CN-121998675-A - Commodity recall method and device and electronic equipment

CN121998675ACN 121998675 ACN121998675 ACN 121998675ACN-121998675-A

Abstract

The invention provides a commodity recall method, device and electronic equipment, the method comprises the steps of obtaining N commodities to be recalled, determining a characterization vector of each commodity to be recalled based on a pre-trained target commodity recall model, layering the characterization vectors of the N commodities to be recalled to generate a layered navigable small world HNSW map, wherein the HNSW map comprises a plurality of layers of navigable small world NSW maps which are arranged up and down, determining target recall commodities from the commodities to be recalled according to the HNSW map and the target commodity recall model, and determining target recall commodities from the commodities to be recalled based on the HNSW map and the target commodity recall model.

Inventors

WU CHANGFA
CHEN MENGXIANG
ZHONG JIYUAN
Ma Mian
ZHAO XIWEI

Assignees

北京沃东天骏信息技术有限公司
北京京东世纪贸易有限公司

Dates

Publication Date: 20260508
Application Date: 20251231

Claims (15)

1. A merchandise recall method, the method comprising: acquiring N to-be-recalled commodities, and determining a characterization vector of each to-be-recalled commodity based on a pre-trained target commodity recall model, wherein N is a natural number greater than or equal to 1; Layering the characterization vectors of the N commodities to be recalled to generate a layered navigable small world HNSW map, wherein the HNSW map comprises a plurality of layers of navigable small world NSW maps arranged up and down; And determining the target recall commodity from the commodity to be recalled according to the HNSW diagram and the target commodity recall model.
2. The method of claim 1, wherein the layering the characterization vectors for the N items to recall to generate a layered navigable small world HNSW map comprises: And taking the characterization vector as a node to be inserted, acquiring a target level of the HNSW graph, inserting the node to be inserted into each previous level of the current level and the target level of the node to be inserted, determining a neighbor node of the node to be inserted in each NSW graph, and connecting the neighbor node with the node to be inserted until the HNSW graph comprises all the characterization vectors to generate the HNSW graph.
3. The method of claim 2, wherein the determining a target recall commodity from the to-be-recalled commodities according to the HNSW diagram and the target commodity recall model comprises: Starting from the top h layer of the HNSW graph, determining a model input vector corresponding to the current i layer and inputting the model input vector into the target commodity recall model to obtain a target predicted value corresponding to the current i layer, wherein the target predicted value is at least one of a click rate predicted value and a click-after-conversion value; Selecting a characterization vector corresponding to a target predicted value meeting a set condition as a first target characterization vector, entering an i-1 layer of the current layer, repeatedly executing the process until the characterization vector corresponding to the target predicted value meeting the set condition is selected as a target characterization vector at the bottom layer of the HNSW graph, and taking a commodity to be recalled corresponding to the target characterization vector as the target recall commodity.
4. The method of claim 3, wherein the determining that current layer i inputs an input vector of the target commodity recall model comprises: Aiming at a top h layer, determining an NSW graph representation vector of the top h layer as an input vector for inputting the target commodity recall model; For non-top layers, determining a second target token vector adjacent to the first target token vector and the first target token vector as input vectors into the target commodity recall model.
5. The method according to any one of claims 3-4, wherein selecting the token vector corresponding to the target predicted value that satisfies the set condition further includes: And sequencing the target predicted values output by each layer of NSW images according to each layer of NSW images of the HNSW images, and taking characterization vectors corresponding to the first K target predicted values sequenced as the characterization vectors meeting the set condition, wherein K is a positive integer.
6. The method of claim 1, wherein the training process of the target commodity recall model comprises: Acquiring a training data set, wherein the training data set comprises a user tag, a user history click sequence, main commodity features and candidate commodity features corresponding to commodity detail pages; And training the initial commodity recall model according to the training data set to obtain a trained target commodity recall model, wherein the target commodity recall model is used for determining the click rate value and the click-to-conversion rate value of the commodity.
7. The method of claim 1, wherein training the commodity recall model based on the training dataset to obtain a trained target commodity recall model comprises: inputting the training data set into the commodity recall model to obtain a predicted click rate value and a predicted conversion rate value of candidate commodities after clicking; and adjusting the commodity recall model according to the click rate predicted value and the click rate actual value, the click rate predicted value and the click rate conversion rate actual value, and continuously training the adjusted commodity recall model until the training ending condition is met, so as to obtain the target commodity recall model.
8. The method of claim 7, wherein training the initial commodity recall model based on the training dataset to obtain a trained target commodity recall model comprises: Determining a first loss function of the commodity recall model according to the click rate predicted value and the click rate actual value, and determining a second loss function of the commodity recall model according to the click rate predicted value and the click rate actual value; And adjusting the commodity recall model according to the first loss function and the second loss function until the training ending condition is met to obtain the target commodity recall model.
9. The method of claim 8, wherein inputting the training dataset into the commodity recall model to obtain a click rate prediction value and a post-click conversion rate prediction value for the candidate commodity comprises: Inputting the training data set into a commodity recall model; Obtaining embedded vectors of various data of the training data set, and fusing the embedded vectors of the various data to obtain a multi-dimensional characterization vector, wherein the multi-dimensional characterization vector comprises a first characterization vector corresponding to a user tag, a second characterization vector corresponding to a main commodity feature, a third characterization vector corresponding to a candidate commodity feature, and a high-order characterization vector corresponding to a user history click sequence; acquiring a first aggregation vector and a second aggregation vector of the high-order characterization vector according to the second characterization vector and the third characterization vector; Performing feature crossing on the second feature vector and the third feature vector to obtain a first feature crossing vector and a second feature crossing vector; And outputting the predicted click rate value and the predicted post-click conversion rate value of the candidate commodity according to the second characterization vector, the third characterization vector, the first aggregation vector, the second aggregation vector, the first characteristic crossing vector and the second characteristic crossing vector.
10. The method of claim 8, wherein the obtaining the embedded vectors of the various types of data in the training data set and fusing the embedded vectors of the various types of data to obtain the multi-dimensional characterization vector comprises: acquiring respective embedded vectors of the user tag, the user history click sequence, the main commodity feature and the candidate commodity feature; Fusing the embedded vectors of the user tag, the user history click sequence, the main commodity feature and the candidate commodity feature to obtain respective characterization vectors; Fusing the embedded vectors of the user tag, the main commodity feature and the candidate commodity feature to obtain the characterization vectors of the user tag, the main commodity feature and the candidate commodity feature; And performing a transform and attention mechanism processing on the embedded vector of the user history click sequence to obtain a high-order characterization vector corresponding to the user history click sequence.
11. The method of claim 10, wherein the obtaining the first and second aggregate vectors of the higher order vectors from the second and third token vectors comprises: according to the second characterization vector, the third characterization vector and the high-order characterization vector, attention values of the second characterization vector and the third characterization vector and the high-order characterization vector are obtained; acquiring the first aggregate vector according to the attention value corresponding to the second characterization vector and the high-order vector; and acquiring the second polymerization vector according to the attention value corresponding to the third characterization vector and the high-order characterization vector.
12. A merchandise recall apparatus, the apparatus comprising: the first acquisition module is used for acquiring N commodities to be recalled, determining the characterization vector of each commodity to be recalled based on a pre-trained target commodity recall model, wherein N is a natural number greater than or equal to 1; the second obtaining module is used for layering the characterization vectors of the N commodities to be recalled to generate a layered navigable small world HNSW diagram, wherein the HNSW diagram comprises a plurality of layers of navigable small world NSW diagrams arranged up and down; And the determining module is used for determining the target recall commodity from the commodity to be recalled according to the HNSW diagram and the target commodity recall model.
13. An electronic device comprising a memory and a processor; wherein the memory and the processor are in communication with each other via an internal connection, the memory being for storing instructions, the processor being for executing the memory-stored instructions and, when the processor executes the memory-stored instructions, causing the processor to perform the method of any one of claims 1-11.
14. A computer readable storage medium, characterized in that the computer readable storage medium stores a computer program which, when run on a computer, implements the method according to any of claims 1-11.
15. A computer program product comprising a computer program which, when executed by a processor, implements the method according to any of claims 1-11.

Description

Commodity recall method and device and electronic equipment Technical Field The disclosure relates to the technical field of information processing, and in particular relates to a commodity recall method, a commodity recall device and electronic equipment. Background At present, commodity recall scenes aiming at commodity detail pages (business detail pages) often comprise recall sources such as homonymous recall, similar recall and Multi-Interest recommendation recall (Multi-Interest Network WITH DYNAMIC, short for MIND) and the like, and the recall sources are used in combination, but the problem that the recalled commodity based on the recall sources is inaccurate, for example, the recalled commodity does not accord with the user expectation, the recalled commodity is irrelevant to the commodity of the business detail pages and the like, so that the user experience is poor, and therefore, how to improve the accuracy of the recalled commodity to improve the user experience is a problem to be solved. Disclosure of Invention The present disclosure provides a commodity recall method, apparatus, electronic device, storage medium, and computer program product. An embodiment of a first aspect of the present disclosure provides a commodity recall method, which includes obtaining N commodities to be recalled, determining a characterization vector of each commodity to be recalled based on a pre-trained target commodity recall model, layering the characterization vectors of the N commodities to be recalled to generate a layered navigable small world HNSW map, wherein the HNSW map includes a plurality of layers of navigable small world NSW maps arranged up and down, and determining a target recall commodity from the commodities to be recalled according to the HNSW map and the target commodity recall model. In the embodiment of the disclosure, N to-be-recalled commodities are obtained, the characterization vector of each to-be-recalled commodity is determined based on a pre-trained target commodity recall model, N is a natural number which is greater than or equal to 1, the characterization vectors of the N to-be-recalled commodities are layered to generate a layered navigable small world HNSW diagram, wherein the HNSW diagram comprises a plurality of layers of navigable small world NSW diagrams which are arranged up and down, the target recall commodity is determined from the to-be-recalled commodities according to the HNSW diagram and the target commodity recall model, the target recall commodity can be determined from the to-be-recalled commodities based on the HNSW diagram and the target commodity recall model, the target recall commodity not only accords with user interests, but also is related to a main commodity of a commodity page, accuracy and reliability of determining the target recall commodity are improved, and experience of a user is improved. An embodiment of a second aspect of the present disclosure provides a commodity recall device, which includes a first obtaining module, a second obtaining module, and a determining module, wherein the first obtaining module is used for obtaining N commodities to be recalled, determining a characterization vector of each commodity to be recalled based on a pre-trained target commodity recall model, N is a natural number greater than or equal to 1, the second obtaining module is used for layering the characterization vectors of the N commodities to be recalled to generate a layered navigable small world HNSW map, the HNSW map includes a plurality of layers of navigable small world NSW maps arranged up and down, and the determining module is used for determining a target recall commodity from the commodities to be recalled according to the HNSW map and the target commodity recall model. An embodiment of a third aspect of the present disclosure provides 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, the instructions being executable by the at least one processor to enable the at least one processor to perform a merchandise recall method as in the embodiment of the first aspect described above. An embodiment of a fourth aspect of the present disclosure proposes a computer-readable storage medium storing computer instructions for causing the computer to perform the merchandise recall method as described above in the embodiment of the first aspect. A fifth aspect embodiment of the present disclosure proposes a computer program product comprising a computer program which, when executed by a processor, implements the merchandise recall method of the first aspect embodiment of the present disclosure. Additional aspects and advantages of the disclosure will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the disclosure. Drawings Th