting model is used for calculating and sorting matching degree between candidate commodities and user interests in commodity recommendation scenes, the network structure of the sorting model comprises an embedded layer for preliminarily mapping input data into vectors and a coding layer for modeling relations between different commodities in a user behavior sequence, and the method comprises the following steps: Constructing an autoregressive-based generated user behavior model, and training the generated user behavior model based on a user history behavior sequence, wherein the generated user behavior model comprises the embedded layer and the coding layer and is used for predicting the next behavior target of a user according to the user behavior sequence; After the training of the generated user behavior model is completed, migrating parameters of the embedded layer in the generated user behavior model to a sequencing model; And in the process of training the sequencing model, fixing the parameters of the embedded layer in the sequencing model, and taking the parameters of the coding layer into training. The method comprises the steps that a periodic training mode is adopted for the generated user behavior model and the sequencing model, and when training is carried out in each period, complete model parameters obtained by training the generated user behavior model in the previous period are migrated to the generated user behavior model to serve as the initialization weight of the current period; after the generated u