Recommendation model training method, recommendation method and related device

Abstract

The invention provides a recommendation model training method, a recommendation method and a related device. The overall recommendation accuracy and surprise performance can be improved. The method comprises the following steps: determining surprise items according to multi-dimensional surprise factors, user interaction historical data and candidate items, wherein the multi-dimensional surprise factors comprise unexpectability, novelty, timeliness and correlation; determining multi-dimensional surprise condition information according to the surprise items; inputting the noise, the multi-dimensional surprise condition information and user interaction historical data into a conditional generative adversarial network of the initial recommendation model to generate a cold start embedded vector; predicting the score of the cold start item according to the cold start embedding vector, the original item embedding vector and the original user-item score to obtain a recommendation list; loss corresponding to the recommendation list is determined according to the recommendation list and a preset loss function, the preset loss function comprises a preset surprise factor function, and the loss comprises surprise loss; and training the initial recommendation model by optimizing the loss to obtain a target recommendation model.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of recommendation, and particularly relates to a training method of a recommendation model, a recommendation method and related devices.

BACKGROUND

[0002] In today's information explosion era, recommendation systems have become an important tool to help users filter personalized needs from massive information, and often face the problems of cold start and filter bubble. The cold start problem is caused by new users or new items being difficult to be accurately recommended due to sparse interaction data, and the system often relies on popular content to fill or map, resulting in homogenization of the recommendation results. The filter bubble problem is caused by the algorithm excessively relying on user historical behavior, falling into a cycle of repeatedly recommending similar content, limiting the possibility of users exploring new interests. Most existing methods consider the solution of these two problems in isolation, while ignoring the negative effects of mutual reinforcement, i.e., cold start recommendation exacerbates filter bubble, and filter bubble further hinders users from discovering potential interests, thus forming a vicious cycle, ultimately leading to reduced overall recommendation accuracy and surprise performance.

SUMMARY

[0003] The present application provides an item recommendation method, device and storage medium, which improves the accuracy of cold start recommendation.

[0004] The first aspect of the embodiment of the present application provides a training method of a recommendation model, comprising:

[0005] determining a surprise item according to a multi-dimensional surprise factor, user interaction history data and a candidate item, wherein the multi-dimensional surprise factor comprises unexpectedness, novelty, timeliness and relevance;

[0006] determining multi-dimensional surprise condition information according to the surprise item;

[0007] inputting noise, the multi-dimensional surprise condition information and the user interaction history data into a conditional generative adversarial network of an initial recommendation model to generate a cold start embedding vector;

[0008] predicting a score of a cold start item according to the cold start embedding vector, an original item embedding vector and an original user-item score to obtain a recommendation list;

[0009] determining a loss corresponding to the recommendation list according to the recommendation list and a preset loss function, wherein the preset loss function comprises a preset surprise factor function, and the loss comprises a surprise degree loss;

[0010] training the initial recommendation model by optimizing the loss to obtain a target recommendation model.

[0011] The second aspect of the embodiment of the present application provides a training device of a recommendation model, comprising:

[0012] The first determining unit is configured to determine a surprise item according to a multi-dimensional surprise factor, user interaction history data and a candidate item, wherein the multi-dimensional surprise factor comprises unexpectedness, novelty, timeliness and relevance.

[0013] The converting unit is configured to determine multi-dimensional surprise condition information according to the surprise item.

[0014] The generating unit is configured to input noise, the multi-dimensional surprise condition information and the user interaction history data into a conditional generative adversarial network of an initial recommendation model to generate a cold start embedding vector.

[0015] The predicting unit is configured to predict a score of a cold start item according to the cold start embedding vector, an original item embedding vector and an original user-item score to obtain a recommendation list.

[0016] The second determining unit is configured to determine a loss corresponding to the recommendation list according to the recommendation list and a preset loss function, wherein the preset loss function comprises a preset surprise factor function, and the loss comprises a surprise degree loss.

[0017] The training unit is configured to train the initial recommendation model by optimizing the loss to obtain a target recommendation model.

[0018] The third aspect of the embodiment of the present application provides a recommendation method, comprising:

[0019] Obtaining user interaction history data;

[0020] Inputting the user interaction history data into a target recommendation model to output a recommendation list, wherein the target recommendation model is a target model obtained by using the method of the first aspect of the embodiment.

[0021] The fourth aspect of the embodiment of the present application provides a computer device, comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to realize the steps of the method of the first aspect.

[0022] The fifth aspect of the embodiment of the present application provides a computer program product, comprising instructions, wherein the instructions are executed by a processor to realize the steps of the method of the first aspect.

[0023] The sixth aspect of the embodiment of the present application provides a computer storage medium, comprising instructions, wherein when the instructions are executed on a computer, the computer executes the steps of the method of the first aspect.

[0024] Compared with the related art, in the embodiment provided by the application, the multi-dimensional surprise factor can solve the filtering bubble problem, avoid filtering the cold start item, predict the score of the cold start item by fully utilizing the information of the original item and the original user-item score, and improve the recommendation accuracy of the cold start item, so as to solve the cold start and filtering bubble problems at the same time and improve the overall recommendation accuracy and surprise performance. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description only constitute some embodiments of the application, and for those skilled in the art, other drawings can also be obtained from the structures shown in the drawings without creative labor.

[0026] Figure 1 The flowchart of the training method of the recommendation model provided by the embodiment of the application;

[0027] Figure 2 The application schematic diagram of the training method of the recommendation model provided by the embodiment of the application;

[0028] Figure 3 The virtual structure schematic diagram of the training device of the recommendation model provided by the embodiment of the application;

[0029] Figure 4 The hardware structure schematic diagram of the server provided by the embodiment of the application.

DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments only constitute some embodiments of the application, not all embodiments.

[0031] The training method of the recommendation model will be described below from the perspective of the training device of the recommendation model. Please refer to Figure 1 The flowchart of the training method of the recommendation model provided by the embodiment of the application, the recommendation method of the recommendation model includes:

[0032] 101、According to the multi-dimensional surprise factor, the user interaction history data and the candidate item, the surprise item is determined, and the multi-dimensional surprise factor specifically includes unexpectedness, novelty, timeliness and relevance.

[0033] In this embodiment, based on the unexpectedness, novelty, timeliness and relevance dimensions, the training device of the recommendation model can determine whether the candidate item is surprising to the user through user interaction history data, which includes user information, original items interacted with the user, original user-item ratings, etc., wherein the original user-item rating represents the user's rating of the original item.

[0034] The unexpectedness dimension is used to mine high-value items across categories. Users' interests often do not only limit to a few specific categories, and they may have potential interest in some cross-category items in the same scenario. The introduction of the unexpectedness dimension enables the recommendation system to break through the conventional recommendation range and recommend some less-interacted but high-value items to the user, thereby increasing the surprise degree of the recommendation.

[0035] The novelty dimension is used to measure the selection frequency of the user for similar items. Users often want to discover some new and different items to meet their changing needs and interests. The introduction of the novelty dimension enables the recommendation system to mine small or newly emerging items that the user may be interested in, thereby satisfying the user's curiosity.

[0036] The timeliness dimension reflects the user's recent interaction behavior. The actual user's needs and interests are constantly changing, and the user does not always desire "surprise" at all times. The introduction of the timeliness dimension enables the recommendation system to adjust according to the user's real-time needs.

[0037] The relevance dimension is used to measure the historical interaction relationship between the user and the item. By analyzing the user's past interaction with the item, such as purchase, browsing, collection, etc., the user's interest in different items can be determined. The introduction of the relevance dimension enables the recommendation system to focus on items that the user has explicitly shown interest in, thereby improving the accuracy of the recommendation.

[0038] In some embodiments, the surprise item includes an unexpected item, a novelty item, a timeliness item and a relevance item, and the surprise item is determined according to the multi-dimensional surprise factor, the user interaction history data and the candidate item, and the specific marking steps can include the following:

[0039] If the number of interactions of the user with the category to which the candidate item belongs is less than or equal to a first preset threshold, and the rating of the candidate item is greater than or equal to a preset rating threshold, the candidate item is marked as an unexpected item;

[0040] If the number of interactions of the user with the candidate item is less than or equal to a second preset threshold, and the category label of the candidate item belongs to the category set corresponding to the original item, the candidate item is marked as a novelty item;

[0041] If the interaction of the candidate item with the user is within a recent threshold number of times, the candidate item is marked as a timeliness item;

[0042] If the number of interactions between the candidate item and the user is greater than or equal to the third preset threshold, the candidate item is marked as a relevance item.

[0043] It can be understood that by marking the candidate item, the multi-dimensional surprise information is constructed according to the unexpectedness item, the novelty item, the timeliness item and the relevance item, and invalid candidate items can be filtered out.

[0044] It should be noted that the training device of the recommendation model can use a matrix as a data carrier, mark the multi-dimensional surprise factor matrix by using a gating network, filter out the candidate items that do not simultaneously satisfy the four surprise factors in the multi-dimensional surprise factor matrix, determine the surprise item according to the multi-dimensional surprise factor, the user interaction history data and the candidate item, and the specific steps further include:

[0045] constructing a user-candidate item interaction matrix and a score matrix according to the user interaction history data and the candidate item;

[0046] constructing a multi-dimensional surprise factor matrix according to the multi-dimensional surprise factor, the user-candidate item interaction matrix and the score matrix, the multi-dimensional surprise factor matrix including an unexpectedness matrix, a novelty matrix, a timeliness matrix and a relevance matrix.

[0047] It can be understood that the sub-item of the unexpectedness matrix is the unexpectedness item, the sub-item of the novelty matrix is the novelty item, the sub-item of the timeliness matrix is the timeliness item, and the sub-item of the relevance matrix is the relevance item, and the specific execution process of the step of constructing the multi-dimensional surprise factor matrix can adopt the marking step.

[0048] Exemplarily, please refer to Figure 2 , an application schematic diagram of the training method of the recommendation model provided by the embodiment of the present application. A user-candidate item interaction matrix R u,i and a score matrix X u,i are taken as an example. The user u is taken as a row, and the candidate item i is taken as a column. The sub-item r u,i of the interaction matrix R u,i marks the interaction behavior. 1 represents that there is interaction, and 0 represents that there is no interaction. The sub-item x u,i of the score matrix X u,i marks the score. The value of x u,i represents the score value. The item category C is obtained from the original item feature set M i . The label data is directly extracted by the category extraction function C u,i = f(M i ).

[0049] The unexpectedness of the candidate item i can be expressed by the following formula as the sub-item of the unexpectedness matrix:

[0050]

[0051] Among them, δ sur represents the preset scoring threshold, β sur Indicates a low interaction threshold. C u Indicates the category of the original item, C u，i Indicates the category of candidate items, selected by k-fold cross validation, sum(C u,i ) represents the total number of interactions of the category to which the candidate item belongs.

[0052] It can be understood that, in this embodiment, the total number of interactions sum (C u,i ), to determine whether it is less than the set low interaction threshold β sur , thereby determining whether candidate item i belongs to the low interaction category item.

[0053] It can also be understood that in the user-candidate item interaction matrix R u,i In the example, 1 means there is interaction and 0 means there is no interaction. Then the total number of interactions of the category to which candidate item i belongs (C u,i ), and is the total number of candidate items i that interact with user u among all candidate items i in the category to which candidate item i belongs.

[0054] Therefore, by analyzing the user interaction frequency and ratings of different items, as well as the categories to which the candidate items belong, we can determine whether candidate item i meets the low interaction and high rating conditions. Sur(u,i) = 1 means that the sub-item corresponding to candidate item i is marked as 1, indicating that candidate item i is an unexpected item; otherwise, it is marked as 0.

[0055] Specifically, the novelty of candidate item i can be expressed as a sub-item of the novelty matrix using the following formula:

[0056]

[0057] It is understandable that by counting the frequency of user u's selections of items in different categories, it is possible to determine whether the category to which candidate item i belongs is a low-frequency selection for user u. New(u,i)=1 indicates that the sub-item corresponding to candidate item i is marked as 1, indicating that candidate item i is a novel item; otherwise, it is marked as 0.

[0058] Specifically, the timeliness of candidate project i can be expressed as a sub-item of the timeliness matrix using the following formula:

[0059]

[0060] Among them, S u is the set of the most recent δ interactions. It can be understood that by judging the interaction r between user u and candidate item i (u,i)Whether the recent interaction count is within the threshold can determine whether the interaction is timely and can bring the desired "surprise" to user u at that time. Tim(u,i) = 1 means that the sub-item corresponding to candidate item i is marked as 1, indicating that candidate item i is timely; otherwise, it is marked as 0.

[0061] It is also understandable that in reality, different users have different interaction frequencies, which makes it difficult to determine the interaction count threshold in a general sense. This is because high-frequency users have a large number of interactions and large fluctuations, while medium and low-frequency users have small fluctuations. Therefore, we need to first divide high-frequency, medium-frequency, and low-frequency interaction users evenly according to the number of interactions, and determine the interaction timeliness threshold for different frequency user groups. Thus, the collection of the most recent δ interactions is constructed according to the timeliness thresholds of different user groups.

[0062] Specifically, the relevance of candidate item i can be expressed as a sub-item of the relevance matrix using the following formula:

[0063]

[0064] It is understandable that by analyzing whether user u has interacted with candidate item i, it is possible to determine whether the user is interested in candidate item i. Rel(u,i)=1 means that the sub-item corresponding to candidate item i is marked as 1, indicating that candidate item i is a relevant item, otherwise it is marked as 0.

[0065] It should be noted that for the candidate item i corresponding to the sub-item marked as 0 above, the corresponding candidate item i can be filtered out by marking it as 0. Furthermore, in real-world cold start scenarios, the matrix R is very sparse, which means that there are many "?" in the matrix. If the original sparse interaction matrix is ​​input into the model, it will greatly reduce the recommendation performance. Therefore, marking all the values ​​in the 0-1 matrix of Sur(u,i), New(u,i), Tim(u,i), and Rel(u,i) as 0 that are also "?" or 0 as 0 can filter out invalid candidate items and enhance the multidimensional surprise factor matrix.

[0066] 102. Determine multi-dimensional surprise condition information based on the surprise item.

[0067] In this embodiment, the training device of the recommendation model can determine multi-dimensional surprise condition information based on the surprise item, and input the conditional generative adversarial network of the initial recommendation model to enrich the generated information. Determining the multi-dimensional surprise condition information based on the surprise item specifically includes:

[0068] Convert surprise items into unexpectedness vector, novelty vector, timeliness vector and relevance vector;

[0069] The unexpectedness vector, the novelty vector, the timeliness vector and the relevance vector are combined and spliced into the surprise condition information.

[0070] The training device of the recommendation model can convert the unexpectedness item, the novelty item, the timeliness item and the relevance item into the unexpectedness vector, the novelty vector, the timeliness vector and the relevance vector through the embedding layer, and can obtain the unexpectedness vector, the novelty vector, the timeliness vector and the relevance vector through the following formula:

[0071]

[0072] In the formula, Embedding-Layer represents the embedding layer.

[0073] The surprise condition information can be obtained through the following formula:

[0074]

[0075] In the formula, Concat is a combination and splicing function.

[0076] 103. The noise, the multi-dimensional surprise condition information and the user interaction history data are input into the conditional generative adversarial network of the initial recommendation model to generate the cold start embedding vector.

[0077] In the embodiment, the training device of the recommendation model can input the noise, the multi-dimensional surprise condition information and the user history interaction data into the conditional generative adversarial network of the initial recommendation model to generate the cold start embedding vector. The noise can be noise of a standard distribution, for example, noise of a Gaussian distribution. The conditional generative adversarial network can be a conditional generative adversarial network for attribute manipulation (CGAR), which guides the generation process of the generator by the multi-dimensional surprise condition information, and the discriminator discriminates the data authenticity by the multi-dimensional surprise condition information. Through the confrontation between the generator and the discriminator, the initial recommendation model generates the cold start embedding vector that meets the multi-dimensional surprise condition information. The noise, the multi-dimensional surprise condition information and the user history interaction data are input into the conditional generative adversarial network of the initial recommendation model to generate the cold start embedding vector, which can specifically include:

[0078] The multi-dimensional surprise condition information is used to guide the sampling noise to generate virtual cold start data;

[0079] The user history interaction data is sampled to generate real cold start data;

[0080] The virtual cold start data and the real cold start data are discriminated whether they meet the surprise condition information to obtain a discrimination result;

[0081] The cold start embedding vector is obtained through the adversarial training of the discrimination result and the preset target function.

[0082] The preset target function is as follows:

[0083]

[0084] Wherein, G represents the generator, R represents the discriminator, Y represents a set of multi-dimensional surprise information samples, y represents real cold start data, Z represents a set of sampling noises, and z represents a sampling noise.

[0085] It can be understood that, V(G, R) represents the expected logarithmic output of the discriminator R on the real cold start data, and the goal of the discriminator R is to maximize this V(G, R), V(G, R) represents the expected logarithmic output of the discriminator R on the virtual cold start data, and the goal of the generator G is to minimize V(G, R). Through the two expected items, the preset target function ensures that the generator and the discriminator can be mutually opposed and promoted in the training process, thereby continuously improving the quality of the generated data and the discrimination ability of the discriminator.

[0086] In some embodiments, the sampling noise can be randomly masked with a preset probability: The robustness and generalization ability of the model training are improved. The specific mask type, proportion, and position can be set according to actual needs, and of course a dynamic adjustment strategy can also be used, which is not limited here.

[0087] 104. The ratings of the cold start items are predicted according to the cold start embedding vector, the original item embedding vector, and the original user-item ratings to obtain a recommendation list.

[0088] In this embodiment, after the training device of the recommendation model generates the cold start embedding vector that meets the multi-dimensional surprise degree condition information through the conditional adversarial network, the pre-trained original item embedding vector is input, and a recommendation list is obtained through the collaborative filtering recommendation method. The ratings of the cold start items are predicted according to the cold start embedding vector, the original item embedding vector, and the original user-item ratings to obtain a recommendation list, which can specifically include:

[0089] The similarity between the cold start embedding vector and the original item embedding vector is calculated;

[0090] The similarity is multiplied by the original user-item ratings to predict the ratings of the cold start items;

[0091] The ratings of the cold start items are sorted to obtain a recommendation list.

[0092] It can be understood that the higher the similarity between the cold start embedding vector and the original item embedding vector, the stronger the correlation between the two items. The score of the cold start item can be predicted based on the similarity between the cold start item and the original item, and the score of the cold start item can be predicted based on the score of the original item. Therefore, for the cold start item lacking user interaction data, the information of the original item can be fully utilized to predict the score of the cold start item, and the new cold start item of interest to the user can be accurately recommended.

[0093] In some embodiments, calculating the similarity between the cold start embedding vector and the original item embedding vector specifically comprises:

[0094]

[0095] wherein G(cold i ) represents the cold start embedding vector, e j represents the original item embedding vector.

[0096] Multiplying the similarity and the original user-item score to predict the score of the cold start item specifically comprises:

[0097]

[0098] wherein M embed represents the embedding vector of the item feature set, x embed represents the original user-item score embedding vector.

[0099] Ranking the scores of the cold start items to obtain a recommendation list specifically comprises:

[0100] Ranking the scores of the cold start items;

[0101] Obtaining the top K cold start items to generate a recommendation list.

[0102] It can be understood that K is a positive integer, and the specific value is set by a person skilled in the art according to the actual situation, which is not limited herein.

[0103] 105. Determining the loss corresponding to the recommendation list according to the recommendation list and a preset loss function.

[0104] In this embodiment, the preset loss function includes a preset surprise factor function and a preset accuracy function, and the loss includes a surprise degree loss and an accuracy loss. After the training device of the recommendation model generates the recommendation list, the performance of the recommendation list generated by the recommendation model in terms of surprise degree can be evaluated through the preset loss function. According to the recommendation list and the preset loss function, the loss corresponding to the recommendation list is determined, specifically comprising:

[0105] determine the unexpectedness loss, the novelty loss, the timeliness loss and the relevance loss of the recommended item in the recommendation list according to the multi-dimensional surprise factor and the user interaction history data;

[0106] perform weighted summation on the unexpectedness loss, the novelty loss, the timeliness loss and the relevance loss to obtain the surprise degree loss.

[0107] The unexpectedness considers cross-category low-interaction items, and the unexpectedness loss can be calculated by the following unexpectedness loss function:

[0108]

[0109] wherein K is the number of recommended items in the recommendation list List, and the value of Sur(u, i) indicates that if the recommended item is a low-interaction and high-score item of the user, it is marked as 1, otherwise it is marked as 0, which is used to calculate the unexpectedness item hit ratio.

[0110] The novelty considers low-interaction items in the same category, and the main goal is to recommend novel items in the same category, i.e. items with low historical interaction times in the recommendation list. The novelty loss can be calculated by the following novelty loss function:

[0111]

[0112] wherein the value of Sur(u, i) indicates that if the recommended item is an item that the original user has not interacted with and is in the same category, it is marked as 1, otherwise it is marked as 0, which is used to calculate the novelty item hit ratio.

[0113] The timeliness considers the user's last δ times of interaction set S u The timeliness loss can be calculated by the following timeliness loss function:

[0114]

[0115] wherein the value of Tim(u, i) indicates that if the recommended item is in S u , it is marked as 1, otherwise it is marked as 0, which is used to calculate the timeliness item hit ratio.

[0116] The relevance considers the prediction accuracy of the user in the interaction items, and the relevance loss can be calculated by the following relevance loss function:

[0117]

[0118] wherein the value of Rel(u, i) indicates that if the recommended item is an item that the user has interacted with, it is marked as 1, otherwise it is marked as 0, which is used to calculate the relevance item hit ratio.

[0119] The surprise degree loss is obtained by weighted summation of unexpectedness loss, novelty loss, timeliness loss and relevance loss, and can be calculated by a preset surprise factor loss function as follows:

[0120] L Multi = w1*L Sur + w2*L New + w3*L Tim + w4*L Rel

[0121] wherein w1, w2, w3 and w4 represent loss weights, used for adjusting the weights of different sub-loss terms.

[0122] The training device of the recommendation model evaluates the recommendation accuracy of the model based on a preset accuracy function, determines the loss corresponding to the recommendation list according to the recommendation list and a preset loss function, and specifically further comprises the following steps of:

[0123] The accuracy loss is calculated by a preset loss function as follows:

[0124] L Acc = V(G, R) + σ*Ω(G, R)

[0125] wherein Ω(G, R) is a regularization loss, and σ is a weight.

[0126] The loss is calculated by the following formula:

[0127] L All = L Acc + L Multi

[0128] 106、The initial recommendation model is trained by optimizing the loss to obtain a target recommendation model.

[0129] In the embodiment, the training device of the recommendation model updates the parameters of the initial model by optimizing the loss under the guidance of the preset surprise factor loss function and the preset accuracy loss function, and obtains a target recommendation model trained. Under the joint optimization of the multi-surprise factor loss, the user will obtain a recommended result with high accuracy and appropriate surprise degree, thereby improving the user experience.

[0130] Compared with the related art, in the embodiments provided by the present application, the multi-dimensional surprise factor can solve the filtering bubble problem and avoid filtering cold start items, the score of the cold start item can be predicted by fully utilizing the information of the original item and the original user-item score, and the recommendation accuracy of the cold start item can be improved, so that the cold start and filtering bubble problems can be solved at the same time, and the overall recommendation accuracy and surprise performance are improved. At the same time, based on the preset accuracy loss function, the recommendation accuracy is further improved.

[0131] The application is described above from the training method of the recommendation model, and is described below from the training device of the recommendation model.

[0132] Please refer to Figure 3 A virtual structure diagram of the training device of the recommendation model is provided for the embodiment of the application, and the training device 200 of the recommendation model comprises:

[0133] The first determination unit 201 is configured to determine a surprise item according to the multi-dimensional surprise factor, the user interaction history data and the candidate item, and the multi-dimensional surprise factor comprises unexpectedness, novelty, timeliness and relevance.

[0134] The conversion unit 202 is configured to determine multi-dimensional surprise condition information according to the surprise item.

[0135] The generation unit 203 is configured to input the noise, the multi-dimensional surprise condition information and the user interaction history data into the conditional generative adversarial network of the initial recommendation model to generate a cold start embedding vector.

[0136] The prediction unit 204 is configured to predict a score of a cold start item according to the cold start embedding vector, an original item embedding vector and an original user-item score to obtain a recommendation list.

[0137] The second determination unit 205 is configured to determine a loss corresponding to the recommendation list according to the recommendation list and a preset loss function, the preset loss function comprises a preset surprise factor function, and the loss comprises a surprise degree loss.

[0138] The training unit 206 is configured to train the initial recommendation model by optimizing the loss to obtain a target recommendation model.

[0139] In a possible design, the surprise item comprises an unexpectedness item, a novelty item, a timeliness item and a relevance item, and the first determination unit 201 is specifically configured to:

[0140] If the number of interactions between the user and the candidate item is less than or equal to a first preset threshold value and the score of the candidate item is greater than or equal to a preset score threshold value, the candidate item is marked as an unexpectedness item.

[0141] If the number of interactions between the user and the candidate item is less than or equal to a second preset threshold value and the category label of the candidate item belongs to a category set corresponding to the original item, the candidate item is marked as a novelty item.

[0142] If the interaction between the candidate item and the user is within a recent threshold number of times, the candidate item is marked as a timeliness item.

[0143] If the number of interactions between the candidate item and the user is greater than or equal to a third preset threshold value, the candidate item is marked as a relevance item.

[0144] In a possible design, the first determining unit 201 is specifically further configured to:

[0145] construct a user-candidate item interaction matrix and a score matrix according to the user interaction history data and the candidate items;

[0146] construct a multi-dimensional surprise factor matrix according to the multi-dimensional surprise factor, the user-candidate item interaction matrix and the score matrix, the multi-dimensional surprise factor matrix including an unexpectedness matrix, a novelty matrix, a timeliness matrix and a relevance matrix.

[0147] In a possible design, the converting unit 202 is specifically configured to:

[0148] convert the surprise item into an unexpectedness vector, a novelty vector, a timeliness vector and a relevance vector;

[0149] combine and splice the unexpectedness vector, the novelty vector, the timeliness vector and the relevance vector into surprise condition information.

[0150] In a possible design, the generating unit 203 is specifically configured to:

[0151] generate virtual cold start data by using the multi-dimensional surprise condition information to guide sampling noise;

[0152] generate real cold start data by sampling user historical interaction data;

[0153] determine whether the virtual cold start data and the real cold start data meet the surprise condition information, to obtain a determination result;

[0154] perform adversarial training by using the determination result and a preset target function, to obtain a cold start embedding vector.

[0155] The preset target function is as follows:

[0156]

[0157] wherein G represents a generator, R represents a discriminator, Y represents a set of multi-dimensional surprise information sampling, y represents real cold start data, Z represents a set of sampling noise, and z represents sampling noise.

[0158] In a possible design, the generating unit 203 can randomly mask the sampling noise with a preset probability:

[0159] In a possible design, the predicting unit 204 is specifically configured to:

[0160] calculate a similarity between the cold start embedding vector and an original item embedding vector;

[0161] multiply the similarity by an original item score, to predict a score of a cold start item.

[0162] The scores of the cold-start items are sorted to obtain a recommendation list.

[0163] In a possible design, the prediction unit 204 calculates the similarity between the cold-start embedding vector and the original item embedding vector by the following formula:

[0164]

[0165] wherein G(cold i ) represents the cold-start embedding vector, e j represents the original item embedding vector.

[0166] In a possible design, the prediction unit 204 predicts the score of the cold-start item by the following formula:

[0167]

[0168] wherein M embed represents the embedding vector of the item feature set, x embed represents the original user-item score embedding vector.

[0169] In a possible design, the prediction unit 204 is further specifically configured to:

[0170] sort the scores of the cold-start items;

[0171] obtain the top K cold-start items to generate the recommendation list.

[0172] In a possible design, the preset loss function includes a preset surprise factor function and a preset accuracy function, the loss includes a surprise degree loss and an accuracy loss, and the second determination unit 205 is specifically configured to:

[0173] determine, according to the multi-dimensional surprise factor and the user interaction history data, the unexpectedness loss, the novelty loss, the timeliness loss, and the relevance loss corresponding to the recommended items in the recommendation list;

[0174] weight-sum the unexpectedness loss, the novelty loss, the timeliness loss, and the relevance loss to obtain the surprise degree loss.

[0175] The second determination unit 205 is specifically further configured to:

[0176] calculate the accuracy loss based on the following formula:

[0177] L Acc = V(G, R) + σ * Ω(G, R)

[0178] wherein Ω(G, R) is a regularization loss, and σ is a weight.

[0179] The loss is calculated based on the following formula:

[0180] L All =L Acc +L Multi

[0181] The embodiment of the present invention further provides a recommendation method, including:

[0182] Get user interaction history data;

[0183] The user interaction history data is input into a target recommendation model to output a recommendation list, wherein the target recommendation model is a target model obtained by using the training method provided by the above embodiment of the present invention.

[0184] Figure 4 This is a schematic diagram of the structure of the server of the present invention, as shown in FIG. Figure 4 As shown, the server 300 of this embodiment includes at least one processor 301, at least one network interface 304 or other user interface 303, a memory 305, and at least one communication bus 302. The server 300 optionally includes a user interface 303, including a display, a keyboard, or a clicking device. The memory 305 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk storage. The memory 305 stores execution instructions. When the server 300 is running, the processor 301 communicates with the memory 305, and the processor 301 calls the instructions stored in the memory 305 to execute the above-mentioned recommendation model training method. The operating system 306 includes various programs for implementing various basic services and processing tasks based on hardware.

[0185] The server provided in the embodiment of the present application can execute the technical solution of the embodiment of the training method of the recommendation model mentioned above. Its implementation principle and technical effects are similar and will not be repeated here.

[0186] Embodiments of the present invention further provide a computer-readable storage medium storing a computer program that, when executed by a computer, implements the method flow associated with the recommendation model training apparatus described in any of the aforementioned method embodiments. Accordingly, the computer may be the recommendation model training apparatus described above.

[0187] Embodiments of the present invention further provide a computer program or a computer program product including a computer program. When executed on a computer, the computer program causes the computer to implement the method flow associated with the recommendation model training device described in any of the aforementioned method embodiments. Accordingly, the computer may be the recommendation model training device described above.

[0188] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware. The foregoing program can be stored in a computer readable storage medium. The program executes to perform the steps of the above-mentioned method embodiments; and the foregoing storage medium includes various storage media that can store program codes, such as ROM, RAM, magnetic disk or optical disk.

[0189] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A training method for a recommendation model, characterized in that: include: Determine a surprise project based on multidimensional surprise factors, user interaction history data, and candidate projects, wherein the multidimensional surprise factors include unexpectedness, novelty, timeliness, and relevance; determining multi-dimensional surprise condition information according to the surprise item; Inputting the noise, the multi-dimensional surprise condition information, and the user interaction history data into a conditional generative adversarial network of an initial recommendation model to generate a cold start embedding vector; Predicting the scores of the cold-start items based on the cold-start embedding vector, the original item embedding vector, and the original user-item score to obtain a recommendation list; Determining a loss corresponding to the recommendation list according to the recommendation list and a preset loss function, wherein the preset loss function includes a preset surprise factor function, and the loss includes a surprise degree loss; The initial recommendation model is trained by optimizing the loss to obtain a target recommendation model.

2. The method according to claim 1, characterized in that The surprise items include unexpected items, novel items, timely items, and relevant items. The process of determining the surprise items based on the multi-dimensional surprise factors, user interaction history data, and candidate items includes: If the number of interactions between the user and the category to which the candidate item belongs is less than or equal to a first preset threshold, and the score of the candidate item is greater than or equal to a preset score threshold, then mark the candidate item as the unexpected item; If the number of interactions between the user and the candidate item is less than or equal to a second preset threshold, and the category label of the candidate item belongs to the category set corresponding to the original item, then mark the candidate item as the novelty item; If the candidate item has interacted with the user within a recent threshold number of times, marking the candidate item as the timely item; If the number of interactions between the candidate item and the user is greater than or equal to a third preset threshold, the candidate item is marked as the relevant item.

3. The method according to claim 1, characterized in that The determining of multi-dimensional surprise condition information according to the surprise item includes: Converting the surprise items into unexpectedness vectors, novelty vectors, timeliness vectors, and relevance vectors; The unexpectedness vector, the novelty vector, the timeliness vector, and the relevance vector are combined and spliced ​​into the surprise condition information.

4. The method according to claim 1, wherein The step of predicting the scores of the cold-start items according to the cold-start embedding vector, the original item embedding vector, and the original user-item score to obtain a recommendation list includes: Calculating the similarity between the cold start embedding vector and the original item embedding vector; Multiplying the similarity by the original user-item score to predict the score of the cold start item; The scores of the cold start items are sorted to obtain the recommendation list.

5. The method according to claim 1, wherein The step of inputting the noise, the multi-dimensional surprise condition information, and the user interaction history data into a conditional generative adversarial network of an initial recommendation model to generate a cold start embedding vector includes: Using the multi-dimensional surprise condition information to guide sampling of the noise to generate virtual cold start data; Sampling the user interaction history data to generate real cold start data; Determining whether the virtual cold start data and the real cold start data meet the surprise condition information to obtain a determination result; Adversarial training is performed using the discrimination result and a preset objective function to obtain the cold start embedding vector.

6. The method according to claim 1, characterized in that The determining the surprise loss corresponding to the recommendation list according to the recommendation list and a preset surprise factor loss function includes: Determining the unexpectedness loss, novelty loss, timeliness loss, and relevance loss corresponding to the recommended items in the recommendation list based on the multi-dimensional surprise factor and the user interaction history data; The unexpectedness loss, the novelty loss, the timeliness loss, and the relevance loss are weightedly summed to obtain the surprise loss.

7. A training device for a recommendation model, characterized in that: include: A first determining unit is configured to determine a surprise project based on a multi-dimensional surprise factor, user interaction history data, and candidate projects, wherein the multi-dimensional surprise factor includes unexpectedness, novelty, timeliness, and relevance; a conversion unit, configured to determine multi-dimensional surprise condition information according to the surprise item; a generating unit, configured to input the noise, the multi-dimensional surprise condition information, and the user interaction history data into a conditional generative adversarial network of an initial recommendation model to generate a cold start embedding vector; a prediction unit, configured to predict scores of cold-start items based on the cold-start embedding vector, the original item embedding vector, and the original user-item score to obtain a recommendation list; A second determining unit is configured to determine a loss corresponding to the recommendation list according to the recommendation list and a preset loss function, wherein the preset loss function includes a preset surprise factor function, and the loss includes a surprise degree loss; A training unit is used to train the initial recommendation model by optimizing the loss to obtain a target recommendation model.

8. A recommendation method, characterized in that: include: Get historical user interaction data; The user interaction history data is input into a target recommendation model to output a recommendation list, wherein the target recommendation model is a target model obtained by using the method according to any one of claims 1 to 6.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

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

Images