Search result sorting method, device and equipment and readable storage medium
By building a neural network model with an ordered estimation layer and combining product relevance levels with user interaction behavior to generate interest scores, the problem of slow sorting of product search results on e-commerce platforms was solved, achieving faster and more accurate sorting and improving user experience.
Patent Information
- Application Number
- CN202510892845.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-30
- Publication Date
- 2025-09-19
AI Technical Summary
In the existing technology, the product search result sorting method of the e-commerce platform cannot efficiently process a large number of product search results, resulting in slow sorting speed and affecting the user experience.
By building a neural network model containing an ordered estimation layer, the target user's interactive behavior and product information are used to rank products. The relevance level of product search results and user interactive behavior are combined to generate interest scores for product ranking.
It improves the speed and accuracy of product search result sorting, reduces computational complexity, and enhances user experience.
Smart Images
Figure CN120672430A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data processing technology, and more specifically, to a search result sorting method, apparatus, device, and readable storage medium. Background Art
[0002] In today's digital age, the e-commerce industry is booming. Major e-commerce platforms offer an increasingly diverse range of products, encompassing everything from daily necessities to high-end electronics, fashion apparel, and various services.
[0003] When consumers enter specific keywords into the search bar on e-commerce platforms, the search system can retrieve a vast number of related products. If these products are displayed based on simple sorting rules, such as random order or chronological order, consumers often have to invest considerable time and effort in sifting through each relevant product individually to find the product that meets their needs.
[0004] To address the problem of unreasonable ranking of product search results, the existing technology adopts a method of comparing two related products, so that relatively better related products are ranked at the top of the search results. The essence of this method is to transform the ranking problem of product search results into a comparison problem between two products. However, in actual application, when the number of related products obtained by the search is large, the number of comparison operations required will increase dramatically. Since a large number of comparisons between multiple products are involved, this will lead to an explosive growth in the amount of comparison data, which will significantly affect the sorting speed of product search results and reduce the search efficiency and user experience of the e-commerce platform. Summary of the Invention
[0005] In view of this, the present application provides a search result sorting method, apparatus, device and readable storage medium to address the shortcoming of slow sorting speed in existing sorting technologies.
[0006] In order to achieve the above objectives, the following solutions are proposed:
[0007] A search result sorting method, comprising:
[0008] Obtaining a ranking model including an ordered estimation layer, wherein the ordered estimation layer is trained based on a plurality of product samples labeled with corresponding relevance levels;
[0009] Obtain multiple product search results that match the search criteria entered by the target user;
[0010] Inputting the target user's interactive behavior, the search conditions, and each product search result into the ranking model, using the order estimation layer to evaluate the product relevance level of each product search result, and using the ranking model to combine the product relevance level of each product search result, the search conditions, and the target user's interactive behavior to score each product search result to obtain an interest score for each product search result;
[0011] Sort the product search results based on their interest scores.
[0012] Optionally, obtaining a ranking model including an ordered estimation layer includes:
[0013] Obtaining an initial ranking model for scoring products based on user transaction characteristics and search information;
[0014] Construct an initial classification module;
[0015] Obtain multiple product samples marked with corresponding relevance levels, each relevance level indicating a probability of consumption intention for the corresponding product sample;
[0016] Using each product sample to train the initial classification module, and using the final initial classification module as the order estimation layer;
[0017] The order estimation layer is added to the initial sorting model to generate a sorting model including the order estimation layer.
[0018] Optionally, obtaining an initial ranking model for scoring products based on user transaction characteristics and search information includes:
[0019] Obtaining a ranking module and multiple triples, each triple containing corresponding product training data, corresponding search information, and transaction characteristics of the corresponding search user, and each triple annotated with a relevance rating value generated based on the corresponding search information and the transaction characteristics of the corresponding search user;
[0020] The ranking module is iteratively trained using each triplet, and the resulting ranking module is an initial ranking model for scoring products based on user transaction characteristics and search information.
[0021] Optionally, obtaining a plurality of product samples marked with corresponding relevance levels includes:
[0022] Obtain multiple product information and determine the relevance level corresponding to each product information based on the corresponding transaction volume, add-to-cart volume, favorite volume, pageview volume, and click volume of each product information;
[0023] The relevance level of each product information is used as a label for the corresponding product information to form a product sample.
[0024] Optionally, the initial classification module is trained using each product sample, and the resulting initial classification module is used as the order estimation layer, including:
[0025] Each product sample is sequentially input into the initial classification module to obtain the predicted grade output by the initial classification module. The initial classification module includes a category tendency assessment layer and a content bias determination layer. The category tendency assessment layer is used to assess the category tendency of each product sample, and the content bias determination layer is used to determine the correspondence between each product sample and multi-dimensional business indicators. The predicted grade of each product sample is generated based on the corresponding category tendency and each corresponding relationship.
[0026] Calculate the cross entropy loss based on the correlation level of each predicted level and its corresponding product sample;
[0027] Based on each cross-entropy loss, combined with the gradient descent method, the parameters of the initial classification module are optimized until the cross-entropy loss meets the preset stopping condition or the product sample iteration is completed. The final trained initial classification module with the content bias judgment layer removed is used as the order estimation layer.
[0028] Optionally, adding the order estimation layer to the initial sorting model to generate a sorting model including the order estimation layer includes:
[0029] The target increasing function is linked to the output end of the order estimation layer, and the order estimation layer linked to the target increasing function is added to the front end of the fully connected layer of the initial sorting model to obtain a sorting model including the order estimation layer.
[0030] Optionally, linking the target increasing function to the output end of the order estimation layer includes:
[0031] A Sigmoid function is connected to the output end of the order estimation layer.
[0032] A search result sorting device, comprising:
[0033] an acquisition module, configured to acquire a ranking model comprising an ordered estimation layer, wherein the ordered estimation layer is trained based on a plurality of commodity samples labeled with corresponding relevance levels;
[0034] A search module is used to obtain multiple product search results that match the search criteria entered by the target user;
[0035] a scoring module for inputting the target user's interactive behavior, the search conditions, and each product search result into the ranking model, using the order estimation layer to evaluate the product relevance level of each product search result, and using the ranking model to combine the product relevance level of each product search result, the search conditions, and the target user's interactive behavior to score each product search result to obtain an interest score for each product search result;
[0036] The sorting module is used to sort the product search results based on the interest scores.
[0037] A search result ranking device, comprising a memory and a processor;
[0038] The memory is used to store programs;
[0039] The processor is used to execute the program to implement each step of the above-mentioned search result sorting method.
[0040] A readable storage medium stores a computer program, which, when executed by a processor, implements the various steps of the above-mentioned search result ranking method.
[0041] It can be seen from the above technical solution that the search result ranking method provided by the present application can obtain a ranking model including an ordered estimation layer, and the ordered estimation layer is trained based on product samples marked with corresponding relevance levels; multiple product search results that match the search conditions input by the target user are obtained; based on this, the present application can recall multiple product search results that need to be ranked through the search conditions; then, the target user's interactive behavior, the search conditions and each product search result can be input into the ranking model, and the ordered estimation layer can be used to evaluate the product relevance level of each product search result, and the ranking model can be used to combine the product relevance level of each product search result, the search conditions and the interactive behavior of the target user to score each product search result to obtain an interest score for each product search result; based on this, the present application can evaluate the relevance level of each product search result, and comprehensively score each product search result one by one based on the interactive behavior characteristics of the search user, the search conditions and the evaluated product relevance level; that is, the product search results are scored based on multiple dimensions to ensure the reliability and accuracy of the interest score. At the same time, by estimating the product relevance level, it is possible to predict the interaction between the corresponding product search results and the user, such as whether the target user purchases the corresponding product search results, clicks on the corresponding product search results, or collects the corresponding product search results. Thus, by comprehensively searching the interactive behavior characteristics of the user, the search conditions, and the product relevance level of the recalled product, it is possible to predict the target user's interest in different product search results, and characterize the interest level corresponding to each product search result as an interest score. Then, the product search results can be sorted based on each interest score. Based on this, the present application can use a sorting model to comprehensively consider the search conditions and interactive behavior of the present application to predict the product relevance level, and then perform interest score scoring. The sorting of each product search result can be completed by the interest score. The amount of data required to be processed is the amount of data for each product search result, interactive behavior, and search conditions. Compared with the existing technology, the amount of data is relatively small. Moreover, scoring by the sorting model can process multiple product search results in parallel, reduce unnecessary calculation steps, reduce calculation complexity, further speed up the scoring calculation speed, and thus further speed up the sorting speed. It can be seen that the present application can speed up the sorting speed while ensuring the accuracy of the sorting results, thereby improving the user's experience. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without any creative work.
[0043] Figure 1 A flow chart of a search result sorting method disclosed in an embodiment of the present application;
[0044] Figure 2 This is a structural block diagram of a search result sorting device disclosed in an embodiment of the present application;
[0045] Figure 3 This is a hardware structure block diagram of a search result sorting device disclosed in an embodiment of the present application. DETAILED DESCRIPTION
[0046] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.
[0047] An embodiment of the present application provides a search result sorting method, which can be applied to various search engine systems or e-commerce platforms, and can also be applied to various computer terminals or smart terminals. Its execution entity can be a processor or server of the computer terminal or smart terminal.
[0048] Next, combine Figure 1 The search result sorting method of this application is introduced in detail, including the following steps:
[0049] Step S1: Obtain a sorting model including an ordered estimation layer.
[0050] Specifically, a trained ranking model may be obtained.
[0051] Considering that the neural network model has strong fitting performance, the sorting model can be a neural network model.
[0052] Among them, the penultimate layer of the sorting model can be a sequence estimation layer.
[0053] The order estimation layer can be pre-trained based on a plurality of commodity samples labeled with corresponding relevance levels.
[0054] The relevance level can be used to indicate the probability of multiple users' consumption intentions for corresponding product samples.
[0055] The higher the correlation level, the higher the probability of consumption intention.
[0056] Different levels of relevance can be represented by different numerical values.
[0057] The larger the value, the higher the level of relevance.
[0058] The numerical value of each correlation level is positively correlated with the probability of consumption intention.
[0059] Therefore, each product sample can be marked with a numerical value corresponding to the relevance level.
[0060] Step S2: Obtain multiple product search results that match the search criteria input by the target user.
[0061] Specifically, in response to the search conditions input by the target user into the search box, multiple related product search results can be recalled based on the search conditions.
[0062] The search condition may be expressed in a variety of different ways. For example, the search condition may be a keyword or a product title.
[0063] Step S3: input the target user's interactive behavior, the search conditions, and each product search result into the ranking model, use the order estimation layer to evaluate the product relevance level of each product search result, and use the ranking model to combine the product relevance level of each product search result, the search conditions, and the target user's interactive behavior to score each product search result to obtain an interest score for each product search result.
[0064] Specifically, the target user's interactive behavior, the input search conditions, and the recalled product search results can be input into the ranking model accordingly.
[0065] There are many ways to obtain the interactive behavior of target users:
[0066] For example, the target user’s interaction preferences can be generated based on their behavioral data and social data;
[0067] Based on the interaction preferences, the interaction type of the target user may be determined;
[0068] The target user's interactive behavior can be determined based on the interactive behaviors of other users who share the same interactive type as the target user.
[0069] Among them, behavioral data may include authorized behaviors such as product browsing history, product search history, product click behavior, product purchase history, and product dwell time that reflect the target user's interest in the product.
[0070] Social data may include topics that users follow, information about their friends, content posted by users, groups that users participate in, and other authorized behaviors that reflect the personal interests of the target user.
[0071] For another example, the target user's interactive behavior can be extracted from the target user's behavioral data.
[0072] The various levels of the ranking model can be used to analyze the target user's interactive behavior, search conditions and various product search results, and the order estimation layer can be used to predict the product relevance level of each product search result.
[0073] Based on the product relevance level, target user's interactive behavior and search conditions, the remaining levels of the ranking model can be combined to perform an interest score on each product search result to generate an interest score for each product search result.
[0074] Step S4: Sort the product search results based on the interest scores.
[0075] Specifically, the product search results may be sorted from large interest scores to small interest scores.
[0076] It can be seen from the above technical solutions that the search result ranking method provided by the present application can obtain a ranking model including an ordered estimation layer, wherein the ordered estimation layer is trained based on product samples marked with corresponding relevance levels; a plurality of product search results that match the search conditions input by the target user are obtained; based on this, the present application can recall a plurality of product search results that need to be ranked by the search conditions; then, the target user's interactive behavior, the search conditions and each product search result can be input into the ranking model, and the ordered estimation layer can be used to evaluate the product relevance level of each product search result, and the ranking model can be used to combine the product relevance level of each product search result, the search conditions and the interactive behavior of the target user to score each product search result to obtain an interest score for each product search result; based on this, the present application can evaluate the relevance level of each product search result, and score each product search result one by one based on the interactive behavior characteristics of the search user, the search conditions and the evaluated product relevance level; that is, the product search results are scored based on multiple dimensions to ensure the reliability and accuracy of the interest score; At the same time, by estimating the product relevance level, it is possible to predict the interaction between the corresponding product search results and the user, such as whether the target user purchases the corresponding product search results, clicks on the corresponding product search results, or collects the corresponding product search results. Thus, by comprehensively searching the interactive behavior characteristics of the user, the search conditions, and the product relevance level of the recalled product, it is possible to predict the target user's interest in different product search results, and characterize the interest level corresponding to each product search result as an interest score. Then, the product search results can be sorted based on each interest score. Based on this, the present application can use a sorting model to comprehensively search for the application's search conditions and interactive behavior to predict the product relevance level, and then perform interest score scoring. The sorting of each product search result can be completed by the interest score. The amount of data required to be processed is the amount of data for each product search result, interactive behavior, and search conditions. Compared with the existing technology, the amount of data is relatively small. Moreover, scoring through the sorting model can process multiple product search results in parallel, reduce unnecessary calculation steps, reduce calculation complexity, further speed up the scoring calculation speed, and thus further speed up the overall sorting speed. It can be seen that the present application can speed up the sorting speed while ensuring the accuracy of the sorting results, thereby improving the user's experience.
[0077] In some embodiments of the present application, the process of step S1, obtaining a ranking model including an ordered estimation layer, is described in detail, and the steps are as follows:
[0078] S10. Obtain an initial ranking model for scoring products based on user transaction characteristics and search information.
[0079] Specifically, a trained initial ranking model may be obtained, where the initial ranking model may be trained based on a plurality of training samples including corresponding product training data, search information, search user transaction characteristics, and rating values.
[0080] S11. Construct the initial classification module.
[0081] Specifically, an initial classification module for performing classification may be constructed.
[0082] S12. Acquire multiple commodity samples marked with corresponding relevance levels, where each relevance level indicates a consumption intention probability of the corresponding commodity sample.
[0083] Specifically, a plurality of commodity samples may be obtained, wherein each commodity sample is marked with a corresponding relevance level.
[0084] The relevance level of each product sample can be used to represent the consumption intention probability of the corresponding product sample.
[0085] Product samples may include product titles, product links, product introductions, product categories, product keywords, and other product information related to the product.
[0086] S13. Use each product sample to train the initial classification module, and use the final initial classification module as the order estimation layer.
[0087] Specifically, each product sample can be input into the initial classification module in turn to obtain the corresponding prediction level of each product sample. Based on the prediction level and the corresponding product sample, the parameters of the initial classification module are adjusted. The final initial classification module is the sequence estimation layer.
[0088] S14. Add the order estimation layer to the initial sorting model to generate a sorting model including the order estimation layer.
[0089] Specifically, the final order estimation layer can be fused into the penultimate layer of the initial sorting model, and after fusion, a sorting model including the order estimation layer is obtained.
[0090] It can be seen from the above technical solution that this embodiment provides an optional way to obtain a sorting model that includes an ordered estimation layer. Through the above method, the present application can train the initial classification module and add the initial classification module to the initial sorting model. Based on this, the present application can transform the existing initial sorting model, and only needs to train the initial classification module to complete the construction of the sorting model, which further accelerates the construction speed of the model of the present application and simplifies the difficulty of constructing the sorting model of the present application.
[0091] In some embodiments of the present application, step S10, the process of obtaining an initial ranking model for scoring products based on user transaction characteristics and search information, is described in detail. The steps are as follows:
[0092] S100. Obtain a ranking module and multiple triples, each triple containing corresponding product training data, corresponding search information, and transaction characteristics of the corresponding search user, and each triple is annotated with a rating value generated based on the corresponding search information and the transaction characteristics of the corresponding search user.
[0093] Specifically, a neural network model can be constructed as a sorting module.
[0094] Each product training data and its corresponding search information and transaction features of the search user can be combined to form a triple, and each triple can be marked with a score value.
[0095] In addition to the rating value, each triplet can also be marked as strongly correlated, weakly correlated, or not correlated to indicate the degree of relevance between the product training data of each triplet and the corresponding search information.
[0096] S101. Using each triplet, iteratively train the ranking module. The resulting ranking module is an initial ranking model for scoring products based on user transaction characteristics and search information.
[0097] Specifically, the sorting module is iteratively trained based on each triple in turn, and the parameter values of the sorting module are adjusted in combination with the gradient descent method.
[0098] The resulting ranking module can be used as an initial ranking model for scoring products based on user transaction characteristics and search information.
[0099] It can be seen from the above technical solution that this embodiment provides an optional way to obtain an initial ranking model. By using the above method, the present application can obtain an initial ranking model through multiple triple training when there is no initial ranking model that can be used for scoring, thereby improving the wide application of the present application.
[0100] In some embodiments of the present application, the process of obtaining multiple product samples marked with corresponding relevance levels in step S12, where each relevance level indicates the consumption intention probability of the corresponding product sample, is described in detail. The steps are as follows:
[0101] S120: Acquire multiple product information and determine the relevance level corresponding to each product information based on the transaction volume, add-to-cart volume, favorite volume, pageview volume, and click volume corresponding to each product information.
[0102] Specifically, multiple product information may be obtained, and each product information may include any one or more of a product link, a product introduction, product keywords, a product category, a product purpose, and a product image of the corresponding product.
[0103] You can determine the transaction volume, add-to-cart volume, favorite volume, page views, and click volume of each product information.
[0104] Determine transaction weight based on the degree to which transaction volume represents the probability of consumption intention;
[0105] Determine the weight of the additional purchase based on the degree to which the additional purchase quantity represents the probability of consumption intention;
[0106] Determine the collection weight based on the degree to which the collection quantity represents the probability of consumption intention;
[0107] Determine the browsing weight based on the degree to which the browsing volume represents the probability of consumption intention;
[0108] Determine the click weight based on the degree to which the click volume represents the probability of consumption intention;
[0109] Calculate the interaction score based on the transaction volume and transaction weight, add-to-cart volume and add-to-cart weight, collection volume and collection weight and browsing weight, pageview volume and browsing weight, click volume and click weight of each product information;
[0110] Based on the preset grading range and the interaction score of each product information, the relevance level of each product information is determined.
[0111] In addition, you can also adjust the relevance level of each product information in combination with the refined sorting order.
[0112] S121. Use the relevance level of each product information as a label for the corresponding product information to form a product sample.
[0113] Specifically, the corresponding relevance level can be marked in each commodity information, and commodity samples can be obtained after marking.
[0114] It can be seen from the above technical solution that this embodiment provides an optional method for constructing multiple product samples. Through the above method, the relevance level of product information can be evaluated based on multiple dimensions such as transaction volume, added purchase volume, collection volume, page views and click volume, further providing the reliability of the relevance level of this application.
[0115] In some embodiments of the present application, the process of step S13, training the initial classification module using each product sample and using the resulting initial classification module as the order estimation layer, is described in detail as follows:
[0116] S130. Input each product sample into the initial classification module in turn to obtain the predicted level output by the initial classification module. The initial classification module includes a category tendency evaluation level and a content bias determination level. The category tendency evaluation level is used to evaluate the category tendency of each product sample, and the content bias determination level is used to determine the correspondence between each product sample and the multi-dimensional business indicators. The predicted level of each product sample is generated based on the corresponding category tendency and each corresponding relationship.
[0117] Specifically, the initial classification module can be used to predict each product sample in turn to generate a prediction grade.
[0118] The category tendency assessment layer of the initial classification module can be used to predict the category tendency of each input product sample, and the content bias determination layer of the initial classification module can be used to determine the correspondence between each product sample and different product interaction behaviors and / or different indicator rankings;
[0119] The initial classification module is used to comprehensively consider the category tendency of each product sample, the corresponding relationship between the interactions between each product, and the corresponding relationship between the ranking of each indicator to generate a predicted score for the corresponding product sample, and based on the predicted score of the corresponding product, generate a predicted level for the corresponding product sample.
[0120] S131. Calculate the cross entropy loss based on the correlation level of each prediction level and its corresponding product sample.
[0121] Specifically, the cross entropy loss function can be combined to calculate the cross entropy loss based on the relevance level of each product sample and its corresponding prediction level.
[0122] S132. Based on each cross entropy loss, combined with the gradient descent method, the parameters of the initial classification module are optimized until the cross entropy loss meets the preset stopping condition or the product sample iteration is completed. The final trained initial classification module with the content bias judgment layer removed is used as the order estimation layer.
[0123] Specifically, based on the cross-entropy loss of each product sample, the parameters of the initial classification module can be adjusted in combination with the gradient descent method until multiple consecutive cross-entropy losses are less than the preset cross-entropy loss threshold. The final initial classification module is used as the order estimation layer.
[0124] It can be seen from the above technical solution that this embodiment provides an optional method for training the initial classification module, and the above method can be used to further improve the reliability of the sequence estimation layer of this application through training.
[0125] In some embodiments of the present application, the process of adding the order estimation layer to the initial sorting model to generate a sorting model including the order estimation layer in step S14 is described in detail, and the steps are as follows:
[0126] S140. Link the target increasing function to the output end of the order estimation layer, and add the order estimation layer linked to the target increasing function to the front end of the fully connected layer of the initial sorting model to obtain a sorting model including the order estimation layer.
[0127] Specifically, an increasing function that can be connected to the neural network model can be determined as the target increasing function.
[0128] For example, the target increasing function may be a linear function, a logistic growth function, an inverse function of a logarithmic function, or a sigmoid function.
[0129] Specifically, the original weight parameter of the penultimate layer of the initial sorting model may be W. When the original input of the penultimate layer is x, the original output of the penultimate layer is Wx. The original weight parameter of the penultimate layer of the initial sorting model is used as the shared parameter of the order estimation layer.
[0130] The ordinal prediction layer estimates the probability of each correlation level as f(Wx+bias_i), where bias_i can be a private parameter of the i-th correlation level.
[0131] It can be seen from the above technical solution that this embodiment provides an optional method of adding the order estimation layer to the initial sorting model to generate a sorting model containing the order estimation layer. Through the above method, the order estimation layer and the initial sorting model can be directly fused through the target increasing function, further simplifying the difficulty of fusing the order estimation layer and the initial sorting model.
[0132] During the application process, the ranking model was validated both offline and online, resulting in a significant increase in online conversion rates by 0.42% and a significant increase in average product clicks by 0.49%. Offline and online model metrics used are Kendall and GAUC. Kendall indicates the degree to which the coarse ranking model fits the fine ranking model. The specific metric is the ratio of Kendall (the positive ordinal number of the coarse ranking relative to the fine ranking - the negative ordinal number of the coarse ranking relative to the fine ranking) to the total ordinal number, with a value between -1 and 1. GAUC measures the model's fit to user preference labels. Whether a user places an order can be used as a label, and each request can be grouped. The AUC within a group is calculated as GAUC.
[0133] Whether offline or online, the kendall and GAUC corresponding to the search result ranking method of this application are higher than those of the control group.
[0134] Furthermore, in some embodiments of the present application, the target increasing function may be a Sigmoid function.
[0135] Next, we will combine Figure 2 The search result sorting device provided in this application is introduced in detail. The search result sorting device provided below can be compared with the search result sorting method provided above.
[0136] See also Figure 2 It can be found that the search result sorting device may include:
[0137] An acquisition module 10 is configured to acquire a ranking model including an ordered estimation layer, wherein the ordered estimation layer is trained based on a plurality of commodity samples labeled with corresponding relevance levels;
[0138] Search module 20, used to obtain multiple product search results that match the search conditions input by the target user;
[0139] Scoring module 30, configured to input the target user's interactive behavior, the search criteria, and each product search result into the ranking model, evaluate the product relevance level of each product search result using the order estimation layer, and score each product search result using the ranking model in combination with the product relevance level of each product search result, the search criteria, and the target user's interactive behavior to obtain an interest score for each product search result;
[0140] The sorting module 40 is configured to sort the product search results based on the interest scores.
[0141] Furthermore, the acquisition module 10 may include:
[0142] An initial ranking model acquisition unit, configured to acquire an initial ranking model for scoring products based on user transaction characteristics and search information;
[0143] An initial classification module unit, used for constructing an initial classification module;
[0144] A product sample acquisition unit, configured to acquire a plurality of product samples labeled with corresponding relevance levels, each relevance level indicating a probability of consumption intention for the corresponding product sample;
[0145] An initial classification module training unit, configured to train the initial classification module using each product sample, and use the resulting initial classification module as a sequence estimation layer;
[0146] The sorting model generating unit is configured to add the order estimation layer to the initial sorting model to generate a sorting model including the order estimation layer.
[0147] Furthermore, the initial sorting model acquisition unit may include:
[0148] A first initial ranking model acquisition subunit is configured to acquire a ranking module and a plurality of triples, each triple containing corresponding product training data, corresponding search information, and transaction characteristics of the corresponding search user, and each triple is annotated with a score value generated based on the corresponding search information and transaction characteristics of the corresponding search user;
[0149] The second initial ranking model acquisition subunit is used to iteratively train the ranking module using each triple. The ranking module finally obtained is an initial ranking model for scoring products according to user transaction characteristics and search information.
[0150] Furthermore, the product sample acquisition unit may include:
[0151] The first product sample acquisition subunit is used to obtain multiple product information and determine the relevance level corresponding to each product information based on the transaction volume, added purchase volume, favorite volume, pageview volume, and click volume corresponding to each product information;
[0152] The second product sample acquisition subunit is configured to use the relevance level of each product information as a label for the corresponding product information to form a product sample.
[0153] Furthermore, the initial classification module training unit may include:
[0154] The first initial classification module training subunit is used to sequentially input each product sample into the initial classification module to obtain a predicted grade output by the initial classification module. The initial classification module includes a category tendency assessment layer and a content bias determination layer. The category tendency assessment layer is used to assess the category tendency of each product sample, and the content bias determination layer is used to determine the correspondence between each product sample and multi-dimensional business indicators. The predicted grade of each product sample is generated based on the corresponding category tendency and each corresponding relationship.
[0155] The second initial classification module training subunit is used to calculate the cross entropy loss based on the correlation level of each predicted level and its corresponding product sample;
[0156] The third initial classification module training subunit is used to optimize the parameters of the initial classification module based on each cross-entropy loss in combination with the gradient descent method until the cross-entropy loss meets the preset stopping condition or the product sample iteration is completed. The final trained initial classification module with the content bias judgment layer removed is used as the order estimation layer.
[0157] Furthermore, the ranking model generating unit may include:
[0158] The target increasing function linking subunit is used to link the target increasing function to the output end of the order estimation layer, and add the order estimation layer linked to the target increasing function to the front end of the fully connected layer of the initial sorting model to obtain a sorting model including the order estimation layer.
[0159] Furthermore, the target increment function linking subunits may include:
[0160] The target increasing function determination component is used to link the Sigmoid function to the output end of the order estimation layer.
[0161] The search result sorting device provided in the embodiment of the present application can be applied to search result sorting devices, such as PC terminals, cloud platforms, servers and server clusters. Figure 3 The hardware structure diagram of the search result sorting device is shown. Figure 3 ,The hardware structure of the search result ranking device may include: at least one processor 1, at least one communication interface 2, at least one memory 3 and at least one communication bus 4;
[0162] In the embodiment of the present application, the number of the processor 1, the communication interface 2, the memory 3, and the communication bus 4 is at least one, and the processor 1, the communication interface 2, and the memory 3 communicate with each other through the communication bus 4;
[0163] The processor 1 may be a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention;
[0164] The memory 3 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory;
[0165] The memory stores a program, and the processor can call the program stored in the memory, wherein the program is used to:
[0166] Obtaining a ranking model including an ordered estimation layer, wherein the ordered estimation layer is trained based on a plurality of product samples labeled with corresponding relevance levels;
[0167] Obtain multiple product search results that match the search criteria entered by the target user;
[0168] Inputting the target user's interactive behavior, the search conditions, and each product search result into the ranking model, using the order estimation layer to evaluate the product relevance level of each product search result, and using the ranking model to combine the product relevance level of each product search result, the search conditions, and the target user's interactive behavior to score each product search result to obtain an interest score for each product search result;
[0169] Sort the product search results based on their interest scores.
[0170] Optionally, the refined functions and extended functions of the program may refer to the above description.
[0171] The present application also provides a readable storage medium, which may store a program suitable for execution by a processor, wherein the program is used to:
[0172] Obtaining a ranking model including an ordered estimation layer, wherein the ordered estimation layer is trained based on a plurality of product samples labeled with corresponding relevance levels;
[0173] Obtain multiple product search results that match the search criteria entered by the target user;
[0174] Inputting the target user's interactive behavior, the search conditions, and each product search result into the ranking model, using the order estimation layer to evaluate the product relevance level of each product search result, and using the ranking model to combine the product relevance level of each product search result, the search conditions, and the target user's interactive behavior to score each product search result to obtain an interest score for each product search result;
[0175] Sort the product search results based on their interest scores.
[0176] Optionally, the refined functions and extended functions of the program may refer to the above description.
[0177] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.
[0178] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.
[0179] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. The various embodiments of the present application may be combined with each other. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A search result sorting method, characterized in that: include: Obtaining a ranking model including an ordered estimation layer, wherein the ordered estimation layer is trained based on a plurality of product samples labeled with corresponding relevance levels; Obtain multiple product search results that match the search criteria entered by the target user; Inputting the target user's interactive behavior, the search conditions, and each product search result into the ranking model, using the order estimation layer to evaluate the product relevance level of each product search result, and using the ranking model to combine the product relevance level of each product search result, the search conditions, and the target user's interactive behavior to score each product search result to obtain an interest score for each product search result; Sort the product search results based on their interest scores.
2. The search result ranking method according to claim 1, characterized in that: The obtaining of the sorting model including the ordered estimation layer includes: Obtaining an initial ranking model for scoring products based on user transaction characteristics and search information; Construct an initial classification module; Obtain multiple product samples marked with corresponding relevance levels, each relevance level indicating a probability of consumption intention for the corresponding product sample; Using each product sample to train the initial classification module, and using the final initial classification module as the order estimation layer; The order estimation layer is added to the initial sorting model to generate a sorting model including the order estimation layer.
3. The search result ranking method according to claim 2, characterized in that: The obtaining of an initial ranking model for scoring products based on user transaction characteristics and search information includes: Obtaining a ranking module and multiple triples, each triple containing corresponding product training data, corresponding search information, and transaction characteristics of the corresponding search user, and each triple annotated with a relevance rating value generated based on the corresponding search information and the transaction characteristics of the corresponding search user; The ranking module is iteratively trained using each triplet, and the resulting ranking module is an initial ranking model for scoring products based on user transaction characteristics and search information.
4. The search result ranking method according to claim 2, characterized in that: The obtaining of multiple product samples marked with corresponding relevance levels includes: Obtain multiple product information and determine the relevance level corresponding to each product information based on the corresponding transaction volume, add-to-cart volume, favorite volume, pageview volume, and click volume of each product information; The relevance level of each product information is used as a label for the corresponding product information to form a product sample.
5. The search result ranking method according to claim 2, characterized in that: The initial classification module is trained using each product sample, and the resulting initial classification module is used as the order estimation layer, including: Each product sample is sequentially input into the initial classification module to obtain the predicted grade output by the initial classification module. The initial classification module includes a category tendency assessment layer and a content bias determination layer. The category tendency assessment layer is used to assess the category tendency of each product sample, and the content bias determination layer is used to determine the correspondence between each product sample and multi-dimensional business indicators. The predicted grade of each product sample is generated based on the corresponding category tendency and each corresponding relationship. Calculate the cross entropy loss based on the correlation level of each predicted level and its corresponding product sample; Based on each cross-entropy loss, combined with the gradient descent method, the parameters of the initial classification module are optimized until the cross-entropy loss meets the preset stopping condition or the product sample iteration is completed. The final trained initial classification module with the content bias judgment layer removed is used as the order estimation layer.
6. The search result ranking method according to claim 2, wherein: Adding the order estimation layer to the initial sorting model to generate a sorting model including the order estimation layer includes: The target increasing function is linked to the output end of the order estimation layer, and the order estimation layer linked to the target increasing function is added to the front end of the fully connected layer of the initial sorting model to obtain a sorting model including the order estimation layer.
7. The search result ranking method according to claim 6, characterized in that: The step of linking the target increasing function to the output end of the order estimation layer comprises: A Sigmoid function is connected to the output end of the order estimation layer.
8. A search result sorting device, characterized in that: include: an acquisition module, configured to acquire a ranking model comprising an ordered estimation layer, wherein the ordered estimation layer is trained based on a plurality of commodity samples labeled with corresponding relevance levels; A search module is used to obtain multiple product search results that match the search criteria entered by the target user; a scoring module for inputting the target user's interactive behavior, the search conditions, and each product search result into the ranking model, using the order estimation layer to evaluate the product relevance level of each product search result, and using the ranking model to combine the product relevance level of each product search result, the search conditions, and the target user's interactive behavior to score each product search result to obtain an interest score for each product search result; The sorting module is used to sort the product search results based on the interest scores.
9. A search result sorting device, characterized in that: including memory and processor; The memory is used to store programs; The processor is configured to execute the program to implement each step of the search result ranking method according to any one of claims 1 to 7.
10. A readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, each step of the search result ranking method according to any one of claims 1 to 7 is implemented.