Material recommendation method, device, electronic device and computer-readable storage medium

Abstract

The present application provides a material recommendation method, device, electronic device and computer-readable storage medium, the method comprising: obtaining the historical behavior characteristics of the user to be recommended for different types of materials; for each material in any type of material, determining multiple similar materials corresponding to the material according to the historical behavior characteristics of the user to be recommended; for any type of material, determining the preference of the user to be recommended for the type of material; according to the preference of the user to be recommended for each type of material, determining a candidate material set from multiple similar materials corresponding to each type of material; inputting the candidate materials and the user characteristics corresponding to the user to be recommended into a prediction model to obtain the predicted click-through rate for different candidate materials; and recommending to the user to be recommended according to the predicted click-through rate of each candidate material. The method can enable users to browse different types of materials of interest, thereby improving user experience.

Description

Technical Field

[0001] The present application relates to the technical field of information recommendation, and in particular to a material recommendation method, device, electronic device and computer-readable storage medium. Background Art

[0002] With the continuous development of Internet technology, network platforms can push various materials to users so that users can browse and thus meet user needs.

[0003] In the related art, there are various types of materials, such as commodity types, non-commodity types, etc. However, network platforms usually only recommend a single type of material, which can easily cause user fatigue and reduce user experience. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide a material recommendation method, device, electronic device and computer-readable storage medium, which are used to determine materials suitable for recommendation to users, so that users can browse different types of materials of interest, thereby improving user experience.

[0005] In the first aspect, the embodiment of the present application provides a material recommendation method, which includes: obtaining the historical behavior characteristics of the user to be recommended for different types of materials; for each material in any type of material, determining multiple similar materials corresponding to the material according to the historical behavior characteristics of the user to be recommended; for any type of material, determining the preference of the user to be recommended for the type of material; according to the preference of the user to be recommended for each type of material, determining a candidate material set from multiple similar materials corresponding to each type of material; for any candidate material in the candidate material set, inputting the candidate material and the user characteristics corresponding to the user to be recommended into a prediction model to obtain a predicted click rate for the candidate material; and recommending to the user to be recommended according to the predicted click rates corresponding to each candidate material. In this way, materials suitable for recommendation to users can be determined from multiple types of materials, so that users can browse different types of materials of interest, thereby improving user experience.

[0006] Optionally, before inputting the candidate material and the user features corresponding to the to-be-recommended user into the prediction model for any candidate material in the candidate material set to obtain the predicted click rate for the candidate material, the method further includes: obtaining the prediction model; wherein the prediction model is pre-trained according to the following steps: obtaining sample materials and sample users; the sample materials include positive sample materials and negative sample materials, the positive sample materials include materials that have been exposed and clicked by sample users, and the negative sample materials include materials that have been exposed but not clicked by sample users; for any sample material, obtaining the sample material features corresponding to the sample material and the sample users who clicked on the sample material; The sample user characteristics corresponding to this user; and, according to the type of the sample material, the sample material characteristics and sample user characteristics corresponding to the sample material are processed by the click rate prediction submodule used to process the sample material of this type in the initial prediction model to obtain the corresponding predicted click rate; for any click rate prediction submodule, determine the loss value between the predicted click rate output by the click rate prediction submodule for different sample materials in the type of material it processes and the corresponding actual click rate; according to the loss values ​​corresponding to each click rate prediction submodule, determine the overall loss value of the initial prediction model, and obtain the prediction model when the overall loss value meets the preset loss requirement. In this way, the prediction model can be used to output the predicted click rates of various types of materials, so as to facilitate the recommendation of different types of materials to the recommended users.

[0007] Optionally, for any sample material, obtaining the sample material features corresponding to the sample material and the sample user features corresponding to the sample user who clicks the sample material includes: for any sample material, obtaining the cross-features, material attribute features and material statistical features corresponding to the sample material; wherein the cross-features characterize the interactive features of the sample material in the click behavior sequence of the sample user who clicks the sample material; and the material statistical features characterize the statistically derived data features related to the sample material. In this way, not only the basic features of the sample material are obtained, but also the cross-features related to the click behavior of the sample material and the sample user are obtained, so that the personalized preferences of the sample user can be better captured, and the prediction accuracy of the prediction model for different recommended users is improved to a certain extent.

[0008] Optionally, for any sample material, obtaining the sample material features corresponding to the sample material and the sample user features corresponding to the sample user who clicks the sample material includes: for any sample material, obtaining the individual behavior features corresponding to the sample user who clicks the sample material, and the overall user attribute features and overall user statistical features corresponding to multiple sample users; using the weighted submodule in the initial prediction model, obtaining the sample user features according to the individual behavior features, the overall user attribute features, and the overall user statistical features. In this way, on the one hand, the prediction model can more accurately capture the actual interests and needs of the recommended users by identifying the common features between the individual behavior features of the users and the common features of the group to which they belong. On the other hand, for some new recommended users or recommended users with less individual behavior data, by referring to the common features of their group, the data sparseness defects caused by insufficient personal data can be compensated, and more reasonable recommendations can be provided. Further, using the common features can not only strengthen the materials of interest known to the recommended users, but also guide the recommended users to discover new points of interest, predict their potential needs, and reduce irrelevant or repeated recommendations, thereby improving user satisfaction and user experience.

[0009] Optionally, for any type of material, determining the preference of the user to be recommended for the type of material includes: for any type of material, determining multiple historical behavior characteristics of the user to be recommended for the type of material; and determining the preference of the user to be recommended for the type of material according to the multiple historical behavior characteristics and the preference weights corresponding to each historical behavior characteristic. In this way, the user's preference can be captured more accurately, the accuracy of the preference degree can be improved, and the accuracy of the recommendation can be improved to a certain extent.

[0010] Optionally, the step of determining a candidate material set from a plurality of similar materials corresponding to each type of material according to the preference of the user to be recommended to each type of material includes: for any type of material, determining the number of candidate materials to be selected from a plurality of similar materials corresponding to each material of the type of material according to the preset number of candidate materials in the candidate material set and the preference of the user to be recommended to the type of material; and, according to the number of candidate materials to be selected, selecting a corresponding number of candidate materials from a plurality of similar materials corresponding to each material of the type of material to obtain the candidate material set. In this way, by processing a preset number of candidate materials and selecting a corresponding number of candidate materials from similar materials corresponding to different types of materials according to the interest of the user to be recommended, it is possible to control the diversity of recommendation information and the quality of recommendation, improve the accuracy of recommendation, and thus improve the user experience.

[0011] In the second aspect, the embodiment of the present application provides a material recommendation device, which includes: an acquisition module for acquiring the historical behavior characteristics of the user to be recommended for different types of materials; a similar material determination module for determining, for each material in any type of material, multiple similar materials corresponding to the material according to the historical behavior characteristics of the user to be recommended; a preference degree determination module for determining, for any type of material, the preference degree of the user to be recommended for the material of that type; a candidate material set determination module for determining a candidate material set from multiple similar materials corresponding to each type of material according to the preference degree of the user to be recommended for each type of material; a prediction module for inputting the candidate material and the user characteristics corresponding to the user to be recommended into a prediction model for any candidate material in the candidate material set to obtain a predicted click rate for the candidate material; and a recommendation module for recommending to the user to be recommended according to the predicted click rates corresponding to each candidate material. In this way, materials suitable for recommendation to users can be determined from multiple types of materials, so that users can browse different types of materials of interest, thereby improving user experience.

[0012] In a third aspect, an embodiment of the present application provides an electronic device, comprising a processor and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the steps in the method provided in the first aspect are performed.

[0013] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps in the method provided in the first aspect are performed.

[0014] In a fifth aspect, an embodiment of the present application provides a computer program product, which includes a computer program or instructions, and when the computer program or instructions are executed by a processor, the method described in the first aspect is performed.

[0015] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or be understood by practicing the embodiments of the present application. The purpose and other advantages of the present application can be realized and obtained by the structures specifically pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 A flow chart of a material recommendation method provided in an embodiment of the present application;

[0018] Figure 2 A schematic diagram of the structure of an initial prediction model provided in an embodiment of the present application;

[0019] Figure 3 A structural block diagram of a material recommendation device provided in an embodiment of the present application;

[0020] Figure 4 A schematic diagram of the structure of an electronic device for executing a material recommendation method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0022] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0023] It should be noted that the embodiments in this application or the technical features in the embodiments may be combined if there is no conflict.

[0024] In the related art, network platforms usually only recommend a single type of material, which can easily cause user fatigue and reduce user experience. In order to solve this problem, the present application provides a material recommendation method, device, electronic device and computer-readable storage medium. Furthermore, according to the preference of the user to be recommended for different types of materials, multiple candidate materials are screened from similar materials of different types of materials of interest, and then recommended to them according to the predicted click-through rate of each candidate material. In this way, among multiple types of materials, materials suitable for recommendation to users can be determined, so that users can browse different types of materials of interest, which improves user experience.

[0025] It should be noted that the defects existing in the solutions in the above-mentioned related technologies are the results obtained by the inventors after practice and careful research. Therefore, the discovery process of the above-mentioned problems and the solutions proposed in the embodiments of the present invention below for the above-mentioned problems should all be the contributions made by the inventors to the present invention in the process of the present invention.

[0026] Please refer to Figure 1 , which shows a flow chart of a material recommendation method provided in an embodiment of the present application. Figure 1 As shown, the material recommendation method includes the following steps 101 to 106.

[0027] Step 101, obtaining historical behavior characteristics of the user to be recommended for different types of materials;

[0028] Among them, different types of materials may include, for example, commodity-type materials and non-commodity-type materials involved in the e-commerce community model. The above non-commodity-type materials may include, for example, video materials and / or graphic materials that are not related to commodities.

[0029] The above-mentioned historical behavior characteristics are also the behavior characteristics that the user to be recommended has performed on different types of materials in the history. For example, it may include the click behavior, browsing behavior, browsing depth (the number of materials exposed during browsing), browsing time, etc. of the type of material within a preset time period. Furthermore, since the user to be recommended can perform multiple behaviors, the historical behavior characteristics can be sequence characteristics, such as historical click sequence characteristics, historical browsing sequence characteristics, etc.

[0030] It should be noted that the network platform obtains the historical behavior characteristics of the user to be recommended only if the user agrees to obtain the historical behavior characteristics. In some application scenarios, for example, the network platform can display a request page for obtaining the historical behavior characteristics of the user to be recommended, and obtain the corresponding historical behavior characteristics only if the user to be recommended clicks a button indicating consent to obtain the historical behavior characteristics.

[0031] Step 102, for each material of any type, determine multiple similar materials corresponding to the material according to the historical behavior characteristics of the user to be recommended;

[0032] In some application scenarios, the network platform can, for example, obtain a similarity determination model, and then input the historical click sequence characteristics of the user to be recommended into the similarity determination model, so as to determine the similarity between other materials to be recommended and each material that has been clicked in history through the similarity determination model. Then, for any material that has been clicked in history, multiple materials to be recommended can be arranged in sequence according to the similarity between other materials to be recommended and the material, and then the first N materials to be recommended with higher similarity can be determined as similar materials corresponding to the material. Alternatively, multiple materials to be recommended with similarity higher than a preset similarity can all be determined as similar materials corresponding to the material.

[0033] The above-mentioned N value or preset similarity can be determined according to actual needs, and this application does not limit this. For example, the N value can include 10, 20, etc., and the preset similarity can include 92%, 95%, etc. Further, the above-mentioned similarity determination model can be modeled by, for example, the ItemCF algorithm (item-based collaborative filtering recommendation algorithm).

[0034] In other application scenarios, for example, the materials with higher click frequencies in the historical click sequences of the users to be recommended may be determined first, and then for any material with higher click frequencies, the similarity between other materials to be recommended and the material may be determined based on the types of the other materials to be recommended, the key information contained, etc. For example, when the types are the same, the higher the similarity between other materials to be recommended with the same or similar key information and the material, the more similar materials that are similar to the current material may be determined.

[0035] Step 103, for any type of material, determining the preference of the user to be recommended for the type of material;

[0036] The above-mentioned preference degree is also the interest degree, which can be characterized by preference score, interest level, etc. in some application scenarios. In these application scenarios, the preference score or interest level can be determined by the click rate, browsing time and / or repeated click number of the type of material by the user to be recommended, for example. For example, the click rate or repeated click number can be directly determined as the preference score.

[0037] Step 104, determining a candidate material set from a plurality of similar materials corresponding to each type of material according to the preference of the user to be recommended to each type of material;

[0038] The candidate material set includes multiple candidate materials, each of which may be recommended to the user to be recommended or may not be recommended.

[0039] In some application scenarios, the network platform may, for example, determine candidate materials from multiple similar materials corresponding to a certain type of material with a higher preference score based on the preference score, and then determine candidate materials from multiple similar materials corresponding to a certain type of material with a lower preference score.

[0040] It should be noted that when the network platform determines the candidate material from multiple similar materials, it can be determined randomly or based on a certain rule, which is not limited here. The above rule can be, for example, screening in sequence according to the click rate. In addition, the number of candidate materials in the candidate material set can be a preset fixed number or a dynamically changing number, which is not limited here either.

[0041] Step 105, for any candidate material in the candidate material set, the candidate material and the user characteristics corresponding to the to-be-recommended user are input into a prediction model to obtain a predicted click rate for the candidate material;

[0042] The above prediction model may include, for example, a deep neural network model, a logistic regression model, etc.

[0043] The above user characteristics may include some or all characteristics related to the user to be recommended, which may include, for example, user attribute characteristics such as gender, age, device used, city, etc., and may further include user statistical characteristics such as activity statistics, purchasing power statistics, etc., and may further include the above historical behavior characteristics.

[0044] Then, the network platform can input some or all of the user-related features and candidate materials into the prediction model to obtain the predicted click rate of the user to be recommended for each candidate material.

[0045] It should be noted that in some application scenarios, for example, a user profile may be first constructed based on user features, and then the user profile and each candidate material may be input into the prediction model, thereby obtaining the corresponding predicted click rate.

[0046] Step 106: Recommend the candidate materials to the user to be recommended according to the predicted click-through rates of the candidate materials.

[0047] In some application scenarios, the network platform may, for example, recommend the candidate material with the highest predicted click-through rate to the user to be recommended, or may recommend any candidate material among several candidate materials with higher predicted click-through rates to the user to be recommended, without limitation here.

[0048] In this implementation, materials suitable for recommendation to users can be determined from multiple types of materials, so that users can browse different types of materials of interest, thereby improving user experience.

[0049] In some optional implementations, the step 103 of determining the degree of preference of the user to be recommended for any type of material may include the following sub-steps:

[0050] Sub-step 1031, for any type of material, determining multiple historical behavior characteristics of the to-be-recommended user for the type of material;

[0051] That is, determine the behavior-related characteristics of the user to be recommended for this type of material, such as browsing depth, browsing time, click-through rate, number of clicks, etc.

[0052] Sub-step 1032, determining the preference degree of the to-be-recommended user for the type of material according to the multiple historical behavior characteristics and the preference weights corresponding to each historical behavior characteristic.

[0053] The preference weights corresponding to the above-mentioned historical behavior characteristics can be preset by the network platform, for example. For example, the network platform can set the preference weights according to the influence of each historical behavior characteristic on the preference degree. For example, the click-through rate has a greater influence on determining the preference degree, so a higher weight can be set for the click-through rate, while the browsing depth has a relatively small influence on determining the preference degree, so a lower weight can be set for it.

[0054] In some application scenarios, the preference degree can be determined by a calculation formula, which can be, for example: S i =w1×imprCnt i +w2×imprTime i +w3×ctr i +w4×clkCnt i +…+w n ×M i ; Among them, S represents the degree of preference (such as preference score), i represents the type index (such as product type or non-product type), w represents the preference weight, imprCnt represents the browsing depth, imprTime represents the browsing time, ctr represents the click-through rate, clkCnt represents the number of clicks, n represents the weight index, and M represents other historical behavior characteristics.

[0055] In this implementation, the preference degree for each type of material can be determined based on the multiple historical behavioral characteristics of the user to be recommended and their corresponding preference weights for each type of material. This can more accurately capture the user's preferences, improve the accuracy of the preference degree, and to a certain extent improve the accuracy of the recommendation.

[0056] In some application scenarios, the network platform can determine candidate material sets in batches to facilitate the management of candidate materials. Then, in some optional implementations, the step 104 described above, according to the preference of the user to be recommended for each type of material, determines the candidate material set from multiple similar materials corresponding to each type of material, including:

[0057] Sub-step 1041, for any type of material, according to the preset number of candidate materials in the candidate material set and the preference of the user to be recommended for the type of material, determine the number of candidate materials to be selected from multiple similar materials corresponding to each material in the type of material;

[0058] The above-mentioned preset number may include, for example, 1000, 2000, etc., which can substantially meet the number required by the user.

[0059] The candidate materials to be selected are similar materials to be selected as candidate materials among multiple similar materials corresponding to various types of materials.

[0060] Then, when the number of candidate materials in the candidate material set is fixed, the network platform can first determine the number of candidate materials that can be screened out from multiple similar materials corresponding to various types of materials. For example, the candidate material set allows 1,000 candidate materials to be accommodated, and the multiple similar materials corresponding to each commodity material in the commodity type are 3,000, and the multiple similar materials corresponding to each non-commodity material in the non-commodity type are 2,000. At this time, the network platform can determine that the number of candidate materials that can be subsequently determined from the 3,000 similar materials corresponding to the commodity type is 700 according to the degree of preference corresponding to the commodity type material. And, the number of candidate materials that can be subsequently determined from the 2,000 similar materials corresponding to the non-commodity type is 300.

[0061] In some application scenarios, for example, the number of candidate materials to be selected corresponding to each type of material can be determined by a softmax function. Furthermore, the calculation formula calculated by the softmax function can be, for example: Among them, mergeCnt represents the number of candidate materials to be selected, TotalCnt represents the preset number of candidate materials in the candidate material set, S represents the degree of preference (such as preference score), i represents the type index, and e represents the base of the natural logarithm.

[0062] In other application scenarios, if the calculated preference score is a positive number, for example, the number of candidate materials to be selected corresponding to each type of material can be determined by the proportional relationship between the preference scores. Among them, a represents the number of types, and the meanings of other parameters are the same as above and will not be repeated here.

[0063] Sub-step 1042, based on the number of candidate materials to be selected, select a corresponding number of multiple candidate materials from multiple similar materials corresponding to each material of the type of material, to obtain the candidate material set.

[0064] Continuing with the above example, after the network platform determines that 700 similar materials can be selected from 3,000 similar materials corresponding to the commodity type and 300 similar materials can be selected from 2,000 similar materials corresponding to the non-commodity type, the corresponding number of candidate materials can be screened out from the similar materials of the corresponding type.

[0065] In some application scenarios, for example, the top 700 similar materials with the highest click rate or number of clicks can be screened out from 3,000 similar materials corresponding to the commodity type, and the top 300 similar materials with the highest click rate or number of clicks can be screened out from 2,000 similar materials corresponding to the non-commodity type. Thus, a candidate material set of 1,000 can be obtained.

[0066] In this implementation, by processing a preset number of candidate materials and selecting a corresponding number of candidate materials from similar materials corresponding to different types of materials according to the degree of interest of the users to be recommended, it is possible to easily control the diversity of recommendation information and the quality of recommendation, improve the accuracy of recommendation, and thus improve the user experience.

[0067] It should be noted that, when the network platform screens candidate materials, in addition to the screening methods mentioned in the above implementation methods, for example, the network platform can also screen out candidate materials based on the degree of preference and the number of similar materials to improve the flexibility of screening. For example, when the number of similar materials corresponding to each type of material is relatively balanced (for example, more or less), the network platform can determine more candidate materials from multiple similar materials corresponding to a certain type of material with a higher preference score, and determine fewer candidate materials from multiple similar materials corresponding to a certain type of material with a lower preference score. In the case where the number of similar materials corresponding to each type of material is unbalanced, for example, when a certain type of material with a higher preference score corresponds to fewer similar materials, and a certain type of material with a lower preference score corresponds to more similar materials, similar materials corresponding to a certain type of material with a higher preference score can also be determined as candidate materials, and some candidate materials can be screened out from similar materials corresponding to a certain type of material with a lower preference score.

[0068] In some optional implementations, for any candidate material in the candidate material set described in step 105 above, the candidate material and the user characteristics corresponding to the to-be-recommended user are input into a prediction model to obtain a predicted click rate for the candidate material, and before the method further includes: obtaining the prediction model.

[0069] It should be noted that the above prediction model can be pre-trained in the network platform, or trained in other devices or platforms, and this application does not limit this. Then, the network platform can, for example, obtain the prediction model from the location where it stores the prediction model, or obtain the prediction model from other devices or platforms.

[0070] In some application scenarios, taking training in a network platform as an example, the prediction model can be trained in advance according to the following steps:

[0071] Step 1, obtaining sample materials and sample users; the sample materials include positive sample materials and negative sample materials, the positive sample materials include materials that have been exposed and clicked by sample users, and the negative sample materials include materials that have been exposed but not clicked by sample users;

[0072] It should be noted that in actual application scenarios, there are often materials that are exposed but not clicked by users. Therefore, in order to be more in line with actual application scenarios, the network platform can obtain positive samples and negative samples to jointly train the prediction model, thereby improving the confidence of the prediction model.

[0073] Step 2: for any sample material, obtain the sample material features corresponding to the sample material and the sample user features corresponding to the sample user who clicks the sample material;

[0074] The above-mentioned sample material characteristics are also characteristics related to the sample materials. In some application scenarios, for example, they may include material attribute characteristics and / or material statistical characteristics corresponding to the sample materials. Among them, the material attribute characteristics may be determined according to its type. For example, for commodity-type materials, its material attribute characteristics may include price, category, brand, series, and other characteristics related to its attributes; material statistical characteristics may include the number of clicks, number of favorites, number of purchases, and other data characteristics related to the material that need to be counted. In addition, for non-commodity-type materials, its material attribute characteristics may include content authors, content categories, content tags, and other characteristics related to its attributes; material statistical characteristics may include the number of clicks, likes, and other data characteristics related to the material that need to be counted.

[0075] The sample user characteristics are characteristics related to the sample users, which may include, for example, user attribute characteristics, user statistical characteristics and / or user behavior characteristics corresponding to the sample users.

[0076] In addition, the sample materials and sample users may be, for example, historical materials that have been truly recommended and users who have truly recommended materials, or constructed materials and users, and this application does not limit this. If both are real materials and real users, then both are obtained only after the user indicates that they agree to obtain them.

[0077] Step 3: According to the type of the sample material, the click rate prediction submodule for processing the sample material of the type in the initial prediction model is used to process the sample material characteristics and sample user characteristics corresponding to the sample material to obtain the corresponding predicted click rate;

[0078] That is to say, the above-mentioned initial prediction model may include a variety of click-through rate prediction submodules, and different click-through rate prediction submodules can be used to process different types of sample materials. In addition, the initial prediction model may further include submodules with other functions, and these submodules may include, for example, submodules for extracting high-dimensional feature vectors. It should be noted that in the initial prediction model, in order to facilitate the processing of various feature data, various feature data are usually converted into vectors for processing, and then it may include a submodule for extracting high-dimensional feature vectors.

[0079] It should be noted that each submodule in the initial prediction model may also be a deep neural network model, for example, so as to achieve the corresponding function.

[0080] Furthermore, the network platform can input the sample material characteristics and sample user characteristics corresponding to each sample material into the click rate prediction submodule, so as to obtain the corresponding predicted click rate.

[0081] Step 4: for any click rate prediction submodule, determine the loss value between the predicted click rates output by the click rate prediction submodule for different sample materials in the type of material processed by the click rate prediction submodule and the corresponding actual click rates.

[0082] It should be noted that the number of click-through rate prediction submodules in the initial prediction model can be the same as the number of sample material types. For example, if there are two types of materials, commodity type and non-commodity type, the initial prediction model can include two click-through rate prediction submodules, which are used to process the two types of materials respectively.

[0083] Then, for any type of click-through rate prediction submodule, the click-through rate prediction submodule can be used to output the predicted click-through rates corresponding to different materials in the type of materials it processes, and then the loss value between each predicted click-through rate and the corresponding actual click-through rate can be determined. For example, for the click-through rate prediction submodule used to process commodity type materials, the click-through rate prediction submodule can be used to output the predicted click-through rates for commodity A and commodity B, and then the loss value currently corresponding to the click-through rate prediction submodule can be determined based on the predicted click-through rate of commodity A and its actual click-through rate. And, the loss value currently corresponding to the click-through rate prediction submodule can be determined based on the predicted click-through rate of commodity B and its actual click-through rate.

[0084] In some application scenarios, the above loss value can be determined, for example, by a cross entropy loss function or a maximum likelihood estimation function.

[0085] Step 5: Determine the overall loss value of the initial prediction model according to the loss values ​​corresponding to the respective click rate prediction submodules, and obtain the prediction model when the overall loss value meets the preset loss requirement.

[0086] In some application scenarios, for example, the loss values ​​corresponding to each click-through rate prediction submodule may be accumulated or multiplied to obtain the overall loss value of the initial prediction model.

[0087] The above-mentioned preset loss requirement may, for example, include being less than a preset loss threshold or a variation range between a plurality of adjacently determined loss values ​​being less than a preset variation range.

[0088] Then, when the overall loss value meets the preset loss requirement, it can be determined that the initial prediction model converges, and thus the prediction model can be obtained.

[0089] In this implementation, a prediction model can be obtained by training the click rate prediction submodules of the initial prediction model to process each type of sample material respectively. Such a prediction model can be used to output the predicted click rates of various types of materials, thereby facilitating the recommendation of different types of materials to the users to be recommended.

[0090] In some optional implementations, the above step 2, for any sample material, obtains the sample material characteristics corresponding to the sample material and the sample user characteristics corresponding to the sample user who clicks the sample material, including: for any sample material, obtains the cross-characteristics, material attribute characteristics and material statistical characteristics corresponding to the sample material; wherein the cross-characteristics characterize the interaction characteristics of the sample material in the click behavior sequence of the sample user who clicks the sample material; and the material statistical characteristics characterize the statistical data characteristics related to the sample material.

[0091] The above cross-features may include, for example, whether the material is in the historical click sequence of the sample user who clicked the sample material, the number of times the material appears in the historical click sequence, the proportion of the material appearing in the historical click sequence, etc.

[0092] In this implementation, it is shown that sample material features can include cross-features, material attribute features, and material statistical features. In this way, not only the basic features of the sample material are obtained, but also the cross-features related to the click behavior of the sample material and the sample user are obtained, so that the personalized preferences of the sample user can be better captured, and the prediction accuracy of the prediction model for different recommended users is improved to a certain extent.

[0093] In some optional implementations, the step 2 described above of obtaining, for any sample material, sample material features corresponding to the sample material and sample user features corresponding to a sample user who clicks the sample material, includes:

[0094] First, for any sample material, obtain the individual behavior characteristics corresponding to the sample user who clicks the sample material, as well as the overall user attribute characteristics and overall user statistical characteristics corresponding to multiple sample users;

[0095] The above-mentioned individual behavior characteristics are characteristics that are unique to the sample user, such as his / her historical click sequence, historical browsing time for various materials, and other characteristics related to the behavior of the sample user.

[0096] The above-mentioned overall user attribute characteristics are attribute characteristics such as gender, age, device model, city, etc., belonging to multiple sample users.

[0097] The above-mentioned overall user statistical characteristics are also statistical data characteristics such as activity, purchasing power, etc. belonging to multiple sample users.

[0098] Then, the weighted submodule in the initial prediction model is used to obtain the sample user characteristics according to the individual behavior characteristics, the overall user attribute characteristics, and the overall user statistical characteristics.

[0099] In some application scenarios, the weighted submodule may, for example, pre-set different weights for individual behavior characteristics, overall user attribute characteristics, and overall user statistical characteristics, so that each characteristic may be accumulated according to the weight corresponding to each characteristic to obtain sample user characteristics.

[0100] It should be noted that if the initial prediction model only considers individual behavior characteristics, but does not consider overall user attribute characteristics and overall user statistical characteristics, the weight of the latter can be set to 0.

[0101] In this implementation, the sample user characteristics obtained take into account both the individual behavioral characteristics corresponding to the sample users and the common characteristics corresponding to the multiple sample users as a whole. In this way, on the one hand, the prediction model can more accurately capture the actual interests and needs of the recommended users by identifying the common characteristics between the individual behavioral characteristics of the users and the common characteristics of the group to which they belong. On the other hand, for some new recommended users or recommended users with less individual behavioral data, by referring to the common characteristics of their group, the data sparsity defects caused by insufficient personal data can be compensated, and more reasonable recommendations can be provided. Furthermore, the use of this common feature can not only strengthen the materials of interest known to the recommended users, but also guide the recommended users to discover new points of interest, predict their potential needs, and reduce irrelevant or repeated recommendations, thereby improving user satisfaction and usage experience.

[0102] To understand the training process of the prediction model, please refer to Figure 2 , which shows a schematic diagram of the structure of an initial prediction model provided by an embodiment of the present application. The initial prediction model includes a commodity click rate prediction submodule for processing commodity type materials and a content click rate prediction submodule for processing non-commodity type materials.

[0103] Taking the commodity click-through rate prediction submodule as an example, after the network platform obtains the sample material features and sample user features corresponding to each sample material, each feature can be converted into a corresponding feature vector through the vector representation layer. Then, the deep neural network used to extract high-dimensional feature vectors can be used to extract the commodity cross feature vector, commodity feature vector, and sample user feature vector of the corresponding dimension allowed by the next deep neural network model. These vectors can then be spliced ​​and finally input into the commodity click-through rate prediction submodule to obtain the predicted click-through rate for each commodity.

[0104] In some application scenarios, after the sample user features pass through the vector representation layer, the individual behavior feature vectors related to the product included therein can be input into the product interest tower (such as a deep neural network model) to extract the high-dimensional feature vector and obtain the individual behavior feature vector of the corresponding dimension. Then, the overall user attribute features and the overall user statistical features can be input into the user commonality tower (such as a deep neural network model) to obtain the commonality feature vector between multiple sample users. Then, the individual behavior feature vector and the commonality feature vector are input into the weighted gating network to obtain the final user interest vector. Then, the user interest vector is passed through a deep neural network model to extract the corresponding high-dimensional feature vector to obtain the user interest vector of the corresponding dimension. The user interest vector of the corresponding dimension is also the current sample user feature vector.

[0105] Furthermore, the weighted gating network can be pre-set with weights for individual behavior feature vectors and weights for common feature vectors, so that when determining the final user interest vector, it can be determined by the corresponding weights, thereby improving the versatility of the initial prediction model. For example, if only individual behavior characteristics are considered, the weights of the overall user attribute feature vector and the overall user statistical feature vector can be set to 0. If the three are considered, different non-zero weights can be set for each.

[0106] In other application scenarios, the initial prediction model may not include the user commonality tower and the weighted gating network. In this case, the individual behavior feature vector can be directly input into the product interest tower to obtain an individual behavior feature vector of the input dimension suitable for the product click rate prediction submodule. The individual behavior feature vector of this dimension is also the current sample user feature vector. In this way, the structure of the initial prediction model can be further simplified.

[0107] It should be noted that after the product click-through rate prediction submodule outputs the corresponding vector, the vector can be processed by a sigmoid activation function to obtain the corresponding predicted click-through rate.

[0108] Then, the loss values ​​of the product click rate prediction submodule and the content click rate prediction submodule can be determined respectively, and the two loss values ​​are added to obtain the overall loss value, and when the overall loss value meets the preset loss requirement, the prediction model can be obtained.

[0109] Those skilled in the art will appreciate that, in the above method of a specific embodiment, the order in which the steps are written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of the steps should be determined by their functions and possible internal logic.

[0110] Please refer to Figure 3 , which shows a structural block diagram of a material recommendation device provided in an embodiment of the present application. The material recommendation device may be a module, program segment or code on an electronic device. It should be understood that the device is similar to the above-mentioned Figure 1 The method embodiment corresponds to and can be executed Figure 1 The method embodiment involves various steps.

[0111] Optionally, the material recommendation device includes an acquisition module 301, a similar material determination module 302, a preference determination module 303, a candidate material set determination module 304, a prediction module 305, and a recommendation module 306. The acquisition module 301 is used to acquire the historical behavior characteristics of the user to be recommended for different types of materials; the similar material determination module 302 is used to determine, for each material in any type of material, multiple similar materials corresponding to the material according to the historical behavior characteristics of the user to be recommended; the preference determination module 303 is used to determine, for any type of material, the preference degree of the user to be recommended for the material of this type; the candidate material set determination module 304 is used to determine the candidate material set from the multiple similar materials corresponding to each type of material according to the preference degree of the user to be recommended for each type of material; the prediction module 305 is used to input the candidate material and the user characteristics corresponding to the user to be recommended into the prediction model for any candidate material in the candidate material set to obtain the predicted click rate for the candidate material; the recommendation module 306 is used to recommend to the user to be recommended according to the predicted click rates corresponding to each candidate material.

[0112] Optionally, the device further comprises a model acquisition module, the module being used for: before inputting the candidate material and the user features corresponding to the to-be-recommended user into the prediction model for any candidate material in the candidate material set to obtain the prediction model; wherein the prediction model is pre-trained by a training module, the training module being used for: obtaining sample materials and sample users; the sample materials comprising positive sample materials and negative sample materials, the positive sample materials comprising materials that have been exposed and clicked by sample users, the negative sample materials comprising materials that have been exposed but not clicked by sample users; for any sample material, obtaining the sample material features corresponding to the sample material in order to obtain the predicted click rate of the candidate material; and sample user characteristics corresponding to the sample users who click on the sample material; and, according to the type of the sample material, using the click-through rate prediction submodule for processing sample materials of this type in the initial prediction model to process the sample material characteristics and sample user characteristics corresponding to the sample material, to obtain the corresponding predicted click-through rate; for any click-through rate prediction submodule, determine the loss value between the predicted click-through rates output by the click-through rate prediction submodule for different sample materials of the type of material it processes and the corresponding actual click-through rates; according to the loss values ​​corresponding to each click-through rate prediction submodule, determine the overall loss value of the initial prediction model, and when the overall loss value meets the preset loss requirement, obtain the prediction model.

[0113] Optionally, the training module is further used to: for any sample material, obtain cross-features, material attribute features and material statistical features corresponding to the sample material; wherein the cross-features represent the interactive features of the sample material in the click behavior sequence of the sample user who clicks the sample material; and the material statistical features represent the statistical data features related to the sample material.

[0114] Optionally, the training module is further used to: for any sample material, obtain the individual behavior characteristics corresponding to the sample users who click on the sample material, as well as the overall user attribute characteristics and overall user statistical characteristics corresponding to multiple sample users; and use the weighted sub-module in the initial prediction model to obtain the sample user characteristics based on the individual behavior characteristics, the overall user attribute characteristics, and the overall user statistical characteristics.

[0115] Optionally, the preference level determination module 303 is further used to: determine, for any type of material, a plurality of historical behavior characteristics of the user to be recommended for that type of material; and determine the preference level of the user to be recommended for that type of material based on the plurality of historical behavior characteristics and the preference weights corresponding to each historical behavior characteristic.

[0116] Optionally, the candidate material set determination module 304 is further used to: for any type of material, determine the number of candidate materials to be selected from multiple similar materials corresponding to each material of this type of material according to a preset number of candidate materials in the candidate material set and the preference degree of the user to be recommended to this type of material; and, according to the number of candidate materials to be selected, screen out a corresponding number of candidate materials from multiple similar materials corresponding to each material of this type of material to obtain the candidate material set.

[0117] It should be noted that those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0118] Please refer to Figure 4 , Figure 4A schematic diagram of the structure of an electronic device for executing a material recommendation method provided in an embodiment of the present application, the electronic device may include: at least one processor 401, such as a CPU, at least one communication interface 402, at least one memory 403 and at least one communication bus 404. Among them, the communication bus 404 is used to realize the direct connection and communication of these components. Among them, the communication interface 402 of the device in the embodiment of the present application is used to communicate signaling or data with other node devices. The memory 403 can be a high-speed RAM memory or a non-volatile memory (non-volatile memory), such as at least one disk storage. The memory 403 can also be at least one storage device located away from the aforementioned processor. Computer-readable instructions are stored in the memory 403. When the computer-readable instructions are executed by the processor 401, the electronic device can execute the methods provided by the above-mentioned method embodiments.

[0119] Understandably, Figure 4 The structure shown is for illustration only, and the electronic device may also include Figure 4 More or fewer components as shown, or with Figure 4 Different configurations are shown. Figure 4 Each component shown in the figure can be implemented by hardware, software or a combination thereof.

[0120] An embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the methods provided by the above-mentioned method embodiments can be executed.

[0121] An embodiment of the present application provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the methods provided by the above-mentioned method embodiments.

[0122] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0123] In addition, the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0124] Furthermore, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0125] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0126] The above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A material recommendation method, characterized in that: include: Obtain the historical behavior characteristics of the user to be recommended for different types of materials; For each material in any type of material, multiple similar materials corresponding to the material are determined according to the historical behavior characteristics of the user to be recommended; For any type of material, determining the preference of the to-be-recommended user for the type of material; According to the preference degree of the to-be-recommended user for each type of material, a candidate material set is determined from a plurality of similar materials corresponding to each type of material; For any candidate material in the candidate material set, the candidate material and the user characteristics corresponding to the to-be-recommended user are input into a prediction model to obtain a predicted click rate for the candidate material; Recommendations are made to the user to be recommended based on the predicted click-through rates corresponding to the candidate materials.

2. The method according to claim 1, characterized in that Before inputting the candidate material and the user characteristics corresponding to the to-be-recommended user into a prediction model for any candidate material in the candidate material set to obtain a predicted click rate for the candidate material, the method further includes: Obtaining the prediction model; The prediction model is pre-trained according to the following steps: Obtain sample materials and sample users; the sample materials include positive sample materials and negative sample materials, the positive sample materials include materials that have been exposed and clicked by sample users, and the negative sample materials include materials that have been exposed but not clicked by sample users; For any sample material, obtain the sample material features corresponding to the sample material and the sample user features corresponding to the sample user who clicks the sample material; and According to the type of the sample material, the click rate prediction submodule for processing the sample material of the type in the initial prediction model is used to process the sample material features and sample user features corresponding to the sample material to obtain the corresponding predicted click rate; For any click rate prediction submodule, determine the loss value between the predicted click rates output by the click rate prediction submodule for different sample materials in the type of material processed by the click rate prediction submodule and the corresponding actual click rates; The overall loss value of the initial prediction model is determined according to the loss values ​​corresponding to the respective click rate prediction submodules, and the prediction model is obtained when the overall loss value meets the preset loss requirement.

3. The method according to claim 2, characterized in that The step of obtaining, for any sample material, sample material features corresponding to the sample material and sample user features corresponding to a sample user who clicks the sample material includes: For any sample material, obtain the cross-features, material attribute features and material statistical features corresponding to the sample material; The cross feature represents the interactive feature of the sample material in the click behavior sequence of the sample user who clicks the sample material; The material statistical characteristics represent statistical data characteristics related to the sample material.

4. The method according to claim 2 or 3, characterized in that: The step of obtaining, for any sample material, sample material features corresponding to the sample material and sample user features corresponding to a sample user who clicks the sample material includes: For any sample material, obtain the individual behavior characteristics corresponding to the sample user who clicks the sample material, as well as the overall user attribute characteristics and overall user statistical characteristics corresponding to multiple sample users; The sample user characteristics are obtained by utilizing the weighted submodule in the initial prediction model according to the individual behavior characteristics, the overall user attribute characteristics, and the overall user statistical characteristics.

5. The method according to claim 1, characterized in that The step of determining the preference of the user to be recommended for any type of material includes: For any type of material, determining multiple historical behavior characteristics of the to-be-recommended user for the type of material; and The preference degree of the to-be-recommended user for the type of material is determined according to the multiple historical behavior characteristics and the preference weights respectively corresponding to the historical behavior characteristics.

6. The method according to claim 1 or 5, characterized in that: The step of determining a candidate material set from a plurality of similar materials corresponding to each type of material according to the preference degree of the to-be-recommended user for each type of material includes: For any type of material, according to the preset number of candidate materials in the candidate material set and the preference of the user to be recommended for the type of material, determine the number of candidate materials to be selected from multiple similar materials corresponding to each material in the type of material; and According to the number of the candidate materials to be selected, a corresponding number of multiple candidate materials are screened out from multiple similar materials corresponding to each material of the type of material, to obtain the candidate material set.

7. A material recommendation device, characterized in that: include: An acquisition module is used to obtain the historical behavior characteristics of the user to be recommended for different types of materials; A similar material determination module is used to determine, for each material of any type of material, a plurality of similar materials corresponding to the material according to the historical behavior characteristics of the user to be recommended; A preference level determination module is used to determine the preference level of the user to be recommended to any type of material for the material; A candidate material set determination module is used to determine a candidate material set from a plurality of similar materials corresponding to each type of material according to the preference of the user to be recommended to each type of material; A prediction module, for inputting the candidate material and the user characteristics corresponding to the to-be-recommended user into a prediction model for any candidate material in the candidate material set, to obtain a predicted click rate for the candidate material; The recommendation module is used to recommend the candidate materials to the user to be recommended according to the predicted click-through rate corresponding to each candidate material.

8. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 6 is executed.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is executed.

10. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the method according to any one of claims 1 to 6 is performed.

Images