Commodity electronic coupon recommendation method and device, equipment and storage medium
By using adaptive fairness filters and fairness discriminators in the electronic voucher recommendation system, the popularity deviation problem is solved, personalized recommendations are achieved, and user experience and service quality are improved.
Patent Information
- Application Number
- CN202411745422.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-29
- Publication Date
- 2025-03-07
AI Technical Summary
There is a bias in the popularity of the electronic coupon recommendation system, which leads to users receiving too many similar popular products recommendations, lacking personalization and freshness, reducing user satisfaction and trust. At the same time, niche products and new user groups do not get the right recommendation opportunities.
By obtaining the preference information of the target user, using the pre-trained electronic voucher preference extraction model, including an adaptive fairness filter and fairness discriminator, the user's biased product characteristics are extracted, and the recommended target products are determined through these characteristics, and electronic vouchers are issued to the user.
Effectively remove popularity deviations, enhance personalized recommendation effects, improve user experience and service quality, and ensure that niche products and new user groups get the right recommendation opportunities.
Smart Images
Figure CN119579247A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of information technology, and more specifically, to a method, device, equipment and storage medium for recommending electronic commodity coupons. Background Art
[0002] With the booming development of e-commerce, competition among e-commerce platforms is becoming increasingly fierce. The electronic voucher system has emerged as the times require, becoming a key means for merchants to stimulate user consumption and enhance user stickiness in various promotional activities, and it occupies an important position in e-commerce operation strategies.
[0003] During the operation of the e-voucher recommendation system, popularity bias occurred. Because some products are popular in the market or some user groups are highly active, these popular products or users are often more likely to be selected in the recommendation system based on their inherent popularity, thereby obtaining more recommendation opportunities. For example, popular online celebrity products or users who shop frequently and spend a large amount of money will frequently appear in the e-voucher recommendation list.
[0004] However, popularity bias will cause most users to frequently receive similar popular product recommendations, which lacks personalization and freshness, and easily leads to boredom, reducing satisfaction and trust in platform recommendations. Moreover, popularity bias prevents some niche products or new user groups from receiving the attention and recommendation opportunities they deserve, resulting in uneven resource allocation, which is not conducive to the diversified promotion of products and the retention and growth of new users. Summary of the invention
[0005] In view of the above problems, the present application provides a method, device, equipment and storage medium for recommending electronic coupons for commodities to remove popularity deviation, enhance personalized recommendation effect, and improve user experience and service quality.
[0006] In order to achieve the above objectives, the specific plan is proposed as follows:
[0007] A method for recommending a commodity electronic voucher, comprising:
[0008] Obtain target user preference information;
[0009] Input the target user's preference information into a pre-trained e-voucher preference extraction model, and output the target user's preferred product features;
[0010] Determining a target product to be recommended to the target user according to the biased product features;
[0011] issuing an electronic voucher for the target product to the target user;
[0012] Wherein, the electronic voucher preference extraction model includes an adaptive fairness filter and a fairness discriminator;
[0013] The training process of the electronic voucher preference extraction model includes:
[0014] Obtaining a user product review and rating information set, wherein the user product review and rating information set includes a user information set, a product information set, and a review and rating information set, wherein the review and rating information set includes multiple review and rating samples, and each review and rating sample includes a single user's review and rating of a single product;
[0015] Extracting user preference features and biased product features based on the user product review rating information set;
[0016] Debiasing and reconstructing the user preference representation and the biased product feature by using the adaptive fairness filter to obtain a user reconstructed preference feature and a product reconstructed feature to train the adaptive fairness filter;
[0017] The fairness discriminator is used to perform popularity deviation detection on the user reconstruction preference feature and the commodity reconstruction feature to train the fairness discriminator.
[0018] Optionally, based on the user product review rating information set, extracting user preference features and biased product features includes:
[0019] Based on the user product review rating information set, initialize the user embedding matrix and the product embedding matrix. The user embedding matrix is:
[0020]
[0021] in, is the user embedding matrix, is the number of users, is the embedding dimension;
[0022] The commodity embedding matrix is:
[0023]
[0024] in, is the product embedding matrix, is the quantity of goods;
[0025] Construct a user product review set, the user product review set is:
[0026]
[0027] in, is the user product review collection, For users to Comments;
[0028] The user embedding matrix and the product embedding matrix are trained using the existing BERT pre-trained model, and comment embedding is generated by encoding all comments in the user product review set;
[0029] In the process of training the user embedding matrix and the product embedding matrix by the BERT pre-training model, the user embedding matrix and the product embedding matrix are updated by extracting scoring features, wherein the scoring features are extracted by the following formula:
[0030]
[0031] in, is the sigmoid activation function, is the comment of user i on product j, is the number of stacked layers of the feature extractor of the BERT pre-trained model, For the The embedding of the jth item in the layer, Neighbor nodes of the i-th product, is the neighbor node of product j, r is the score in the preset score set, Based on the rating The first learnable weight matrix of the layer, is the first scalar weight, Based on the rating The learnable second weight matrix of the layer, is the second scalar weight;
[0032] The embedding of target item j is learned from neighboring user i using the following formula:
[0033]
[0034] Using the following formula, the polymerization User embedding and product embedding of the layer:
[0035]
[0036] in, For the User i embedding at layer, For the The product j of the layer is embedded, For the The learnable weight matrix of the layer, For all neighbor products with a rating of r for user i, For all neighbor users who rated product i as r;
[0037] The user embedding of the last layer of the feature extractor is determined as a user preference representation, and the product embedding of the last layer of the feature extractor is determined as a biased product feature.
[0038] Optionally, debiasing and reconstructing the user preference representation and the biased product feature by the adaptive fairness filter to obtain a user reconstructed preference feature and a product reconstructed feature to train the adaptive fairness filter includes:
[0039] The adaptive fairness filter is used to perform nonlinear transformation on the user preference representation and the biased commodity feature through the following formula:
[0040]
[0041] in, is the commodity feature after nonlinear transformation of the biased commodity, is the product feature after nonlinear transformation. It is The product weight matrix of the layer can be learned. It is The user-learnable weight matrix of the layer, is the learnable first bias, is the learnable second bias;
[0042] Based on the user information set and the comment rating information set, a user scaling operator and a movement operator are generated, and based on the product information set and the comment rating information set, a product scaling operator and a movement operator are generated;
[0043] updating a user learnable weight matrix of the adaptive fairness filter by using a scaling operator of the user and a moving operator of the user, and updating a commodity learnable weight matrix of the adaptive fairness filter by using a scaling operator of the commodity and a moving operator of the commodity, so as to train the adaptive fairness filter;
[0044] Using the following formula, the user preference representation is reconstructed according to the specific weight matrix of the user and the specific bias parameter of the user i to obtain the user reconstructed preference feature, and the biased product feature is reconstructed according to the specific weight matrix of the product and the characteristic bias parameter of the product to obtain the product reconstructed feature:
[0045]
[0046] in, is the specific weight matrix of user i, is the specific bias parameter of user i, is the specific weight matrix of product j, is the specific bias parameter for product j, is the user preference representation, is the biased commodity feature, Reconstruct preference features for users, Reconstruct features for products.
[0047] Optionally, generating a user's scaling operator and a moving operator based on the user information set and the comment rating information set includes:
[0048] The user's scaling operator and movement operator are generated using the following formula:
[0049]
[0050] in, is the scaling operator of the user, is the user's mobile operator, is a learnable scaling bias, is the learnable mobile bias.
[0051] Optionally, updating the user learnable weight matrix of the adaptive fairness filter by using the user's scaling operator and the user's movement operator comprises:
[0052] Using the user's scaling operator and the user's movement operator, the user's learnable weight matrix is updated by the following formula:
[0053]
[0054] in, is the specific weight matrix of user i, is the specific bias parameter of user i;
[0055] Optionally, the fairness discriminator includes an explicit fairness discriminator and an implicit fairness discriminator;
[0056] The fairness discriminator is used to perform popularity deviation detection on the user reconstruction preference feature and the commodity reconstruction feature to train the fairness discriminator, including:
[0057] The popularity deviation detection is performed on the user reconstruction preference feature by the explicit fairness discriminator:
[0058]
[0059] in, is the group prediction probability of user i, is the activation function, is the learnable third weight matrix, is a learnable third bias;
[0060] The popularity deviation detection is performed on the commodity reconstruction feature by the explicit fairness discriminator:
[0061]
[0062] in, is the predicted probability of grouping of product j, is the learnable fourth weight matrix, is a learnable fourth bias;
[0063] The explicit fairness discriminator is trained under the maximum logarithmic probability formula, and the maximum logarithmic probability formula is:
[0064]
[0065] in, is the true value of the user group to which user i belongs, is the number of all user groups, for The mth element in represents the probability that user i belongs to the mth user group;
[0066] The existing large language model is used to extract the user's historical preferences from the user's historical interactive product reviews, and the user's historical preferences are used to enhance the user's reconstructed preference features to obtain enhanced user's reconstructed preference features;
[0067] Based on the user reconstruction preference features, a user prediction score is determined, and the user prediction score is normalized using the implicit fairness discriminator to train the implicit fairness discriminator.
[0068] Optionally, determining a user prediction score based on the user reconstruction preference feature, and normalizing the user prediction score using the implicit fairness discriminator to train the implicit fairness discriminator includes:
[0069] A reference scale is obtained by averaging the ratings of all products given by the user;
[0070] The user preference representation and unbiased product features are used as input to predict the rating of user i on product j, and the predicted rating is obtained. ;
[0071] Constructing a user scale based on the predicted ratings of the user i for all commodities;
[0072] Based on the reference scale and the user scale, construct a scoring sample set;
[0073] Using the implicit fairness discriminator, it is determined whether the scoring samples in the scoring sample set are derived from the user scale by the following formula:
[0074]
[0075] in, is the probability that the rating sample comes from the user scale, is the learnable fifth weight matrix, To learn the fifth bias, is the scoring sample set;
[0076] Under the maximum log-likelihood formula, the implicit fairness discriminator is updated and trained. The maximum log-likelihood formula is:
[0077]
[0078] in, It is the value corresponding to the actual source to which the scoring sample belongs.
[0079] A commodity electronic voucher recommendation device, comprising:
[0080] A preference information acquisition unit, used to acquire preference information of a target user;
[0081] A biased commodity feature output unit, used to input the preference information of the target user into a pre-trained e-voucher preference extraction model, and output the biased commodity features of the target user;
[0082] a target commodity determining unit, configured to determine a target commodity to be recommended to the target user according to the biased commodity feature;
[0083] An electronic voucher issuing unit, used for issuing an electronic voucher for the target product to the target user;
[0084] Wherein, the electronic voucher preference extraction model includes an adaptive fairness filter and a fairness discriminator;
[0085] An information set acquisition unit is used to acquire a user product review and rating information set, wherein the user product review and rating information set includes a user information set, a product information set, and a review and rating information set, wherein the review and rating information set includes a plurality of review and rating samples, and each review and rating sample includes a single user's review and rating of a single product;
[0086] A preference extraction unit, used to extract user preference features and biased product features based on the user product review rating information set;
[0087] a debiasing and reconstruction unit, configured to debias and reconstruct the user preference representation and the biased commodity feature through the adaptive fairness filter to obtain a user reconstructed preference feature and a commodity reconstructed feature to train the adaptive fairness filter;
[0088] The deviation detection unit is used to perform popularity deviation detection on the user reconstruction preference feature and the commodity reconstruction feature through the fairness discriminator to train the fairness discriminator.
[0089] A commodity electronic voucher recommendation device, comprising a memory and a processor;
[0090] The memory is used to store programs;
[0091] The processor is used to execute the program to implement the various steps of the above-mentioned commodity electronic voucher recommendation method.
[0092] A storage medium stores a computer program, which, when executed by a processor, implements the various steps of the above-mentioned method for recommending electronic commodity coupons.
[0093] Through the above technical solution, this application extracts user preference features and product features from user comments and interaction history through deep learning and natural language processing technology, uses embedding technology to achieve accurate modeling, and introduces adaptive fairness filters to perform nonlinear transformations on user preferences and product features. At the same time, it is equipped with explicit and implicit fairness discriminators to continuously monitor and adjust the fairness of recommendations, thereby effectively removing popularity bias, extracting users' real preferences through large language models, and further enhancing the personalized effect of recommendations. It can be seen that it not only improves user experience and promotes the fairness of services, but also helps to improve the accuracy of commercial electronic voucher recommendations. BRIEF DESCRIPTION OF THE DRAWINGS
[0094] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:
[0095] Figure 1 A schematic diagram of a process for implementing a product electronic voucher recommendation provided in an embodiment of the present application;
[0096] Figure 2 A schematic diagram of a process for implementing a training e-voucher preference extraction model provided in an embodiment of the present application;
[0097] Figure 3 It is a schematic diagram of the logical relationship for implementing the recommendation of electronic coupons for commodities provided in an embodiment of the present application.
[0098] Figure 4 It is a structural diagram of the electronic voucher preference extraction for mall users provided in an embodiment of the present application.
[0099] Figure 5 It is a schematic diagram of the structure of the adaptive fairness filter provided in the embodiment of the present application to remove feature bias.
[0100] Figure 6 It is a structural diagram of the electronic voucher debiased recommendation based on prompt enhancement provided in an embodiment of the present application.
[0101] Figure 7 A schematic diagram of a device structure for implementing a product electronic voucher recommendation provided in an embodiment of the present application;
[0102] Figure 8 A schematic diagram of the structure of a device for implementing electronic commodity voucher recommendations provided in an embodiment of the present application. DETAILED DESCRIPTION
[0103] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0104] The present application solution can be implemented based on a terminal with data processing capabilities, which can be a computer, cloud, server, etc.
[0105] Next, combine Figure 1 The commodity electronic voucher recommendation method of the present application may include the following steps:
[0106] Step S110: Obtain preference information of the target user.
[0107] Step S120: input the target user's preference information into a pre-trained e-voucher preference extraction model, and output the target user's preferred product features.
[0108] Step S130: Determine the target product to be recommended to the target user according to the biased product feature.
[0109] Step S140: issuing an electronic voucher for the target product to the target user.
[0110] Among them, the training process of the electronic voucher preference extraction model is as follows: Figure 2 As shown, the training process may include:
[0111] Step S210: Obtain a user product review rating information set.
[0112] The user product review and rating information set includes a user information set, a product information set and a review and rating information set. The review and rating information set includes multiple review and rating samples, and each review and rating sample includes a single user's review and rating of a single product.
[0113] Step S220: extracting user preference features and biased product features based on the user product review rating information set.
[0114] Step S230: Debiasing and reconstructing the user preference representation and the biased product feature through the adaptive fairness filter to obtain a user reconstructed preference feature and a product reconstructed feature to train the adaptive fairness filter.
[0115] Step S240: Perform popularity deviation detection on the user reconstructed preference features and the product reconstructed features through the fairness discriminator to train the fairness discriminator.
[0116] For example, the implementation process of this application for recommending product electronic coupons can be referred to Figure 3 . First, for the extraction of e-voucher preferences of mall users, basic user attributes, user consumption preferences, user association relationships and user behavior preferences can be collected when modeling user portraits. Basic product attributes, product labels, product review texts and product value scores can be collected when modeling product descriptions. The user information set and product information set are handed over to the user preference extraction module, which extracts biased user preference features and biased product features. Then, the adaptive fairness filter removes feature bias and obtains unbiased user preference features and unbiased product features. Finally, debiased recommendations are performed based on the suggestive enhancement module.
[0117] The electronic voucher recommendation method for commodities provided in the present embodiment obtains the preference information of the target user, inputs the preference information of the target user into a pre-trained electronic voucher preference extraction model, outputs the biased commodity features of the target user, determines the target commodity recommended to the target user based on the biased commodity features, and issues the electronic voucher of the target commodity to the target user, wherein the electronic voucher preference extraction model obtains the user commodity review and rating information set, extracts the user preference features and the biased commodity features based on the user commodity review and rating information set, debias-reconstructs the user preference representation and the biased commodity features through the adaptive fairness filter, obtains the user reconstructed preference features and the commodity reconstructed features, and trains the adaptive fairness filter, performs popularity deviation detection on the user reconstructed preference features and the commodity reconstructed features through the fairness discriminator, and trains the fairness discriminator to obtain. As can be seen, firstly, through deep learning and natural language processing technology, user preference features and product features are extracted from user comments and interaction history, and an adaptive fairness filter is introduced. This module performs nonlinear transformation on user preferences and product features, and is equipped with explicit and implicit fairness discriminators to continuously monitor and adjust the fairness of recommendations, effectively remove popularity bias, enhance personalized recommendation effects, and improve user experience and service quality. .
[0118] In some embodiments of the present application, the process of extracting user preference features and biased product features based on the user product review rating information set in step S220 is introduced, and the process may include:
[0119] S1. Initialize a user embedding matrix and a product embedding matrix based on the user product review rating information set.
[0120] It is understandable that this application can be oriented to mall users, and the electronic voucher preference extraction model can extract biased user preferences and biased product features from the user's review text and the product review text, such as Figure 4 shown.
[0121] Among them, the user embedding matrix is:
[0122]
[0123] in, is the user embedding matrix, is the number of users, is the embedding dimension;
[0124] The commodity embedding matrix is:
[0125]
[0126] in, is the product embedding matrix, The quantity of goods.
[0127] Specifically, represents the embedding of the i-th user, Represents the embedding of the jth item.
[0128] S2. Construct a user product review set, where the user product review set is:
[0129]
[0130] in, is the user product review collection, For users to Comments.
[0131] S3. Use the existing BERT pre-trained model to train the user embedding matrix and the product embedding matrix, and generate comment embedding by encoding all comments in the user product review set.
[0132] In the process of training the user embedding matrix and the product embedding matrix by the BERT pre-training model, the user embedding matrix and the product embedding matrix are updated by extracting scoring features, wherein the scoring features are extracted by the following formula:
[0133]
[0134] in, is the sigmoid activation function, is the comment of user i on product j, is the number of stacked layers of the feature extractor of the BERT pre-trained model, For the The embedding of the jth item in the layer, Neighbor nodes of the i-th product, is the neighbor node of product j, r is the score in the preset score set, Based on the rating The first learnable weight matrix of the layer, is the first scalar weight, Based on the rating The learnable second weight matrix of the layer, is the second scalar weight.
[0135] It is understandable that during training, the embedding matrices of users and items are learnable, so user embeddings and item embeddings are updated, and since the reviews are generated by the pre-trained model, the review embeddings will be frozen to reduce training time and memory usage.
[0136] In the data, each interaction between a user and an item includes the user's rating and comments on the item. Compared with numerical ratings, textual comments provide more detailed semantic information and have the following advantages: First, the comment semantics contain rich user preferences and item attributes, which helps to learn the embedding representation of users and items more effectively. Second, detailed comment semantics can help the model understand the user's preference for items, and then be used to readjust the relationship weight between users and items. Therefore, we use the user's comments on items and neighbor items j to update user i.
[0137] S4. Learn the embedding of target product j from neighbor user i using the following formula:
[0138] .
[0139] It can be understood that the embedding of target item j is learned from neighboring user i in a similar way to S3.
[0140] S5, using the following formula, aggregate the User embedding and product embedding of the layer:
[0141]
[0142] in, For the User i embedding at layer, For the The product j of the layer is embedded, For the The learnable weight matrix of the layer, For all neighbor products with a rating of r for user i, For all neighbor users who rated product i as r.
[0143] S6. Determine the user embedding of the last layer of the feature extractor as a user preference representation, and determine the product embedding of the last layer of the feature extractor as a biased product feature.
[0144] In some embodiments of the present application, the above step S230, the process of debiasing and reconstructing the user preference representation and the biased product features through the adaptive fairness filter to obtain the user reconstructed preference features and the product reconstructed features to train the adaptive fairness filter is introduced. The logic of the adaptive fairness filter to debias and reconstruct the extracted biased user preference features and the biased product features can refer to Figure 5 Specifically, the process may include:
[0145] S1. Using the adaptive fairness filter, a nonlinear transformation is performed on the user preference representation and the biased commodity feature by the following formula:
[0146]
[0147] in, is the commodity feature after nonlinear transformation of the biased commodity, is the product feature after nonlinear transformation. It is The product weight matrix of the layer can be learned. It is The user-learnable weight matrix of the layer, is the learnable first bias, is the learnable second bias.
[0148] S2. Generate a user's zoom operator and a movement operator based on the user information set and the comment rating information set, and generate a product's zoom operator and a movement operator based on the product information set and the comment rating information set.
[0149] It is understandable that considering that using a unified and shared nonlinear change to remove popularity bias for all user or item features is neither flexible nor targeted, a scaling and shifting operator is generated for each user and item, and the parameters of the original multi-layer perceptron are adjusted to adapt to different user or item inputs. The operator can be specifically generated based on the original user preferences and product features.
[0150] S3. Updating the user learnable weight matrix of the adaptive fairness filter by using the scaling operator of the user and the movement operator of the user, and updating the commodity learnable weight matrix of the adaptive fairness filter by using the scaling operator of the commodity and the movement operator of the commodity, so as to train the adaptive fairness filter.
[0151] S4. Using the following formula, the user preference representation is reconstructed according to the specific weight matrix of the user and the specific bias parameter of the user i to obtain the user reconstructed preference feature, and the biased product feature is reconstructed according to the specific weight matrix of the product and the characteristic bias parameter of the product to obtain the product reconstructed feature:
[0152]
[0153] in, is the specific weight matrix of user i, is the specific bias parameter of user i, is the specific weight matrix of product j, is the specific bias parameter for product j, is the user preference representation, is the biased commodity feature, Reconstruct preference features for users, Reconstruct features for products.
[0154] In some embodiments of the present application, the process of generating the user's scaling operator and moving operator based on the user information set and the comment rating information set mentioned in the above embodiment is introduced, and the process may include:
[0155] The user's scaling operator and movement operator are generated using the following formula:
[0156]
[0157] in, is the scaling operator of the user, is the user's mobile operator, is a learnable scaling bias, is the learnable mobile bias.
[0158] Based on this, the method of updating the user learnable weight matrix of the adaptive fairness filter by using the user's scaling operator and the user's movement operator mentioned in the above embodiment may specifically include:
[0159] Using the user's scaling operator and the user's movement operator, the user's learnable weight matrix is updated by the following formula:
[0160]
[0161] in, is the specific weight matrix of user i, is the specific bias parameter of user i.
[0162] In some embodiments of the present application, the fairness discriminator mentioned in the above embodiments is further introduced. The fairness discriminator may include an explicit fairness discriminator and an implicit fairness discriminator.
[0163] Among them, the explicit fairness discriminator pursues fairness from a local perspective and directly detects the residual popularity deviation of each user or product. The implicit fairness discriminator takes into account that different users have different rating standards, which will especially affect the fairness of the product in terms of exposure (for example, because a product with a high predicted score will still have a relatively low score for a user who is accustomed to giving a lower score). Therefore, the implicit fairness discriminator can normalize the user's predicted score and reduce the popularity deviation caused by different scales from a global perspective.
[0164] Based on this, the process of performing popularity deviation detection on the user reconstructed preference features and the product reconstructed features by the fairness discriminator mentioned in the above embodiment to train the fairness discriminator is introduced. The process may include:
[0165] S1. Performing popularity deviation detection on the user reconstruction preference feature through the explicit fairness discriminator:
[0166]
[0167] in, is the group prediction probability of user i, is the activation function, is the learnable third weight matrix, is a learnable third bias.
[0168] S2. Performing popularity deviation detection on the commodity reconstruction features through the explicit fairness discriminator:
[0169]
[0170] in, is the predicted probability of grouping of product j, is the learnable fourth weight matrix, is a learnable fourth bias.
[0171] S3. Training the explicit fairness discriminator under the maximum logarithmic probability formula, the maximum logarithmic probability formula is:
[0172]
[0173] in, is the true value of the user group to which user i belongs, is the number of all user groups, for The mth element in represents the probability that user i belongs to the mth user group.
[0174] S4. Utilize the existing large language model to extract the user's historical preferences from the user's historical interactive product reviews, and enhance the user's reconstructed preference features through the user's historical preferences to obtain enhanced user's reconstructed preference features.
[0175] Specifically, user historical preferences can represent the user's real preferences. Figure 6 As shown in the figure, the preference feature filtering of users and products without popularity bias can adopt the strategy of manually designing user preference mining templates. These templates cover users' comments on products and their evaluations of specific products, so that the model can further utilize the key preference information contained in the comment text. On this basis, the pre-trained general knowledge large language model is used to generate a clear natural language description of user preferences. Among them, a specific template example input to the large language model is as follows:
[0176] "user The following comments were made on the product: , who has a high rating for the following items , please output his preferences . "
[0177] Finally, LLM can generate a natural language description of preferences for each user, such as user i’s preference description is , contains natural language descriptions of N preferences. We also use the existing BERT pre-trained model to encode N preference descriptions and generate preference description embeddings. The preference description embedding of user i is defined as The embeddings are pooled and the potential preference description of user i is aggregated by the following formula:
[0178]
[0179] Then, a multi-layer perceptron is used to capture the unbiased user preference features and the complex nonlinear relationship between the latent preference descriptions, and the two are fused to generate a more complex combined representation.
[0180] S5. Determine a user predicted score based on the user reconstructed preference feature, and normalize the user predicted score using the implicit fairness discriminator to train the implicit fairness discriminator.
[0181] Specifically, the process of determining the user predicted score based on the user reconstruction preference feature, and normalizing the user predicted score by using the implicit fairness discriminator to train the implicit fairness discriminator may include:
[0182] S51. Obtain a reference scale by averaging the ratings given by the users to all commodities.
[0183] S52: Taking the user preference representation and the unbiased product features as input, predicting the rating of user i on product j, and obtaining the predicted rating .
[0184] S53: construct a user scale based on the predicted ratings of the user i for all commodities.
[0185] S54: Construct a scoring sample set based on the reference scale and the user scale.
[0186] S55: Using the implicit fairness discriminator, determine whether the scoring samples in the scoring sample set are derived from the user scale by the following formula:
[0187]
[0188] in, is the probability that the rating sample comes from the user scale, is the learnable fifth weight matrix, To learn the fifth bias, is the scoring sample set.
[0189] S56. Update and train the implicit fairness discriminator under the maximum log-likelihood formula, where the maximum log-likelihood formula is:
[0190]
[0191] in, It is the value corresponding to the actual source to which the scoring sample belongs. If the sample comes from the user scale, it is 1; if the sample comes from the reference scale, it is 0.
[0192] For example, for user i, we first average the ratings of all products given by user i to obtain a reference scale, and then combine the user preferences described by the fusion preference and the unbiased product features. As input, the multi-layer perceptron predicts the rating of user i on product j, and all predicted ratings of user i are represented as a user scale. Subsequently, a rating set with a sample size of K is uniformly constructed from the reference scale and the user scale. , and then use the implicit fairness discriminator to distinguish whether the rating samples are from the reference scale or the user scale.
[0193] The following is a description of an apparatus for implementing electronic voucher recommendation for goods provided in an embodiment of the present application. The apparatus for implementing electronic voucher recommendation for goods described below and the method for implementing electronic voucher recommendation for goods described above can be referenced to each other.
[0194] See also Figure 7 , Figure 7 This is a schematic diagram of the structure of a device for implementing electronic voucher recommendations for commodities disclosed in an embodiment of the present application.
[0195] like Figure 7 As shown, the device may include:
[0196] A preference information acquisition unit 11, used to acquire preference information of a target user;
[0197] The biased commodity feature output unit 12 is used to input the preference information of the target user into a pre-trained electronic voucher preference extraction model, and output the biased commodity features of the target user;
[0198] A target product determination unit 13 is used to determine a target product to be recommended to the target user according to the biased product feature;
[0199] An electronic voucher issuing unit 14, configured to issue an electronic voucher for the target product to the target user;
[0200] Wherein, the electronic voucher preference extraction model includes an adaptive fairness filter and a fairness discriminator;
[0201] An information set acquisition unit 15 is used to acquire a user product review and rating information set, wherein the user product review and rating information set includes a user information set, a product information set, and a review and rating information set, wherein the review and rating information set includes a plurality of review and rating samples, and each review and rating sample includes a single user's review and rating of a single product;
[0202] A preference extraction unit 16, configured to extract user preference features and biased product features based on the user product review rating information set;
[0203] A debiasing and reconstruction unit 17, configured to debias and reconstruct the user preference representation and the biased product feature through the adaptive fairness filter to obtain a user reconstructed preference feature and a product reconstructed feature to train the adaptive fairness filter;
[0204] The deviation detection unit 18 is used to perform popularity deviation detection on the user reconstruction preference feature and the commodity reconstruction feature through the fairness discriminator to train the fairness discriminator.
[0205] The description of other units refers to the introduction of the corresponding method part of this application, which will not be repeated here.
[0206] The device for recommending electronic commodity coupons provided in the embodiment of the present application can be applied to electronic commodity coupon recommendation devices, such as terminals: mobile phones, computers, etc. Optionally, Figure 8 The hardware structure diagram of the commodity electronic voucher recommendation device is shown. Figure 8 ,The hardware structure of the commodity electronic voucher recommendation device may include: at least one processor 1, at least one communication interface 2, at least one memory 3 and at least one communication bus 4;
[0207] In the embodiment of the present application, the number of the processor 1, the communication interface 2, the memory 3, and the communication bus 4 is at least one, and the processor 1, the communication interface 2, and the memory 3 communicate with each other through the communication bus 4;
[0208] The processor 1 may be a central processing unit CPU, or an application-specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present invention, etc.;
[0209] The memory 3 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory;
[0210] The memory stores a program, and the processor can call the program stored in the memory, wherein the program is used to:
[0211] Obtain target user preference information;
[0212] Input the target user's preference information into a pre-trained e-voucher preference extraction model, and output the target user's preferred product features;
[0213] Determining a target product to be recommended to the target user according to the biased product features;
[0214] issuing an electronic voucher for the target product to the target user;
[0215] Wherein, the electronic voucher preference extraction model includes an adaptive fairness filter and a fairness discriminator;
[0216] The training process of the electronic voucher preference extraction model includes:
[0217] Obtaining a user product review and rating information set, wherein the user product review and rating information set includes a user information set, a product information set, and a review and rating information set, wherein the review and rating information set includes multiple review and rating samples, and each review and rating sample includes a single user's review and rating of a single product;
[0218] Extracting user preference features and biased product features based on the user product review rating information set;
[0219] Debiasing and reconstructing the user preference representation and the biased product feature by using the adaptive fairness filter to obtain a user reconstructed preference feature and a product reconstructed feature to train the adaptive fairness filter;
[0220] The fairness discriminator is used to perform popularity deviation detection on the user reconstruction preference feature and the commodity reconstruction feature to train the fairness discriminator.
[0221] Optionally, the detailed functions and extended functions of the program may refer to the above description.
[0222] The embodiment of the present application further provides a storage medium, which may store a program suitable for execution by a processor, wherein the program is used to:
[0223] Obtain target user preference information;
[0224] Input the target user's preference information into a pre-trained e-voucher preference extraction model, and output the target user's preferred product features;
[0225] Determining a target product to be recommended to the target user according to the biased product features;
[0226] issuing an electronic voucher for the target product to the target user;
[0227] Wherein, the electronic voucher preference extraction model includes an adaptive fairness filter and a fairness discriminator;
[0228] The training process of the electronic voucher preference extraction model includes:
[0229] Obtaining a user product review and rating information set, wherein the user product review and rating information set includes a user information set, a product information set, and a review and rating information set, wherein the review and rating information set includes multiple review and rating samples, and each review and rating sample includes a single user's review and rating of a single product;
[0230] Extracting user preference features and biased product features based on the user product review rating information set;
[0231] Debiasing and reconstructing the user preference representation and the biased product feature by using the adaptive fairness filter to obtain a user reconstructed preference feature and a product reconstructed feature to train the adaptive fairness filter;
[0232] The fairness discriminator is used to perform popularity deviation detection on the user reconstruction preference feature and the commodity reconstruction feature to train the fairness discriminator.
[0233] Optionally, the detailed functions and extended functions of the program may refer to the above description.
[0234] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.
[0235] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can refer to each other.
[0236] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for recommending electronic commodity coupons, characterized in that: include: Obtain target user preference information; Input the target user's preference information into a pre-trained e-voucher preference extraction model, and output the target user's preferred product features; Determining a target product to be recommended to the target user according to the biased product features; issuing an electronic voucher for the target product to the target user; Wherein, the electronic voucher preference extraction model includes an adaptive fairness filter and a fairness discriminator; The training process of the electronic voucher preference extraction model includes: Obtaining a user product review and rating information set, wherein the user product review and rating information set includes a user information set, a product information set, and a review and rating information set, wherein the review and rating information set includes multiple review and rating samples, and each review and rating sample includes a single user's review and rating of a single product; Extracting user preference features and biased product features based on the user product review rating information set; Debiasing and reconstructing the user preference representation and the biased product feature by using the adaptive fairness filter to obtain a user reconstructed preference feature and a product reconstructed feature to train the adaptive fairness filter; The fairness discriminator is used to perform popularity deviation detection on the user reconstruction preference feature and the commodity reconstruction feature to train the fairness discriminator.
2. The method according to claim 1, characterized in that Based on the user product review rating information set, extract user preference features and biased product features, including: Based on the user product review rating information set, initialize the user embedding matrix and the product embedding matrix. The user embedding matrix is: in, is the user embedding matrix, is the number of users, is the embedding dimension; The commodity embedding matrix is: in, is the product embedding matrix, is the quantity of goods; Construct a user product review set, the user product review set is: in, is the user product review collection, For users to Comments; The user embedding matrix and the product embedding matrix are trained using the existing BERT pre-trained model, and comment embedding is generated by encoding all comments in the user product review set; In the process of training the user embedding matrix and the product embedding matrix by the BERT pre-training model, the user embedding matrix and the product embedding matrix are updated by extracting scoring features, wherein the scoring features are extracted by the following formula: in, is the sigmoid activation function, is the comment of user i on product j, is the number of stacked layers of the feature extractor of the BERT pre-trained model, For the The embedding of the jth item in the layer, Neighbor nodes of the i-th product, is the neighbor node of product j, r is the score in the preset score set, Based on the rating The first learnable weight matrix of the layer, is the first scalar weight, Based on the rating The learnable second weight matrix of the layer, is the second scalar weight; The embedding of target item j is learned from neighboring user i using the following formula: ； Using the following formula, the polymerization User embedding and product embedding of the layer: in, For the User i embedding at layer, For the The product j of the layer is embedded, For the The learnable weight matrix of the layer, For all neighbor products with a rating of r for user i, For all neighbor users who rated product i as r; The user embedding of the last layer of the feature extractor is determined as a user preference representation, and the product embedding of the last layer of the feature extractor is determined as a biased product feature.
3. The method according to claim 2, characterized in that Debiasing and reconstructing the user preference representation and the biased product feature by the adaptive fairness filter to obtain a user reconstructed preference feature and a product reconstructed feature to train the adaptive fairness filter, including: The adaptive fairness filter is used to perform nonlinear transformation on the user preference representation and the biased commodity feature through the following formula: in, is the commodity feature after nonlinear transformation of the biased commodity, is the product feature after nonlinear transformation. It is The product weight matrix of the layer can be learned. It is The user-learnable weight matrix of the layer, is the learnable first bias, is the learnable second bias; Based on the user information set and the comment rating information set, a user scaling operator and a movement operator are generated, and based on the product information set and the comment rating information set, a product scaling operator and a movement operator are generated; updating a user learnable weight matrix of the adaptive fairness filter by using a scaling operator of the user and a moving operator of the user, and updating a commodity learnable weight matrix of the adaptive fairness filter by using a scaling operator of the commodity and a moving operator of the commodity, so as to train the adaptive fairness filter; Using the following formula, the user preference representation is reconstructed according to the specific weight matrix of the user and the specific bias parameter of the user i to obtain the user reconstructed preference feature, and the biased product feature is reconstructed according to the specific weight matrix of the product and the characteristic bias parameter of the product to obtain the product reconstructed feature: in, is the specific weight matrix of user i, is the specific bias parameter of user i, is the specific weight matrix of product j, is the specific bias parameter for product j, is the user preference representation, is the biased commodity feature, Reconstruct preference features for users, Reconstruct features for products.
4. The method according to claim 3, characterized in that Based on the user information set and the comment rating information set, a user scaling operator and a moving operator are generated, including: The user's scaling operator and movement operator are generated using the following formula: in, is the scaling operator of the user, is the user's mobile operator, is a learnable scaling bias, is the learnable mobile bias.
5. The method according to claim 4, characterized in that Updating the user learnable weight matrix of the adaptive fairness filter by using the user's scaling operator and the user's movement operator includes: Using the user's scaling operator and the user's movement operator, the user's learnable weight matrix is updated by the following formula: in, is the specific weight matrix of user i, is the specific bias parameter of user i.
6. The method according to claim 3, characterized in that The fairness discriminator includes an explicit fairness discriminator and an implicit fairness discriminator; The fairness discriminator is used to perform popularity deviation detection on the user reconstruction preference feature and the commodity reconstruction feature to train the fairness discriminator, including: The popularity deviation detection is performed on the user reconstruction preference feature by the explicit fairness discriminator: in, is the group prediction probability of user i, is the activation function, is the learnable third weight matrix, is a learnable third bias; The popularity deviation detection is performed on the commodity reconstruction feature by the explicit fairness discriminator: in, is the predicted probability of grouping of product j, is the learnable fourth weight matrix, is a learnable fourth bias; The explicit fairness discriminator is trained under the maximum logarithmic probability formula, and the maximum logarithmic probability formula is: in, is the true value of the user group to which user i belongs, is the number of all user groups, for The mth element in represents the probability that user i belongs to the mth user group; The existing large language model is used to extract the user's historical preferences from the user's historical interactive product reviews, and the user's historical preferences are used to enhance the user's reconstructed preference features to obtain enhanced user's reconstructed preference features; Based on the user reconstruction preference features, a user prediction score is determined, and the user prediction score is normalized using the implicit fairness discriminator to train the implicit fairness discriminator.
7. The method according to claim 6, characterized in that Determining a user prediction score based on the user reconstruction preference feature, and normalizing the user prediction score using the implicit fairness discriminator to train the implicit fairness discriminator, including: A reference scale is obtained by averaging the ratings of all products given by the user; The user preference representation and unbiased product features are used as input to predict the rating of user i on product j, and the predicted rating is obtained. ; Constructing a user scale based on the predicted ratings of the user i for all commodities; Based on the reference scale and the user scale, construct a scoring sample set; Using the implicit fairness discriminator, it is determined whether the scoring samples in the scoring sample set are derived from the user scale by the following formula: in, is the probability that the rating sample comes from the user scale, is the learnable fifth weight matrix, To learn the fifth bias, is the scoring sample set; Under the maximum log-likelihood formula, the implicit fairness discriminator is updated and trained. The maximum log-likelihood formula is: in, It is the value corresponding to the actual source to which the scoring sample belongs.
8. A commodity electronic voucher recommendation device, characterized in that: include: A preference information acquisition unit, used to acquire preference information of a target user; A biased commodity feature output unit, used to input the preference information of the target user into a pre-trained e-voucher preference extraction model, and output the biased commodity features of the target user; a target commodity determining unit, configured to determine a target commodity to be recommended to the target user according to the biased commodity feature; An electronic voucher issuing unit, used for issuing an electronic voucher for the target product to the target user; Wherein, the electronic voucher preference extraction model includes an adaptive fairness filter and a fairness discriminator; An information set acquisition unit is used to acquire a user product review and rating information set, wherein the user product review and rating information set includes a user information set, a product information set, and a review and rating information set, wherein the review and rating information set includes a plurality of review and rating samples, and each review and rating sample includes a single user's review and rating of a single product; A preference extraction unit, used to extract user preference features and biased product features based on the user product review rating information set; a debiasing and reconstruction unit, configured to debias and reconstruct the user preference representation and the biased commodity feature through the adaptive fairness filter to obtain a user reconstructed preference feature and a commodity reconstructed feature to train the adaptive fairness filter; The deviation detection unit is used to perform popularity deviation detection on the user reconstruction preference feature and the commodity reconstruction feature through the fairness discriminator to train the fairness discriminator.
9. A commodity electronic voucher recommendation device, characterized in that: including memory and processor; The memory is used to store programs; The processor is used to execute the program to implement each step of the commodity electronic voucher recommendation method as described in any one of claims 1 to 7.
10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, each step of the method for recommending electronic merchandise vouchers as described in any one of claims 1 to 7 is implemented.