Rating Prediction Method and System for Recommendation System Based on Hierarchical Gating and Discriminant Network

Abstract

The present invention proposes a method and system for predicting ratings of a recommendation system based on hierarchical gating and discriminant networks, which relates to the technical field of recommendation systems. The problems addressed are as follows: In the prior art, the noise impact of irrelevant comments is not considered, the complexity and hierarchy between users and items are not captured, and there is a lack of supervision signals, resulting in poor accuracy of the rating prediction model. The method obtains information of the recommendation system; constructs a rating prediction model using a hierarchical gating network and a hierarchical discriminant network, and trains the rating prediction model based on the obtained information of the recommendation system; uses the trained rating prediction model to complete the final rating prediction of a user for an item. The present invention solves the problem of the noise impact of irrelevant comments, can capture the complexity and hierarchy between users and items, and at the same time generates sub-supervision signals through contrastive learning to improve the accuracy of rating prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of recommendation systems, and particularly relates to a recommendation system score prediction method and system based on hierarchical gating and discriminant networks. Background Technique

[0002] The statements in this part only provide background technical information related to the present invention and do not necessarily constitute prior art.

[0003] A recommendation system is a technology that screens and recommends potential associated items for users by analyzing user behavior, item attributes, and context information. Its core mechanisms include collaborative filtering (recommendation based on the similarity of user-item interactions), content filtering (matching based on item features), and hybrid methods (integrating multi-source data). Typical application scenarios include e-commerce (such as Taobao product recommendations), streaming platforms (such as iQIYI movie recommendations), and social media (such as Douyin video recommendations). The core goal of the recommendation system is to help users efficiently obtain information of interest through personalized recommendations, while improving the service quality and user stickiness of the platform.

[0004] As a core sub-task of the recommendation system, score prediction aims to predict the score value (usually a discrete value, such as 1-5 points) of an item that a user has not interacted with based on the user's historical behavior (such as ratings, clicks, purchases) and auxiliary information (such as user profiles, item descriptions, review content). Its input is a user-item pair, and the output is the predicted score. The modeling goal is to minimize the error between the predicted value and the true score (such as mean square error). The accuracy of score prediction directly determines the performance of the recommendation system.

[0005] Traditional recommendation systems mainly rely on explicit interactions (such as ratings, clicks) or implicit feedback (such as browsing duration) between users and items, and model user preferences through collaborative filtering or matrix factorization techniques. However, such methods usually face two core challenges: data sparsity and cold start. The introduction of user reviews provides a solution to the above challenges.

[0006] The fine-grained semantic information contained in reviews (such as "the headphone sound quality is clear and the bass is shocking" reflects the user's sensitivity to sound effects; "the battery life is long but the charging is slow" reveals the advantages and disadvantages of the product's functions) can supplement sparse interaction signals. By leveraging the rich semantic information in reviews, more comprehensive user and item representations can be developed, thereby improving the accuracy of recommendations. Therefore, the expressiveness of historical reviews in capturing user preferences has attracted extensive attention in the field of recommendation systems, and many review-based recommendation systems have emerged.

[0007] In the early stages of review-based recommendation systems, neural networks were used to independently generate review-based representations for users and items, and then score prediction was performed through user-item interaction modeling. More advanced review-based methods incorporate interaction modeling into the feature extraction stage, enabling the capture of the text semantics of specific user-item pairs and enhancing performance. For example, DeepCoNN uses a pair of convolutional neural networks to generate embeddings for users and items, and then combines these embeddings to predict the final score. Additionally, DAML focuses on improving the modeling of the interaction between users and items by extracting text features that are both local and mutual in the reviews. The following methods emphasize identifying informative text signals unique to each user-item pair. To address the problem of coarse-grained historical reviews, some methods utilize target reviews to assist in modeling user-item interactions.

[0008] The effectiveness of graph learning methods in learning representations from graph structures has been well demonstrated. It aims to reveal the high-order relationships between users and items, thereby providing a more refined modeling of user-item relationships for recommendation systems. In review-based recommendation systems, the user-item relationship is represented as a bipartite graph that connects users and items. To enhance the recommendation system, advanced techniques utilize graph learning to leverage the inherent graph structure. For example, GC-MC uses relational graph convolution to enhance the embeddings of users and items, treating score prediction as a link prediction problem in the user-item bipartite graph. RMG introduces a powerful framework that integrates the data of the user-item graph with the review content. This framework uses hierarchical attention networks to extract important features from the hierarchical structure of the reviews. Meanwhile, graph neural networks (GNNs) are used to map the interactions between users and items in the graph, with the aim of clarifying their mutual relationships. Similarly, RGCL proposes a score prediction framework centered on contrastive learning, leveraging the graph structure and incorporating review data. In this model, the reviews serve as connections in the information propagation process, facilitating the comprehensive utilization of the unique user-item graph structure. Additionally, the model combines contrastive learning to generate supervision signals, further enhancing the graph learning process using review data.

[0009] Although the above research work has made significant progress in improving rating prediction by leveraging the textual and structural features among users, projects, and reviews, it still faces three major challenges. The first challenge is the existence of noisy reviews. Due to the diversity and variability of user behavior, historical reviews may not always provide valuable signals for predicting specific targets. Although some work has proposed attention mechanisms to softly focus on important reviews, irrelevant text signals may still overwhelm useful signals, thus weakening the performance of the model. The second challenge lies in the adoption of a flat approach that treats historical reviews indiscriminately and simply concatenates them to represent users / projects. In fact, the correlation between a user's future behavior and historical behavior changes dynamically, which makes it challenging to accurately evaluate the usefulness of review information in predicting user behavior. Therefore, the flat approach may limit the discriminative understanding of users. The third challenge is the lack of supervision signals for selecting informative reviews. Existing work selects reviews in an attention-based manner by learning from the data, but the lack of supervision signals indicating review relevance may lead to unexplainable review selection and is prone to overfitting to the training data.

[0010] In summary, there are still some deficiencies in the prior art:

[0011] (1) Irrelevant reviews may introduce noise, which in turn has an adverse effect on the model performance.

[0012] (2) Existing methods regard review features as static single-grained representations and do not model hierarchical semantic associations, reducing the usefulness of review information and the accuracy of model evaluation.

[0013] (3) Due to the lack of sparse supervision signals, it is difficult for existing review-based recommendation system rating prediction techniques to effectively evaluate the usefulness of reviews. Summary of the Invention

[0014] To overcome the above deficiencies of the prior art, the present invention provides a rating prediction method and system for a recommendation system based on hierarchical gating and discriminant networks, which combines hierarchical gating and hierarchical discriminant networks. By hierarchical gating, noisy reviews are filtered and relevant text information is extracted. By hierarchical discriminant networks, supervision signals are generated at the local and global levels to ensure accurate review selection, thereby improving the accuracy of rating prediction.

[0015] To achieve the above object, one or more embodiments of the present invention provide the following technical solutions:

[0016] The first aspect of the present invention provides a rating prediction method for a recommendation system based on hierarchical gating and discriminant networks, including:

[0017] Obtain recommendation system information, where the recommendation system information includes user information, item information, the user's historical reviews, item-related historical reviews, the user's historical ratings for items, and the user's target reviews for items;

[0018] Construct a rating prediction model using a hierarchical gated network and a hierarchical discriminative network, and train the rating prediction model based on the obtained recommendation system information;

[0019] Use the trained rating prediction model to obtain the final predicted rating of the user for the item, that is, complete the rating prediction of the recommendation system;

[0020] Among them, the rating prediction model includes a hierarchical gated network, a hierarchical discriminative network, and a multi-layer perceptron;

[0021] During the training process, input the recommendation system information into the hierarchical gated network to obtain the user's preference vector and item embedding vector;

[0022] Through the hierarchical discriminative network, generate self-supervised signals at the local and global levels to optimize the user's preference vector and item embedding vector, and obtain the user's optimized preference vector and item optimized vector;

[0023] Input the user's optimized preference vector and item optimized vector into the multi-layer perceptron to obtain the final predicted rating.

[0024] As an implementation, input the recommendation system information into the hierarchical gated network to obtain the user's preference vector and item embedding vector, where the hierarchical gated network includes a local gated module and a global gated module, and the specific process is as follows:

[0025] Input the propulsion system information into the local gated module to obtain the user's summary review vector and item-related summary review vector respectively;

[0026] Input the user information, item information, the user's summary review vector, and item-related summary review vector into the global gated module to obtain the user's preference vector and item embedding vector respectively.

[0027] As an implementation, obtain the user's summary review vector and item-related summary review vector, and the specific process is as follows:

[0028] Convert each comment in the historical reviews into the user's comment vector and item-related historical review vector respectively;

[0029] Calculate the correlation score between the target item and the user's comment;

[0030] Based on the correlation score, through the local gating function, obtain the user's summary review vector and item-related summary review vector.

[0031] As an implementation, obtain the user's preference vector and item embedding vector. The specific process is as follows:

[0032] In the global gating module, apply a fully connected layer to the user's summary comment vector and item embedding vector to respectively obtain the activation signal of the user's historical comments and the activation signal of the item-related historical comments;

[0033] According to the activation signal of the user's historical comments and the activation signal of the item-related historical comments, through the global gating function, obtain the user's preference vector and item embedding vector.

[0034] As an implementation, through a hierarchical discriminative network, generate self-supervised signals at the local and global levels to optimize the user's preference vector and item embedding vector, and obtain the user's optimized preference vector and item optimized vector. Among them, the hierarchical discriminative network includes a local discriminative network and a global discriminative network. The specific process is as follows:

[0035] According to the user's target comment on the item, through the local discriminative network, optimize the user's summary comment vector and the item-related summary comment vector to obtain the user's optimized summary comment vector and the item-related optimized summary comment vector;

[0036] According to the user's optimized summary comment vector and the item-related optimized summary comment vector, through the global discriminative network, optimize the user's preference vector and item embedding vector to obtain the user's optimized preference vector and item optimized vector.

[0037] As an implementation, according to the user's optimized summary comment vector and the item-related optimized summary comment vector, through the global discriminative network, optimize the user's preference vector and item embedding vector. The specific process is as follows:

[0038] According to the user's optimized summary comment vector and the item-related optimized summary comment vector, optimize the activation signal of the user's historical comments and the activation signal of the item-related historical comments to obtain the optimized user activation signal and the item-related activation signal;

[0039] According to the optimized user activation signal and the item-related activation signal, through the global gating function, obtain the user's optimized preference vector and item optimized vector.

[0040] As an implementation, input the user's optimized preference vector and item optimized vector into a multi-layer perceptron to obtain the final predicted score. The formula is:

[0041]

[0042] Among them, represents the final predicted score, Represents the optimized preference vector of the user; Represents the project optimization vector, Represents a multi-layer perceptron MLP with two hidden layers and one prediction layer.

[0043] The second aspect of the present invention provides a recommendation system score prediction system based on a hierarchical gating and discriminant network, including:

[0044] A data acquisition module for acquiring recommendation system information, where the recommendation system information includes user information, item information, the user's historical reviews, item-related historical reviews, the user's historical scores for items, and the user's target reviews for items;

[0045] A model training module for constructing a score prediction model using a hierarchical gating network and a hierarchical discriminant network, and training the score prediction model based on the acquired recommendation system information;

[0046] Among them, the score prediction model includes a hierarchical gating network, a hierarchical discriminant network, and a multi-layer perceptron;

[0047] During the training process, input the recommendation system information into the hierarchical gating network to obtain the user's preference vector and the item embedding vector;

[0048] Through the hierarchical discriminant network, generate self-supervised signals at the local and global levels to optimize the user's preference vector and the item embedding vector, and obtain the user's optimized preference vector and the project optimization vector;

[0049] Input the user's optimized preference vector and the project optimization vector into the multi-layer perceptron to obtain the final predicted score;

[0050] A prediction score module for using the trained score prediction model to complete the final score prediction of the user for the item.

[0051] The third aspect of the present invention provides a computer device, including a memory, a processor, and a program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in the method described in the first aspect of the present invention.

[0052] The fourth aspect of the present invention aims to provide a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements the steps in the method described in the first aspect of the present invention.

[0053] The above one or more technical solutions have the following beneficial effects:

[0054] In this embodiment, a hierarchical gating network is adopted, which consists of a local gating module and a global gating module. The local gating module selectively identifies informative comments by using personalized trainable thresholds, effectively filtering out noisy comments; the global gating module fully considers the dynamic changes in the correlation between the user's future behavior and historical behavior by evaluating the overall utility of text signals when integrating comment representations, and improves the prediction accuracy.

[0055] In this embodiment, a hierarchical discriminant network is adopted to provide signals to guide the learning process of the hierarchical gating network, generating supervision signals at both the local and global levels to guide the learning of the gating function, solving the problem that it is difficult for existing comment-based recommendation system rating prediction technologies to effectively evaluate the usefulness of comments, ensuring accurate comment selection, and thus achieving better rating prediction.

[0056] Advantages of additional aspects of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The accompanying drawings forming a part of this specification are used to provide a further understanding of the present invention. The schematic embodiments and descriptions thereof of the present invention are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0058] Figure 1 It is a schematic diagram of the overall framework of a recommendation system rating prediction method based on a hierarchical gating and discriminant network for Embodiment 1;

[0059] Figure 2 It is a comparative analysis of MRR and MRR-W of HGD and ATT in Embodiment 1 on different data sets;

[0060] Figure 3 It is the RR distribution of each data set in the test set of Embodiment 1;

[0061] Figure 4 For the user-item pair (1063, 56) in Embodiment 1, the visualization includes and ;

[0062] Figure 5 For the user-item pair (15961, 10371) in Embodiment 1, the visualization includes and ;

[0063] Figure 6 For the user-item pair (23018, 3297) in Embodiment 1, the visualization includes and ;

[0064] Figure 7 For the user - item pair (19, 16431) in the first embodiment, the visualization includes and ;

[0065] Figure 8 For the user - item pair (11172, 13982) in the first embodiment, the visualization includes and . Detailed implementation manners

[0066] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0067] It should be noted that the terms used herein are only for describing the specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention.

[0068] Without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0069] First embodiment

[0070] This embodiment discloses a method for predicting the score of a recommendation system based on a hierarchical gating and discriminant network.

[0071] To more clearly illustrate this embodiment, the implementation process of the method for predicting the score of a recommendation system based on a hierarchical gating and discriminant network can be specifically described as follows:

[0072] A method for predicting the score of a recommendation system based on a hierarchical gating and discriminant network, comprising:

[0073] S1. Obtain the recommendation system information, where the recommendation system information includes user information, item information, the user's historical reviews, item - related historical reviews, the user's historical scores for items, and the user's target reviews for items;

[0074] S2. Use a hierarchical gating network and a hierarchical discriminant network to construct a score prediction model, and train the score prediction model based on the obtained recommendation system information;

[0075] Among them, the score prediction model includes a hierarchical gating network, a hierarchical discriminant network, and a multi - layer perceptron;

[0076] During the training process, input the recommendation system information into the hierarchical gating network to obtain the user's preference vector and the item embedding vector;

[0077] Generate self-supervised signals at the local and global levels through a hierarchical discrimination network, optimize the user's preference vector and item embedding vector, and obtain the optimized preference vector and item optimized vector of the user;

[0078] Input the optimized preference vector and item optimized vector of the user into a multi-layer perceptron to obtain the final predicted score;

[0079] S3. Use the trained score prediction model to obtain the final predicted score of the user for the item, that is, complete the score prediction of the recommendation system.

[0080] As Figure 1 shown, in step S1, obtain the recommendation system information, where the recommendation system information includes user information, item information, the user's historical comments, item-related historical comments, the user's historical scores for the item, and the user's target comments on the item.

[0081] In this embodiment, four publicly accessible datasets are used, and these datasets come from different fields of Amazon 5-core: "Musical Instruments", "Grocery and Gourmet Food", "Video Games", and "Sports and Outdoors". "5-core" means that each user / item pair has at least five interaction records, so there are at least five comments corresponding to it.

[0082] Each dataset is divided into a training set, a validation set, and a test set according to the ratio of 8:1:1. To reduce the impact of division bias, the dataset division and experiments are repeated 10 times, and the results are the average values of these 10 experiments.

[0083] Table 1. Statistical information of the four datasets

[0084]

[0085] S1-1. User information, item information, the user's historical scores for the item, and the user's target comments on the item.

[0086] In the standard task of predicting scores, from the four major datasets, given a user set and an item set, it is expressed as:

[0087]

[0088] (1)

[0089] Where and respectively represent the i-th user and the j-th item, where M and N respectively represent the total number of users and items in the dataset, and 。

[0090] Rating matrix organizes explicit rating records, and each entry represents the user 's rating of the item , indicating the user's preference for that specific item.

[0091] Target comment , and the goal is to predict a real-valued score within a specified range (e.g., 1, 2, 3, 4, 5). The rating is formally associated with the corresponding comment made by the user on the item to express their evaluation of the item.

[0092] S1-2, Obtain the user's historical comments and item-related historical comments.

[0093] In this embodiment, the user's historical comments are represented as a set:

[0094] (2)

[0095] The comments related to the item are represented as a set:

[0096] (3)

[0097] where, represents the number of the user's historical comments, represents the number of comments related to the item 。

[0098] The main goal of the prediction task is, given the user and the item , to generate an estimated rating using past ratings and comments, representing the user's expected evaluation of the item.

[0099] In step S2, a rating prediction model is constructed using a hierarchical gated network and a hierarchical discriminative network, and the rating prediction model is trained based on the obtained user information, item information, corresponding historical comments, and ratings.

[0100] S2-1, Construct a rating prediction model.

[0101] In this embodiment, a scoring prediction model is constructed using a hierarchical gating, hierarchical discriminant network, and multi-layer perceptron. Among them, the hierarchical gating network includes a local gating module and a global gating module. The discriminant network includes a local discriminant network and a global discriminant network.

[0102] The hyperparameter configuration of the evaluation prediction model is shown in Table 2. The trade-off hyperparameters and are systematically adjusted within the range [0, 1] to study the impact of the local and global discriminant learning modules. We train the model using mini-batch stochastic order, set the batch size to 16, and complete a total of 10 training epochs. In the smaller "MusicalInstruments" dataset, the batch size and the number of epochs are set to 16 and 10 respectively, while for the larger dataset, these parameters are set to 128 and 5. The Adam optimizer is used during the optimization process, and the initial learning rate is set to . The embedding dimension and the latent dimension are set to and respectively. The user-item interaction is modeled by a 2-layer multi-layer perceptron (MLP) with hidden layer sizes of 64 and 32 respectively. The model is developed using TensorFlow2, and we use an NVIDIA TITAN graphics card with 24GB of video memory during the experiment. The relevant source code has been made public to support further research.

[0103] Table 2 Hyperparameter Configuration

[0104]

[0105] Among them, the model performance is evaluated using these values. For the smaller "Musical Instruments" dataset, the number of epochs is set to 10, while for the larger dataset it is set to 5. For the "MusicalInstruments" dataset, the specified smaller batch size is 16, while the larger dataset uses a batch size of 128. These values represent the hidden dimensions of the initial layer and subsequent layers of the multi-layer perceptron (MLP).

[0106] Take the observed ratings and historical reviews as inputs. For each user-item pair, the historical reviews are converted into review vectors. The model is built around two basic components. The first component optimizes these review vectors by adopting a local gating module and a global gating module. These mechanisms filter out noisy reviews and extract relevant text information to improve the accuracy of rating prediction. The local gating module identifies the most relevant reviews in a differentiable manner, while the global gating module evaluates the validity of the review data and integrates it into the user-item interaction modeling. In the second component, the supervision signals from the reviews are utilized to develop a discriminative hierarchical gating module. Specifically, in the local gating module, a discriminative local gating threshold is derived by maximizing the margin between the target review and randomly sampled reviews. In contrast, in the global gating module, an anchor-guided contrastive learning process is introduced to optimize the global gating, enabling it to effectively utilize the overall utility of the review features and thus enhancing the prediction task. Due to the hierarchical gating mechanism and discriminative network, compared with existing work, the proposed model can more effectively discover and utilize relevant text information, thereby improving the prediction accuracy.

[0107] S2-2. During the training process, input the recommendation system information into the hierarchical gating network to obtain the user's preference vector and item embedding vector.

[0108] Compared with the traditional soft attention mechanism, the "hard" gating in this module can completely block the influence of noisy reviews.

[0109] In the local gating module, it focuses on the interaction between a specific user and an item ( , ), checks the user's historical reviews and item-related historical reviews , identifies and prioritizes the most relevant reviews, thereby enhancing the user / item representation.

[0110] S2-2-1. Input the propulsion system information into the local gating module to obtain the user's summary review vector and item-related summary review vector respectively.

[0111] (1) Convert each review in the historical reviews into a user's review vector and an item-related review vector respectively.

[0112] In this embodiment, BERT-whitening is used to convert each user's historical review and item-related review into a fixed-size vector and , representing the user's review vector and item-related review vector respectively.

[0113] During the training phase, the comment vector remains static, and this strategy has proven to be effective in saving time and memory resources while maintaining model performance.

[0114] (2) Calculate the correlation between the target item and the user's comment.

[0115] Target item and the user The correlation between comments is calculated by the formula:

[0116] (4)

[0117] Among them, and respectively represent the embedding vectors of the user and the item , , , , are trainable parameters, represents the correlation score of the user's comment.

[0118] (3) Based on the correlation, filter out irrelevant comment information through a gating function.

[0119] In this embodiment, a user-specific threshold is introduced, which is learnable and ensures that any comment with a correlation lower than this value is excluded from the user representation learning process. Filter out irrelevant comment information through the gating function, and the gating function formula is:

[0120] (5)

[0121] Among them, as an extensible hyperparameter, the function as a gating mechanism, when is large enough and , is close to 1, and is 0 under other conditions.

[0122] (4) Based on the gating function, obtain the user's summary comment vector and the item-related summary comment vector.

[0123] By using this gating function, normalize the correlation score to obtain the normalized correlation score; based on the correlation score, obtain the user's summary comment vector.

[0124] In this embodiment, according to the gating function, normalize the correlation score through the softmax function to obtain the normalized correlation score.

[0125] Normalize the relevant scores through the softmax function to obtain the normalized correlation scores , and then use them to summarize the comment vectors to obtain the user's summarized comment vector. The formulas are as follows:

[0126]

[0127] (6)

[0128] where is the correlation weight based on gated filtering normalization, represents the user's summarized comment vector.

[0129] In formula (6), the user's summarized comment vector is mainly composed of comments with correlation scores exceeding the threshold.

[0130] Similarly, the summarized comment vector of item can be obtained through the following formula:

[0131] ;

[0132] ;

[0133] ;

[0134] (7)

[0135] where h 2 , W 3 , W 4 and b 2 represent model parameters, is the trainable parameter related to user . βt represents the item-comment correlation score, which is used to measure the importance of item comments for the target prediction task. represents the comment weight based on gated filtering normalization, which is used to generate the summarized representation of the item. represents the gating function, which is used to filter out noisy comments. represents the item-related summarized comment vector.

[0136] The gating function has two key advantages. First, it is personalized and can be adjusted according to each user's data to achieve better performance. Second, this design ensures that the downstream parameters (such as W 1 , W 2 ) are differentiable with respect to the objective function. This enables the model to support end-to-end learning and simplifies the training process.

[0137] After the above steps, the hierarchical gating network can accurately select comments in an end-to-end manner, effectively filtering out noisy comments and improving prediction accuracy. For traditional flat methods, the concatenation of all historical comments may easily introduce text noise, while the proposed hierarchical mechanism carefully utilizes useful information at different granularity levels, thus achieving better prediction results.

[0138] S2-2-2. Input the user information, item information, the summary comment vector of the user, and the summary comment vector related to the item into the global gating module to obtain the preference vector of the user and the item embedding vector respectively.

[0139] The local-level module focuses on the relevance of each comment, while the global-level module evaluates the overall value of text features for the prediction task.

[0140] Existing research has adopted manually designed features, such as information entropy, to evaluate the overall information content of input instances; however, these intuitive methods have limitations in capturing complex patterns, which may be crucial for measuring the usefulness of comment features.

[0141] (1) In the global gating module, apply a fully connected layer to the summary comment vector of the user and the item embedding vector to obtain the activation signal of the user's historical comments and the activation signal of the item-related historical comments respectively.

[0142] In the global gating module, apply a fully connected layer to the summary comment vector of the user to obtain the accumulation of the usefulness activation signal from historical comments. The formula is:

[0143] (8)

[0144] where and represent parameters, is the sigmoid activation function, , represents the accumulation of the usefulness activation signal obtained from historical comments.

[0145] After the above steps, the overall usefulness of comment features is modeled and effectively evaluated, providing a prerequisite for regulating the overall flow of comment information related to the prediction task.

[0146] (2) According to the activation signal of the user's historical comments and the activation signal of the item-related historical comments, obtain the preference vector of the user and the item embedding vector through the global gating function.

[0147] In this embodiment, according to the accumulation , to adjust the overall flow of comment information related to the prediction task and obtain the user's preference vector, the formula is:

[0148] (9)

[0149] where, ∘ represents element-wise multiplication, represents the user 's preference vector.

[0150] In formula (9), acts as a global gating mechanism. The basic principle of the global gating module is to explore the global comment features , and selectively pick the information dimensions that contribute to forming the user's preferences.

[0151] (3) Similarly, according to step S2-2-2, the global gating module of item is expressed as:

[0152]

[0153] (10)

[0154] where, and are trainable parameters, represents the item embedding vector.

[0155] In this embodiment, once the learned user's preference vector and the item embedding vector are obtained, they will be used as inputs into a multi-layer perceptron to estimate the predicted score .

[0156] S2-3. Through a hierarchical discriminative network, self-supervised signals are generated at the local and global levels to optimize the user's preference vector and the item embedding vector, and obtain the user's optimized preference vector and the item optimized vector.

[0157] The local and global gating modules aim to utilize valuable comments to enhance the representations of users and items, thereby improving the prediction performance. However, due to the lack of supervision signals, the performance of the model may become unstable. Specifically, in the absence of supervision labels indicating the relevance of comments, the local gating module may have difficulty distinguishing relevant comments from irrelevant comments. Due to the lack of supervision signals regarding the overall effectiveness of comments, it also becomes challenging to evaluate the effectiveness of text features in the global gating module. To address this issue, a contrastive learning paradigm is adopted, and different local and global gating mechanisms are developed using self-supervised signals.

[0158] S2-3-1. Optimize the user's summary comment vector and the project-related summary comment vector through the local discriminant network according to the user's target comment on the project, and obtain the optimized summary comment vector of the user and the optimized summary comment vector related to the project.

[0159] In this embodiment, the core concept of local contrast learning involves extracting a guiding signal from the target comment where the encoded sentiment has a strong correlation with the target score . Although it cannot be obtained during the test phase , it can be used during the training process to learn a unique local gating function. Specifically, the target comment should have a higher correlation score compared to randomly sampled comments.

[0160] During the training phase, the target comment is used as a positive sample, and its correlation score is calculated through formula (4) and compared with the score of the randomly sampled noisy comment to generate the local contrast loss formula (11) and optimize the user-side threshold ; the optimized threshold is used in the local gating function formula (5) to filter the noise in the historical comments, and finally the optimized summary comment vector of the user is generated through formula (6). The specific formulas are as follows:

[0161] ;

[0162] ;

[0163] (11)

[0164] where, represents the importance score of the target comment of user for predicting its score, which is used to measure the predictive value of the target comment for the current user-project pair, represents the pseudo-correlation score of the irrelevant comment randomly sampled from the training set, represents the local contrast loss function on the user side, , , , , are trainable parameters, is the user-specific threshold, is the sigmoid activation function, , specifically referring to a comment randomly selected from the training set by the sigmoid activation function.

[0165] Among them, the local contrast loss function on the user side optimizes the user-specific threshold by maximizing the correlation difference between the target review and the noise review. ; The loss value The smaller it is, the higher the discrimination degree of the target review, and the better the model optimization effect.

[0166] During the training phase, the target review generates the local contrast loss on the item side through formula (12) to optimize the item-specific threshold. ; The summary review vector related to the item is calculated by formula (7), and its weight is generated by normalizing the filtered reviews.

[0167] Similarly, according to formula (11), the objective function of the target review is obtained, and according to the objective function, the discriminative item correlation threshold is determined. The formula is:

[0168] ;

[0169] ;

[0170] (12)

[0171] Among them, represents the importance score of the target review of the item for predicting the user score, represents the pseudo-correlation score of the irrelevant item review randomly sampled from the training set, represents the local contrast loss function on the item side, h 2 、W 3 、W 4 and b 2 represent model parameters.

[0172] The local contrast loss function formula (12) on the item side optimizes the item-specific threshold by maximizing the correlation difference between the target review and the noise review. ; The smaller the loss value, the higher the discrimination degree of the target review, and the better the model optimization effect.

[0173] Although the target review cannot be obtained in the test phase, using it during the training process can provide discriminative signals. This method aims to propagate self-supervised signals through the target review to identify reviews that are informative for the prediction task. Since for a user-item pair, the target review has a strong indication for the final prediction, it should be ranked highest among the historical reviews.

[0174] S2-3-2. Optimize the user's preference vector and item embedding vector through the global discriminant network based on the user's optimized summary comment vector and the project-related optimized summary comment vector, and obtain the user's optimized preference vector and item optimization vector.

[0175] In the global gating module, an anchor-guided contrastive learning process is introduced to optimize the global gating, enabling it to effectively utilize the overall utility of comment features, thereby enhancing the prediction task.

[0176] To develop a discriminative global gating module, we introduce an anchor-guided contrastive learning method.

[0177] In this embodiment, use to represent the global gating function defined in formulas (8) and (9), and use to represent the global gating function described in formula (10), and obtain the user's optimized preference vector and item optimization vector.

[0178] In the global gating module, a global discriminant mechanism is adopted to generate self-supervised signals through contrastive learning, guiding the global gating module to select more informative comment features and avoid noise interference.

[0179] S2-3-3. Input the user's optimized preference vector and item optimization vector into a multi-layer perceptron to obtain the final predicted score.

[0180] In this embodiment, once the user's optimized preference vector and item optimization vector are learned, the user's optimized preference vector and item optimization vector are input into a multi-layer perceptron as inputs to estimate the predicted score , and the formula is:

[0181] (13)

[0182] where represents the final predicted score, represents the user's optimized preference vector; represents the item optimization vector, represents a multi-layer perceptron MLP with two hidden layers and one prediction layer.

[0183] At the same time, let represent the negative global comment vector obtained by setting the local gating value to 1. The global contrast loss function is expressed as follows:

[0184]

[0185]

[0186]

[0187] (14)

[0188] Among them, represents the margin value, and the positive and negative sample pairs are distinguished by the margin value ; is a global contrast loss, where for each user-item pair, is the anchor, is the positive sample, is the negative sample. The optimization method of the global gating module is to make the score corresponding to the anchor close to the score of the positive sample and far from the score of the negative sample.

[0189] S2-3-4. Optimize and learn the scoring model through the objective function, and the objective function is defined by three components: scoring prediction loss, local contrast loss, and global contrast loss.

[0190] For the scoring prediction loss, the mean squared error (MSE) between the predicted score and the true score is used to establish the learning objective, and the formula is:

[0191] (15)

[0192] Among them, represents the user 's actual score for the item , represents the predicted score of the model for the same user-item pair.

[0193] After combining the local and global contrast loss components, the target loss function of the model is specifically defined as:

[0194] (16)

[0195] Among them, D represents the training dataset, including the tuples composed of the user , the item , their corresponding actual scores and the possible distance or similarity metric ; The hyperparameters and are used to balance between the scoring prediction loss and the local / global contrast loss; represents the evaluation prediction loss, represents the local contrast loss on the user side, represents the local contrast loss on the item side, represents the global contrast loss.

[0196] Optimize the scoring model through the target loss to make the model's scoring prediction more accurate.

[0197] In step S3, use the trained scoring prediction model to obtain the user's final predicted score for the item, that is, complete the scoring prediction of the recommendation system.

[0198] In this embodiment, when predicting the user's score for the target commodity, the input data includes user information (such as age, gender, historical preferences), item information (such as commodity category, brand), and historical comments (such as the user's evaluation of similar commodities "The sound quality is very good"). The hierarchical gated network calculates the correlation score between the comment and the target commodity through the local gated module (for example, the score of the "sound quality" related comment is 0.85), and uses the personalized threshold to filter out low-correlation comments (such as the score of "slow delivery" 0.15 is excluded); the global gated module fuses the filtered comment vector with the user / item embedding to generate a preference representation. Finally, the multi-layer perceptron (MLP) outputs the predicted score based on the optimized vector (such as 4.3 points). Even if the target comment is missing, the model can still accurately infer the user's preference through historical data.

[0199] In this embodiment, evaluate the model and compare it with the existing models.

[0200] 1. Use the mean absolute error (MAE) as the evaluation metric for evaluating the existing model.

[0201] MAE is a commonly used scoring prediction metric in recommendation systems. To reduce random bias, each experiment is repeated 10 times, and the average results of these 10 runs are reported for effective model comparison. It is worth mentioning that the performance variation between different runs is very small.

[0202] Among them, the existing models are:

[0203] PMF: A traditional matrix factorization method that predicts scores through the dot product of user and item vectors.

[0204] NeuMF: Combines generalized matrix factorization and multi-layer neural networks for score estimation.

[0205] DeepCoNN: Uses comment-based user and item representations to complete the prediction task through factorization machines.

[0206] NARRE: Introduces an attention mechanism based on DeepCoNN to more effectively utilize comment information.

[0207] DAML: Proposes to simultaneously learn comment features, combining the unique and shared attributes of convolutional neural networks (CNNs).

[0208] GC-MC: Using graph convolution techniques, it focuses on the relationship between users and items, and takes the connection prediction between users and items in the bipartite graph as the recommendation task.

[0209] RMG: Combines graph data and review data to predict ratings.

[0210] RGCL: Uses graph contrastive learning methods to improve the understanding of graph architectures and address challenges related to user behavior.

[0211] These baseline models are selected for representative comparison. For example: The PMF and NeuMF models achieve recommendations through user-item interactions. DeepCoNN uses review semantics to solve the cold start problem and improve model performance. NARRE and DAML introduce attention mechanisms to model text relevance and learn comprehensive user representations. Finally, GC-MC, RMG, and RGCL propose advanced graph learning methods to model the higher-order connections between users, items, and reviews, enabling more accurate predictions.

[0212] 2. Overall Comparison

[0213] Table 3 summarizes the comparison results for four datasets. The second-best value of MAE is underlined. Δ represents the performance improvement of HGD compared to the leading baseline model, where a performance improvement of more than 1% is considered significant.

[0214] Table 3 Model Comparison on Different Datasets Based on MAE

[0215]

[0216] Based on these data, several key observations can be drawn.

[0217] (1) Obviously, PMF consistently performs the worst on all datasets. This poor performance is mainly attributed to its reliance on the user-item inner product to model rating behavior, which makes it particularly vulnerable to the inherent data sparsity in recommendation problems. In contrast, NeuMF has a significant improvement compared to PMF in terms of the MAE metric. This improvement benefits from its adoption of a multi-layer neural network, which more efficiently captures and models complex user-item interactions.

[0218] (2)The model that incorporates reviews into the framework (Table 3, lines 3 - 5) always outperforms the latent factor - based models (such as PMF and NeuMF). This indicates the efficiency of review features in modeling user behavior and item attributes, demonstrating the important role of these features in improving prediction accuracy. NARRE has a better MAE performance than DeepCoNN on all datasets. This finding emphasizes the advantage of analyzing text information at different granularities. Although DeepCoNN aggregates all past reviews into a comprehensive document, this approach introduces text - related noise and irrelevant information. NARRE uses an attention mechanism to identify and select the most informative reviews, enabling more accurate predictions. Compared with other review - based models, DAML is proven to be very effective. It emphasizes the ability to successfully detect pairwise connections between different user - item pairs, and this approach dynamically reveals the specific complex interactions for each user - item pair, providing a more accurate basis for rating prediction.

[0219] (3)The graph - based models (Table 3, lines 6 - 8) always significantly outperform the review - based models. This can be attributed to the ability of graph learning itself to discover high - order connections in the user - item bipartite graph. This ability can generate more comprehensive representations, thus enhancing the prediction accuracy. Among the competing models, RGCL has the highest performance ranking, which emphasizes the advantage of enriching the user - item graph with review attributes and using contrastive learning for self - supervision. This analysis provides valuable insights into understanding the relative performance of various models on the dataset.

[0220] (4)The proposed HGD model demonstrates superior performance to the baseline models on all four datasets. Specifically, compared with the best-performing baseline model RGCL, the HGD model shows a relative improvement of 1.1% in MAE on the "Musical Instruments" dataset, 2.6% on the "Grocery and Gourmet Food" dataset, 3.9% on the "Video Games" dataset, and 3.8% on the "Sports and Outdoors" dataset. Compared with interaction-based models such as PMF and NeuMF, the proposed model utilizes historical reviews for comprehensive user / item modeling, thus enhancing the prediction accuracy. Different from review-based methods such as DeepCoNN, which concatenates all reviews indiscriminately into a single document, the proposed model can leverage the relevance of reviews to achieve more accurate predictions. For attention-based methods such as DAML, HGD suppresses review noise through a hierarchical gating network that identifies the most relevant reviews in a "hard" manner via personalized thresholds. Finally, compared with graph-based methods such as RGCL, the advantage of HGD lies in its ability to provide supervision signals, thereby enhancing the discovery of useful reviews. Instead of learning high-order general representations from interaction data, HGD guides the exploration of historical reviews through a hierarchical discriminative network to support target-specific predictions.

[0221] In summary, there are two main reasons for the performance improvement of the HGD model. First, the hierarchical gating module plays a crucial role in reducing irrelevant noise in reviews and identifying the most relevant reviews. This gating mechanism effectively addresses the noise problem in reviews, enabling the model to fully utilize relevant reviews to complete the prediction task. Second, the hierarchical discriminative network we introduced makes a significant contribution to the success of the model. This network transfers knowledge from the target reviews, delivering self-supervised signals at different granularity levels, thus helping to form a local discriminative gating function and guiding the HGD model to focus on relevant text signals.

[0222] 3. Performance of Each Sub-module in the HGD Model

[0223] For this purpose, HGD is compared with the following variants:

[0224] HGD-local: This is a version of HGD that does not include the local discriminative module. Therefore, there is no supervision signal during the learning of the discriminative local gating module, and it is treated as a traditional attention mechanism.

[0225] HGD-global: This is the HGD version without the global discriminative module. Therefore, there is no supervised learning mechanism for the discriminative global gating module, and it is regarded as a traditional gating network.

[0226] Table 4. Comparison of different model variants

[0227]

[0228] Table 4 shows the comparison results, and the following conclusions can be drawn:

[0229] (1) The performance of HGD is better than all variants, which highlights the effectiveness of its sub-modules in filtering noise and utilizing the supervision signal.

[0230] (2) The performance degradation of HGD-local is more significant than that of HGD-global, indicating that it is very necessary to denoise at the local review level.

[0231] (3) The comparison between HGD and its variants further confirms the effect of extracting insights from the target reviews to generate supervision signals and establishing a cross-level discriminative gating mechanism. It can be inferred that without the supervision signal indicating the review relevance, the hierarchical gating network may not be able to extract the review features informative for the prediction task, which may lead to poor learning of the relevant parameters.

[0232] To provide a more comprehensive understanding, the effectiveness of HGD in selecting relevant reviews is compared with the traditional attention mechanism (ATT). ATT is a traditional attention mechanism that does not use self-supervised signals when selecting reviews. In this comparison, the performance is evaluated by two metrics: the mean reciprocal rank (MRR) and its weighted version MRR-W. MRR represents the average ratio of the number of selected relevant reviews to the number of relevant reviews whose scores are higher than the threshold. MRR-W combines the scoring weights of relevance on the basis of MRR. Higher MRR and MRR-W scores indicate a larger proportion of selected relevant reviews. The following are the definitions of MRR and MRR-W:

[0233] ,

[0234] (17)

[0235] where represents the test set represents the historical reviews is in the same review category as the target review

[0236] ​As shown in Figure 2, the results clearly show that in terms of both MRR and MRR-W, HGD is always superior to ATT. This indicates that compared with the traditional attention mechanism, HGD is more effective in selecting relevant reviews, further highlighting the advantage of HGD in using historical reviews to provide accurate rating predictions.

[0237] The distribution of RR (Relevant Ratio) on each dataset in the test set is also shown. Figure 3 shows the results, where the x-axis represents RR and the y-axis represents the number of records in the test set that reach the corresponding RR. From the chart, the following observations can be made:

[0238] (1) The number of records of HGD at higher RR is more than that of ATT, while the number of records at lower RR is less than that of ATT, which confirms the effectiveness of HGD in discovering more relevant reviews to improve rating predictions.

[0239] (2) The improvement of RR on the "Musical Instrument" dataset is small, which also explains why the improvement of HGD compared with HGD-local that also uses the traditional attention mechanism to select reviews is not obvious.

[0240] (3) Compared with ATT, HGD has more records in the RR interval [0, 0.2]. A possible explanation is that HGD adopts a relatively "strict" way to select reviews, which may accidentally discard some actually informative reviews.

[0241] 4. Case Analysis

[0242] To more deeply understand the effectiveness of HGD in identifying relevant reviews for prediction, we show some visualization examples that illustrate the relevance scores of historical reviews during the rating estimation process. To achieve this goal, user-item entries in the test subset of the "Sports and Outdoors" dataset were randomly selected. Then, for each user and their historical reviews, charts of (see formula (4)) and (see formula (5)) were plotted. As Figures 4 to 8 shown, these visualizations show three user-item pairs. In these figures, the x-axis represents historical reviews, and the y-axis represents the corresponding values. In addition, the personalized threshold is included to assist in the analysis. In the illustration, "gray" bars represent reviews related to the target review, meaning that their sentiment is consistent with the target review; on the contrary, "white" bars are used to represent irrelevant reviews.

[0243] As can be seen from the figure, compared with irrelevant comments, relevant comments usually obtain higher relevance scores. In addition, relevant comments are usually prioritized above the threshold, while irrelevant comments are usually below the threshold. These findings indicate that the introduced discriminative learning module can effectively utilize the supervision signal of the target comments and successfully distinguish relevant comments from irrelevant comments. The proposed HGD model demonstrates a powerful ability to enhance prediction accuracy by effectively utilizing relevant comments. This ability benefits from the adoption of a hierarchical gating module, which can automatically integrate informative semantics while suppressing irrelevant noise.

[0244] In this embodiment, a hierarchical gating and discriminative framework is constructed for score prediction. It solves many challenges existing in traditional recommendation systems, such as text noise, lack of supervision information, and the flat structure of comment features. A hierarchical gating network is adopted, which is specifically used to extract features from comments. The task of the local gating module is to selectively identify comments that are useful and relevant for the target prediction; while the global gating module evaluates the effectiveness of the overall comment signal. To guide the learning process of the hierarchical gating network, a discriminative hierarchical network is constructed, which utilizes the sentiment alignment relationship between the target comment and its corresponding score. This network generates self-supervision signals to help learn discriminative gating at the local and global levels. The hierarchical gating and discriminative networks work together, enabling the model to extract valuable comment features, thereby improving prediction accuracy. Through detailed experiments on a public dataset, the superior effect of the HGD method is verified, showing that it outperforms the existing state-of-the-art benchmarks.

[0245] Embodiment 2

[0246] The purpose of this embodiment is to provide a recommendation system score prediction system based on a hierarchical gating and discriminative network, including:

[0247] A data acquisition module, configured to acquire recommendation system information, where the recommendation system information includes user information, item information, the user's historical comments, item-related historical comments, the user's historical scores for items, and the user's target comments on items;

[0248] A model training module, configured to construct a score prediction model using a hierarchical gating network and a hierarchical discriminative network, and train the score prediction model based on the acquired recommendation system information;

[0249] Wherein, the score prediction model includes a hierarchical gating network, a hierarchical discriminative network, and a multi-layer perceptron;

[0250] During the training process, the recommendation system information is input into the hierarchical gating network to obtain the user's preference vector and the item embedding vector;

[0251] Generate self-supervised signals at both local and global levels through a hierarchical discrimination network, optimize the user's preference vector and item embedding vector, and obtain the optimized preference vector and item optimized vector of the user.

[0252] Input the optimized preference vector and item optimized vector of the user into a multi-layer perceptron to obtain the final predicted score.

[0253] The prediction score module is used to complete the final score prediction of the user for the item by using the trained score prediction model.

[0254] Based on a recommendation system score prediction system based on hierarchical gating and discrimination network, implement the method steps in Embodiment 1.

[0255] Embodiment 3

[0256] The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the above method are implemented.

[0257] Embodiment 4

[0258] The purpose of this embodiment is to provide a computer-readable storage medium.

[0259] A computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps of the above method are executed.

[0260] Embodiment 5

[0261] The purpose of this embodiment is to provide a computer program product containing instructions. When it runs on a computer, it enables the computer to execute the methods and functions involved in any one of the above embodiments.

[0262] The steps involved in the devices of the above embodiments correspond to those of Method Embodiment 1. For specific implementation manners, reference may be made to the relevant description part of Embodiment 1. The term "computer-readable storage medium" should be understood to include a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0263] Those skilled in the art should understand that the above-mentioned modules or steps of the present invention can be implemented by a general-purpose computer device. Optionally, they can be implemented by program codes executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0264] Although the specific implementation manners of the present invention have been described above in conjunction with the accompanying drawings, it is not a limitation to the protection scope of the present invention. Those skilled in the art should understand that, based on the technical solutions of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A recommendation system rating prediction method based on hierarchical gating and discriminant network, characterized in that: include: Acquire recommendation system information, wherein the recommendation system information includes user information, project information, user's historical comments, project-related historical comments, user's historical ratings of the project, and user's target comments on the project; A rating prediction model is constructed using a hierarchical gating network and a hierarchical discriminant network, and the rating prediction model is trained based on the obtained recommendation system information; Using the trained rating prediction model, we can get the user's final predicted rating for the project, thus completing the rating prediction of the recommendation system. Among them, the score prediction model includes hierarchical gating network, hierarchical discriminant network and multi-layer perceptron; During the training process, the recommendation system information is input into the hierarchical gating network to obtain the user's preference vector and item embedding vector; Through the hierarchical discriminant network, self-supervisory signals are generated at the local and global levels to optimize the user's preference vector and the item embedding vector, and the user's optimized preference vector and the item optimized vector are obtained; The user's optimization preference vector and the item optimization vector are input into the multi-layer perceptron to obtain the final prediction score.

2. A recommendation system rating prediction method based on hierarchical gating and discriminant network as claimed in claim 1, characterized in that: The recommendation system information is input into the hierarchical gating network to obtain the user's preference vector and the item embedding vector. The hierarchical gating network includes a local gating module and a global gating module. The specific process is as follows: Input the promotion system information into the department control module to obtain the user's summary comment vector and the project-related summary comment vector respectively; User information, project information, user summary comment vector and project-related summary comment vector are input into the global gating module to obtain the user preference vector and project embedding vector respectively.

3. A recommendation system rating prediction method based on hierarchical gating and discriminant network as claimed in claim 2, characterized in that: Get the user's summary comment vector and the project-related summary comment vector. The specific process is: Convert each comment in the historical comments into the user's comment vector and the project-related historical comment vector respectively; Calculate the relevance score between the target item and the user's review; Based on the relevance score, the user's summary comment vector and the project-related summary comment vector are obtained through the local gating function.

4. A recommendation system rating prediction method based on hierarchical gating and discriminant network as claimed in claim 2, characterized in that: Get the user's preference vector and item embedding vector. The specific process is: In the global gating module, the user's summary comment vector and the project embedding vector are applied to the fully connected layer to obtain the activation signal of the user's historical comments and the activation signal of the project-related historical comments respectively; According to the activation signals of the user's historical comments and the activation signals of the project-related historical comments, the user's preference vector and the project embedding vector are obtained through the global gating function.

5. The recommendation system rating prediction method based on hierarchical gating and discriminant network as claimed in claim 1, characterized in that: Through the hierarchical discriminant network, self-supervisory signals are generated at the local and global levels to optimize the user's preference vector and the project embedding vector, and the user's optimized preference vector and the project optimization vector are obtained. The hierarchical discriminant network includes a local discriminant network and a global discriminant network. The specific process is as follows: According to the user's target comments on the project, the user's summary comment vector and the project-related summary comment vector are optimized through the local discriminant network to obtain the user's optimized summary comment vector and the project-related optimized summary comment vector; According to the user's optimized summary comment vector and the project-related optimized summary comment vector, the user's preference vector and the project embedding vector are optimized through the global discriminant network to obtain the user's optimized preference vector and the project optimization vector.

6. A recommendation system rating prediction method based on hierarchical gating and discriminant network as claimed in claim 5, characterized in that: According to the optimized summary comment vector of the user and the optimized summary comment vector related to the project, the user's preference vector and the project embedding vector are optimized through the global discriminant network. The specific process is as follows: According to the optimized summary comment vector of the user and the optimized summary comment vector related to the project, the activation signal of the user's historical comments and the activation signal of the project-related historical comments are optimized to obtain the optimized user activation signal and the project-related activation signal; According to the optimized user activation signal and project-related activation signal, the user's optimized preference vector and project optimization vector are obtained through the global gating function.

7. The recommendation system rating prediction method based on hierarchical gating and discriminant network as claimed in claim 1, characterized in that: The user's optimization preference vector and the project optimization vector are input into the multi-layer perceptron to obtain the final prediction score. The formula is: ； in, represents the final prediction score, represents the user's optimized preference vector; represents the project optimization vector, represents a multilayer perceptron MLP with two hidden layers and one prediction layer.

8. A recommendation system rating prediction system based on hierarchical gating and discriminant network, characterized in that: include: The data acquisition module is used to acquire recommendation system information, wherein the recommendation system information includes user information, project information, user's historical comments, project-related historical comments, user's historical ratings of projects, and user's target comments on projects; A model training module is used to construct a rating prediction model using a hierarchical gating network and a hierarchical discriminant network, and train the rating prediction model based on the obtained recommendation system information; Among them, the score prediction model includes hierarchical gating network, hierarchical discriminant network and multi-layer perceptron; During the training process, the recommendation system information is input into the hierarchical gating network to obtain the user's preference vector and item embedding vector; Through the hierarchical discriminant network, self-supervisory signals are generated at the local and global levels to optimize the user's preference vector and the item embedding vector, and the user's optimized preference vector and the item optimized vector are obtained; Input the user's optimization preference vector and the item optimization vector into the multi-layer perceptron to obtain the final prediction score; The prediction and rating module is used to use the trained rating prediction model to complete the user's final rating prediction for the project.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are performed.

Images