Recommendation system score prediction method and system based on hierarchical gating and discrimination network

Abstract

The invention provides a recommendation system score prediction method and system based on hierarchical gating and a discriminant network, relates to the technical field of recommendation systems, and aims to solve the problems that in the prior art, noise influence of irrelevant comments is not considered, complexity and hierarchy between users and projects are not captured, meanwhile, supervision signals are lacked, and a score prediction model is poor in accuracy. The method comprises the steps of obtaining recommendation system information; constructing a score prediction model by using a hierarchical gating network and a hierarchical discrimination network, and training the score prediction model based on the obtained recommendation system information; and the trained score prediction model is used to complete the final score prediction of the user on the project. According to the method, the noise influence problem of irrelevant comments is solved, the complexity and hierarchy between users and items can be captured, the sub-supervision signals are generated through contrast learning, and the score prediction accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of recommendation systems, and in particular to a recommendation system rating prediction method and system based on hierarchical gating and discriminant networks. Background Art

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

[0003] Recommendation system is a technology that screens and recommends potentially related items to users by analyzing user behavior, item attributes and contextual information. Its core mechanisms include collaborative filtering (similarity recommendations based on user-item interactions), content filtering (matching based on item features) and hybrid methods (integration of 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 TikTok video recommendations). The core goal of the recommendation system is to help users efficiently obtain information of interest through personalized recommendations, while improving the platform's service quality and user stickiness.

[0004] Rating prediction is a core subtask of the recommendation system. It aims to predict the user's rating value (usually a discrete value, such as 1-5 points) for uninteracted items based on the user's historical behavior (such as rating, click, purchase) and auxiliary information (such as user portrait, item description, and comment content). Its input is the user-item pair, and the output is the predicted rating. The modeling goal is to minimize the error (such as mean square error) between the predicted value and the actual rating. The accuracy of the rating prediction directly determines the performance of the recommendation system.

[0005] Traditional recommendation systems mainly rely on explicit interactions between users and items (such as ratings, clicks) or implicit feedback (such as browsing time) to model user preferences through collaborative filtering or matrix decomposition 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 the comments (such as "the earphones sound clear and the bass is powerful" reflects the user's sensitivity to sound effects; "long battery life but slow charging" reveals the functional advantages and disadvantages of the product) can supplement the sparse interaction signals. By utilizing the rich semantic information in the comments, more comprehensive user and item representations can be developed, thereby improving the accuracy of recommendations. Therefore, the expressiveness of historical comments in capturing user preferences has attracted widespread attention in the field of recommendation systems, and many recommendation systems based on comments have emerged.

[0007] In the early days of review-based recommender systems, neural networks were used to generate review-based representations for users and items independently, and then modeled user-item interactions for rating prediction. More advanced review-based methods incorporate interaction modeling into the feature extraction stage, thereby capturing the textual semantics of specific user-item pairs and improving performance. For example, DeepCoNN uses a pair of convolutional neural networks to generate embeddings for users and items, and then merges these embeddings to predict the final rating. In addition, DAML focuses on improving the modeling of interactions between users and items by extracting text features in reviews that are both local and mutual. The following methods emphasize identifying unique and informative textual signals for each user-item pair. To address the problem of coarse-grained historical reviews, some methods use 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. They aim to reveal high-order relationships between users and items, thereby providing more refined user-item relationship modeling for recommendation systems. In review-based recommendation systems, user-item relationships are represented as a bipartite graph that connects users and items. To enhance recommendation systems, advanced techniques use graph learning to exploit the inherent graph structure. For example, GC-MC uses relational graph convolution to enhance the embedding of users and items, and regards rating prediction as a link prediction problem in the bipartite graph of users and items. RMG introduces a powerful framework to integrate the data of user-item graphs with review content. The framework uses a hierarchical attention network to extract important features from the hierarchical structure of reviews. Meanwhile, graph neural networks (GNNs) are used to map the interactions between users and items in the graph, aiming to clarify their mutual relationships. Similarly, RGCL proposes a rating prediction framework with contrastive learning as the core, exploiting the graph structure and incorporating review data. In this model, reviews serve as connections in the information propagation process, which helps to fully exploit the unique user-item graph structure. In addition, the model incorporates contrastive learning to generate supervisory signals, which further enhances the graph learning process using review data.

[0009] Although the above research works have made significant progress in leveraging textual and structural features between users, items, and reviews to improve rating prediction, they still face three major challenges. The first challenge is the existence of noisy reviews. Due to the diversity and variability of user behaviors, historical reviews may not always provide valuable signals to predict specific targets. Although some works have proposed attention mechanisms to softly focus on important reviews, irrelevant textual signals may still overwhelm useful signals, thereby weakening the performance of the model. The second challenge lies in the flat approach that treats historical reviews indiscriminately and simply connects them to represent users / items. 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 supervisory signals for selecting informative reviews. Existing works select reviews in an attentive manner by learning from data, but lack supervisory signals indicating review relevance, which may lead to unexplainable review selection and prone to overfitting to training data.

[0010] In summary, the existing technology still has some shortcomings: (1) Irrelevant comments may introduce noise, which in turn has an adverse impact on model performance.

[0011] (2) Existing methods treat review features as static single-granular representations and do not model hierarchical semantic associations, which reduces the usefulness of review information and the accuracy of model evaluation.

[0012] (3) Due to the lack of sparse supervisory signals, existing review-based recommendation system rating prediction techniques are difficult to effectively evaluate the usefulness of reviews. Summary of the invention

[0013] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a recommendation system rating prediction method and system based on hierarchical gating and discriminant network, which combines hierarchical gating with hierarchical discriminant network, filters noisy comments through hierarchical gating, extracts relevant text information, and generates supervision signals at the local and global levels through hierarchical discriminant network to ensure accurate comment selection, thereby improving the accuracy of rating prediction.

[0014] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: A first aspect of the present invention provides a recommendation system rating prediction method based on hierarchical gating and discriminant network, comprising: 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.

[0015] As an implementation method, the recommendation system information is input into a hierarchical gating network to obtain a user preference vector and an item embedding vector, wherein 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.

[0016] As an implementation method, the user's summary comment vector and the project-related summary comment vector are obtained. The specific process is as follows: 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 comments; 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.

[0017] As an implementation method, the user's preference vector and the item embedding vector are obtained. The specific process is as follows: 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.

[0018] As an implementation method, a self-supervisory signal is generated at the local and global levels through a hierarchical discriminant network, and the user's preference vector and the project embedding vector are optimized to obtain the user's optimized preference vector and the project optimization vector. 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.

[0019] As an implementation method, based on 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 a 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.

[0020] As an implementation method, the user's optimization preference vector and the project optimization vector are input into a multi-layer perceptron to obtain the final prediction score, and the formula is:

[0021] 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.

[0022] A second aspect of the present invention provides a recommendation system rating prediction system based on hierarchical gating and discriminant network, comprising: 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.

[0023] A third aspect of the present invention provides a computer device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method described in the first aspect of the present invention are implemented.

[0024] The fourth aspect of the present invention aims to provide a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the method described in the first aspect of the present invention.

[0025] One or more of the above technical solutions have the following beneficial effects: In this embodiment, a hierarchical gating network is adopted, which is composed of a local gating module and a global gating module. The local gating module selectively identifies informative comments by using personalized trainable thresholds, and effectively filters out noisy comments; the global gating module fully considers the dynamic changes of the correlation between the user's future behavior and historical behavior and improves the prediction accuracy by evaluating the overall utility of the text signal when integrating the comment representation.

[0026] In this embodiment, a hierarchical discriminant network is used to provide signals to guide the learning process of the hierarchical gating network, generate supervision signals at both the local and global levels, and guide the learning of the gating function, thereby solving the problem that the existing review-based recommendation system rating prediction technology is difficult to effectively evaluate the usefulness of reviews, ensuring accurate review selection, and thus achieving better rating prediction.

[0027] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0029] Figure 1 Schematic diagram of the overall framework of a recommendation system rating prediction method based on hierarchical gating and discriminant network in the first embodiment of the present invention; Figure 2 This is a comparative analysis of MRR and MRR-W of HGD and ATT on different data sets in Example 1; Figure 3 RR distribution of each data set in the test set of the first embodiment; Figure 4 For the user-item pair (1063, 56) in the first embodiment, the visualization includes and ; Figure 5 For the user-item pair (15961, 10371) in the first embodiment, the visualization includes and ; Figure 6 For the user-item pair (23018, 3297) of the first embodiment, the visualization includes and ; Figure 7 For the user-item pair (19, 16431) in the first embodiment, the visualization includes and ; Figure 8 For the user-item pair (11172, 13982) in the first embodiment, the visualization includes and . DETAILED DESCRIPTION

[0030] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0031] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0032] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0033] Embodiment 1 This embodiment discloses a recommendation system rating prediction method based on hierarchical gating and discriminant network.

[0034] In order to more clearly illustrate this embodiment, the implementation process of the recommendation system rating prediction method based on hierarchical gating and discriminant network can be specifically described as follows: A recommendation system rating prediction method based on hierarchical gating and discriminant network, comprising: S1. Obtain 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; S2. Use the hierarchical gating network and the hierarchical discriminant network to build a rating prediction model, 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; S3. Use the trained rating prediction model to obtain the user's final predicted rating for the project, thus completing the recommendation system rating prediction.

[0035] like Figure 1 As shown, in step S1, recommendation system information is obtained, and 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.

[0036] In this example, four publicly accessible datasets are used, which are from different areas of Amazon 5-core: "Musical Instruments", "Grocery and GourmetFood", "Video Games", and "Sports and Outdoors". "5-core" means that each user / item pair has at least five interaction records, and therefore corresponds to at least five reviews.

[0037] Each dataset is divided into training set, validation set and test set in the ratio of 8:1:1. In order to reduce the impact of the partition bias, the dataset partition and experiment are repeated 10 times, and the results are the average of the 10 experiments.

[0038] Table 1. Statistics of the four datasets

[0039] S1-1, user information, project information, user's historical ratings of the project and user's target comments on the project.

[0040] In the standard task of predicting ratings, given a set of users and a set of items from four major datasets, it is represented as:

[0041] (1) in and denote the i-th user and j-th item respectively, M and N denote the total number of users and items in the dataset respectively, and .

[0042] Rating Matrix Organize explicit rating records, each entry Indicates user About Project The rating indicates the user's preference for that particular item.

[0043] Target Comments , the goal is to predict a real number score within a specified range (e.g., 1, 2, 3, 4, 5). With the user About Project Corresponding comments Formal association used to express their evaluation of the project.

[0044] S1-2. Obtain the user's historical comments and project-related historical comments.

[0045] In this embodiment, the user The historical comments of are represented as a collection: (2) With Project Related comments are represented as a collection: (3) in, Indicates user The number of historical comments, Representation and Project The number of related comments.

[0046] The main goal of the prediction task is to and Projects , using past ratings and reviews to generate estimated ratings , indicating that the user About Project expected evaluation.

[0047] In step S2, 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 acquired user information, project information, corresponding historical comments and ratings.

[0048] S2-1. Construct a rating prediction model.

[0049] In this embodiment, hierarchical gating, hierarchical discriminant network and multi-layer perceptron are used to construct a score prediction model, wherein the hierarchical gating network includes a local gating module and a global gating module, and the discriminant network includes a local discriminant network and a global discriminant network.

[0050] The hyperparameter configurations of the evaluated prediction models are shown in Table 2. Weighing Hyperparameters and is systematically adjusted in the range [0, 1] to study the impact of local and global discriminative learning modules. We train the model using mini-batches in random order, with a batch size of 16 and a total of 10 training epochs. In the smaller MusicalInstruments dataset, the batch size and number of epochs are set to 16 and 10, respectively, while for the larger dataset, these parameters are set to 128 and 5, respectively. The Adam optimizer is used for optimization, and the initial learning rate is set to The embedding dimension and latent dimension are set to and 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 during the experiment we used an NVIDIA TITAN graphics card with 24GB of video memory. The relevant source code has been made public to support further research.

[0051] Table 2 Hyperparameter configuration

[0052] 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 smaller batch size specified is 16, while the larger dataset uses a batch size of 128. These values ​​represent the hidden dimensions of the initial and subsequent layers of the Multilayer Perceptron (MLP).

[0053] The observed ratings and historical reviews are taken as input. 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 textual 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 effectiveness of the review data and integrates it into the user-item interaction modeling. In the second component, the supervisory signals from the reviews are used 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 the randomly sampled reviews. In contrast, in the global gating module, an anchor-guided contrastive learning process is introduced to optimize the global gating, which enables it to effectively exploit the overall utility of the review features to enhance the prediction task. Due to the hierarchical gating mechanism and the discriminative network, the proposed model can more effectively discover and exploit relevant textual information compared to existing works, thereby improving the prediction accuracy.

[0054] S2-2. 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.

[0055] Compared with the traditional soft attention mechanism, the “hard” gating in this module can completely block the influence of noisy comments.

[0056] In the local gating module, the focus is on specific users and projects ( , ) to check the user's historical comments Historical comments related to the project , identifying and prioritizing the most relevant reviews, thus enhancing the representation of users / items.

[0057] S2-2-1. 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.

[0058] (1) Convert each comment in the historical comments into a user comment vector and a project-related comment vector respectively.

[0059] In this embodiment, BERT-whitening is used to transform each user’s historical comments Comments related to the project Convert to a fixed-size vector and , respectively represent the user's comment vector and the item-related comment vector.

[0060] During the training phase, the review vectors are kept static, a strategy that has been shown to be effective in saving time and memory resources while maintaining model performance.

[0061] (2) Calculate the correlation between the target item and the user’s comments.

[0062] Target Project With users The correlation between the comments is calculated as: (4) in, and Respectively represent users and Projects The embedding vector of , , , is a trainable parameter, Represents the relevance score of the user review.

[0063] (3) Based on relevance, filter out irrelevant comment information through the gating function.

[0064] In this embodiment, a user-specific threshold is introduced , this threshold is learnable and ensures that any comments below this relevance are excluded from the user representation learning process. The gating function is used to filter out irrelevant comment information. The gating function formula is: (5) in, As a scalable hyperparameter, the function As a gating mechanism, Large enough and hour, close to 1, and 0 under other conditions.

[0065] (4) Based on the gating function, the user’s summary comment vector and the project-related summary comment vector are obtained.

[0066] By using this gating function, the relevance score is normalized to obtain a normalized relevance score; based on the relevance score, the user's summary comment vector is obtained.

[0067] In this embodiment, according to the gating function, the correlation score is calculated by the softmax function. Normalization is performed to obtain a normalized correlation score.

[0068] The relevant scores are calculated through the softmax function Normalize and get the normalized correlation score , and then use it to summarize the comment vector to get the user's summary comment vector. The formulas are:

[0069] (6) in, Based on the normalized relevance weights of the gated filter, Represents the user's summary comment vector.

[0070] In formula (6), the user's summary comment vector Mainly consists of reviews with relevance scores above a threshold.

[0071] Similarly, projects The summary comment vector can be obtained by the following formula: ; ; ; (7) Among them, h2, W3, W4 and b2 represent model parameters, Is with the user The relevant trainable parameters, βt, represent the item-review relevance score, which is used to measure the importance of item reviews to the target prediction task. represents the review weights normalized based on gated filtering, which is used to generate the summary representation of the project. represents the gating function used to filter out noisy comments, Represents the summary review vector associated with an item.

[0072] The gating function has two key advantages. First, it is personalized and can be adjusted based on each user's data to achieve better performance. Second, the design ensures that downstream parameters (such as W1, W2) are differentiable with respect to the objective function. This enables the model to support end-to-end learning and simplifies the training process.

[0073] After the above steps, the hierarchical gating network can accurately select comments in an end-to-end manner, effectively filter out noisy comments and improve 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 to achieve better prediction results.

[0074] S2-2-2. Input user information, project information, user summary comment vector and project-related summary comment vector into the global gating module to obtain the user preference vector and project embedding vector respectively.

[0075] The local-level module focuses on the relevance of each review, while the global-level module evaluates the overall value of text features for the prediction task.

[0076] Existing studies have adopted hand-crafted features, such as information entropy, to assess the overall informativeness of an input instance; however, these intuitive approaches have limitations in capturing complex patterns that may be crucial in measuring the usefulness of review features.

[0077] (1) 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.

[0078] In the global gating module, a fully connected layer is applied to the user's summary comment vector to obtain the accumulation of usefulness activation signals from historical comments. The formula is: (8) in, and Indicates the parameters, is the sigmoid activation function, , Represents the accumulation of helpfulness activation signals obtained from historical reviews.

[0079] After the above steps, the overall usefulness of the review features is modeled and effectively evaluated, and a prerequisite is provided for regulating the overall flow of review information related to the prediction task.

[0080] (2) Based on 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 a global gating function.

[0081] In this embodiment, the usefulness activation signal is accumulated based on the historical reviews. , to regulate the overall flow of comment information related to the prediction task and obtain the user's preference vector, the formula is: (9) Among them, ∘ represents the element-by-element product, Indicates user The preference vector of .

[0082] In formula (9), It acts as a global gating mechanism. The basic principle of the global gating module is to explore the global review features , and selectively picks information dimensions that help form user preferences.

[0083] (3) Similarly, according to step S2-2-2, project The global gating module is expressed as:

[0084] (10) in, and is a trainable parameter, Represents the item embedding vector.

[0085] In this embodiment, once the user's preference vector is learned and the item embedding vector Once obtained, it will be passed as input to the multilayer perceptron to estimate the predicted score .

[0086] S2-3. 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 to obtain the user's optimized preference vector and the project optimization vector.

[0087] The local and global gating modules aim to leverage valuable reviews and enhance the representation of users and items, thereby improving the prediction performance. However, due to the lack of supervisory signals, the performance of the models may become unstable. Specifically, in the absence of supervisory labels indicating the relevance of reviews, the local gating module may have difficulty distinguishing relevant from irrelevant comments. It also becomes challenging to evaluate the effectiveness of text features in the global gating module due to the lack of supervisory signals about the overall effectiveness of the reviews. To address this issue, different local and global gating mechanisms are developed using self-supervisory signals using a contrastive learning paradigm.

[0088] S2-3-1. 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.

[0089] In this embodiment, the core concept of local contrastive learning involves The guidance signal is extracted from has a strong correlation. Although it is not available during the testing phase , but it can be used to learn a unique local gating function during training. Specifically, the target review should have a higher relevance score than the randomly sampled reviews.

[0090] During the training phase, the target comments As a positive sample, its relevance score is calculated by formula (4) , and the score of randomly sampled noisy reviews Compare and generate the local contrast loss formula (11) to optimize the user-side threshold ; Optimized threshold The local gating function formula (5) is used to filter out the noise in historical comments, and finally the optimized summary comment vector of the user is generated through formula (6) The specific formula is as follows: ; ; (11) in, Indicates user Target Comments An importance score for predicting its rating, which is used to measure the predictive value of the target review for the current user-item pair. Represents irrelevant comments randomly sampled from the training set The pseudo-correlation score of represents the local contrast loss function on the user side, , , , , are trainable parameters, User-specified thresholds, is the sigmoid activation function, , specifically a comment randomly selected from the training set by the sigmoid activation function.

[0091] The local contrast loss function on the user side optimizes the user-specific threshold by maximizing the correlation difference between the target comments and the noise comments. ; Loss value The smaller it is, the higher the discrimination of the target comments is and the better the model optimization effect is.

[0092] During the training phase, the target review generates the item-side local contrast loss through formula (12) and optimizes the item-specific threshold ; Project-related summary comment vector Calculated by formula (7), its weight Generated by normalizing the filtered comments.

[0093] Similarly, according to formula (11), the objective function of the target review is obtained. According to the objective function, the discriminative item relevance threshold is determined, and the formula is: ; ; (12) in, represents the importance score of the target review of the item to predict the user rating, represents the pseudo-relevance score of irrelevant item reviews randomly sampled from the training set, represents the local contrast loss function on the item side, and h2, W3, W4 and b2 represent model parameters.

[0094] 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 of the target comments and the better the model optimization effect.

[0095] Although the target review is not available during the testing phase, using it during training can provide a discriminative signal. This method aims to propagate self-supervisory signals through the target reviews to identify reviews that are informative for the prediction task. Since the target review has a strong indicative effect on the final prediction for a user-item pair, it should be ranked highest among the historical reviews.

[0096] S2-3-2. Based on 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.

[0097] In the global gating module, an anchor-guided contrastive learning process is introduced to optimize the global gating, which enables it to effectively utilize the overall utility of review features and thus enhance the prediction task.

[0098] To develop a discriminative global gating module, we introduce an anchor-guided contrastive learning approach.

[0099] In this embodiment, represents the global gating function defined in equations (8) and (9), and is expressed as Represents the global gating function described in formula (10), and obtains the user's optimized preference vector and the item optimization vector.

[0100] In the global gating module, a global discrimination mechanism is adopted to generate self-supervision signals through contrastive learning to guide the global gating module to select comment features with higher information content and avoid noise interference.

[0101] S2-3-3. Input the user's optimization preference vector and the project optimization vector into the multi-layer perceptron to obtain the final prediction score.

[0102] In this embodiment, once the user's optimized preference vector is learned and project optimization vector , the user's optimized preference vector and the item's optimized vector are passed as input to the multilayer perceptron to estimate the predicted rating , the formula is: (13) 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.

[0103] At the same time, represents the negative global review vector obtained by setting the local gating value to 1. The global contrast loss function is expressed as follows:

[0104]

[0105]

[0106] (14) in, Represents the marginal value, the positive and negative sample pairs are separated by the marginal value to distinguish; is a global contrastive loss where for each user-item pair, It is the anchor point. is a positive sample, is a negative sample. The global gating module is optimized in such a way that the score corresponding to the anchor point is close to the score of the positive sample and away from the score of the negative sample.

[0107] S2-3-4. The scoring model is optimized and learned through the objective function, which is defined by three components: score prediction loss, local contrast loss, and global contrast loss.

[0108] For the rating prediction loss, the mean square error (MSE) between the predicted rating and the true rating is used to establish the learning objective, as follows: (15) in, Indicates user About Project The actual rating of Represents the model's predicted score for the same user-item pair.

[0109] After combining the local and global contrast loss components, the objective loss function of the model is specifically defined as: (16) Where D represents the training data set, including ,project , their corresponding actual scores and possible distance or similarity measures A tuple consisting of: hyperparameters and Used to trade off between score prediction loss and 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 project side, represents the global contrast loss.

[0110] The scoring model is optimized through target loss to make the model scoring prediction more accurate.

[0111] In step S3, the trained rating prediction model is used to obtain the user's final predicted rating for the project, thus completing the recommendation system rating prediction.

[0112] In this embodiment, when predicting the user's rating of the target product, the input data includes user information (such as age, gender, historical preferences), project information (such as product category, brand) and historical comments (such as the user's evaluation of similar products as "the sound quality is very good"). The hierarchical gating network calculates the relevance score between the comments and the target product through the local gating module (for example, the score of the comments related to "sound quality" is 0.85), and uses personalized thresholds to filter low-relevance comments (such as "slow delivery" with a score of 0.15 is eliminated); the global gating module fuses the filtered comment vector with the user / project embedding to generate a preference representation. Finally, the multi-layer perceptron (MLP) outputs a predicted score (such as 4.3 points) based on the optimized vector. Even if the target comment is missing, the model can still accurately infer user preferences through historical data.

[0113] In this embodiment, the model is evaluated and compared with existing models.

[0114] 1. The mean absolute error (MAE) is used as an evaluation indicator to evaluate the existing model.

[0115] MAE is a commonly used rating prediction metric in recommendation systems. To reduce random bias, each experiment was repeated 10 times and the average results of these 10 runs were reported for effective model comparison. It is worth mentioning that the performance variation between different runs is very small.

[0116] Among them, the existing models are: PMF: A traditional matrix factorization method that predicts ratings by taking the dot product of user and item vectors.

[0117] NeuMF: Combining generalized matrix factorization and multi-layer neural networks for rating estimation.

[0118] DeepCoNN: Use review-based user and item representations for prediction tasks via factorization machines.

[0119] NARRE: Introduced an attention mechanism based on DeepCoNN to more effectively utilize comment information.

[0120] DAML: Proposes to simultaneously learn review features that combine the unique and shared properties of convolutional neural networks (CNNs).

[0121] GC-MC: Uses graph convolution technology to focus on the relationship between users and items, and takes the connection prediction between users and items in the bipartite graph as the recommendation task.

[0122] RMG: Combining graph data and review data to predict ratings.

[0123] RGCL: Improving the understanding of graph architectures and addressing challenges related to user behavior using graph contrastive learning methods.

[0124] These baseline models are selected for representative comparison. For example, PMF and NeuMF models implement 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 high-order connections between users, items, and reviews, thereby achieving more accurate predictions.

[0125] 2. Overall comparison Table 3 summarizes the comparison results on the four datasets. The second-ranked MAE value is underlined. Δ represents the performance improvement of HGD compared to the leading baseline model, where performance improvement of more than 1% is considered significant.

[0126] Table 3 Comparison of models based on different datasets based on MAE

[0127] Based on these data, several key observations can be made.

[0128] (1) It is clear that 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 susceptible to the inherent data sparsity in recommendation problems. In contrast, NeuMF has a significant improvement over PMF in terms of MAE. This improvement is due to its use of a multi-layer neural network that more efficiently captures and models complex user-item interactions.

[0129] (2) Models that incorporate reviews into the framework (Table 3, rows 3-5) consistently outperform latent factor-based models such as PMF and NeuMF. This demonstrates the effectiveness of review features in modeling user behavior and item attributes, demonstrating the important role of these features in improving prediction accuracy. NARRE outperforms DeepCoNN in MAE on all datasets. This finding highlights the advantage of analyzing text information at different granularities. Although DeepCoNN aggregates all past reviews into a comprehensive document, this approach introduces noise and irrelevant information associated with the text. NARRE uses an attention mechanism to identify and select the most informative reviews, resulting in more accurate predictions. DAML proves to be very effective compared to other review-based models. It highlights the ability to successfully detect pairwise connections between different user-item pairs. This approach dynamically reveals the complex interactions specific to each user-item pair, providing a more accurate basis for rating prediction.

[0130] (3) Graph-based models (Table 3, rows 6-8) consistently outperform review-based models significantly. This can be attributed to the inherent ability of graph learning, namely the ability to discover high-order connections in the user-item bipartite graph. This ability enables the generation of more comprehensive representations, thereby improving prediction accuracy. Among the competing models, RGCL ranks highest, which highlights the advantages of enriching the user-item graph with review attributes and leveraging contrastive learning for self-supervision. This analysis provides valuable insights into understanding the relative performance of various models on the dataset. (4) The proposed HGD model outperforms the baseline models on all four datasets. Specifically, compared with the best performing baseline model RGCL, the HGD model achieves 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 leverages historical reviews for comprehensive user / item modeling, thereby improving prediction accuracy. Unlike review-based methods such as DeepCoNN, which indiscriminately concatenates all reviews into one document, the proposed model is able to 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 through personalized thresholds. Finally, compared to graph-based methods such as RGCL, HGD has the advantage of providing supervisory signals to enhance the discovery of useful reviews. Instead of learning high-level general representations from interaction data, HGD guides the exploration of historical reviews through a hierarchical discriminative network to support the prediction of specific goals.

[0131] 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 the comments and identifying the most relevant comments. This gating mechanism effectively solves the noise problem in the comments and enables the model to fully utilize relevant comments to complete the prediction task. Second, the hierarchical discriminant network we introduced has made a significant contribution to the success of the model. The network transfers knowledge from the target comment and transmits self-supervisory signals at different granularity levels, which helps to form the local discriminative gating function and guides the HGD model to focus on relevant text signals.

[0132] 3. Performance of each submodule in the HGD model For this purpose, HGD is compared with the following variants: HGD-local: This is a version of HGD without the local discriminative module. Therefore, there is no supervisory signal when learning the discriminative local gating module, which is treated as a traditional attention mechanism.

[0133] HGD-global: This is a version of HGD without a global discriminative module. Therefore, there is no supervised learning mechanism for the discriminative global gating module and it is treated as a traditional gating network. Table 4. Comparison of different model variants

[0134] Table 4 shows the comparison results, and the following conclusions can be drawn: (1) HGD outperforms all variants, which highlights the effectiveness of its sub-modules in filtering noise and leveraging supervisory signals.

[0135] (2) The performance of HGD-local drops more significantly than that of HGD-global, indicating that denoising at the local review level is very necessary.

[0136] (3) The comparison between HGD and its variants further confirms the effectiveness of extracting insights from target reviews to generate supervisory signals and establish a cross-level discriminative gating mechanism. It can be inferred that if there is a lack of supervisory signals indicating review relevance, the hierarchical gating network may fail to extract review features that are informative for the prediction task, which may lead to poor learning of relevant parameters.

[0137] 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-supervision signals when selecting reviews. In this comparison, the performance is evaluated by two metrics: mean relevance ratio (MRR) and its weighted version MRR-W. MRR represents the average ratio of the number of relevant reviews selected to the number of relevant reviews with a score above a threshold. MRR-W combines the score weight of relevance on top of MRR. Higher MRR and MRR-W scores indicate that a larger proportion of relevant reviews are selected. Below are the definitions of MRR and MRR-W: , (17) in Represents the test set Show historical comments Comments with target Belong to the same rating category.

[0138] As shown in Figure 2, the results clearly show that HGD consistently outperforms ATT in both MRR and MRR-W. This indicates that HGD is more effective in selecting relevant reviews than traditional attention mechanisms, further highlighting the advantage of HGD in leveraging historical reviews to provide accurate rating predictions.

[0139] 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. Based on the chart, the following observations can be made: (1) HGD has more records than ATT at higher RR, but less records than ATT at lower RR, which confirms the effectiveness of HGD in discovering more relevant reviews to improve rating prediction.

[0140] (2) The RR improvement on the “Musical Instrument” dataset is relatively small, which also explains why HGD does not show significant improvement over HGD-local, which also uses the traditional attention mechanism to select comments.

[0141] (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” approach to select reviews, which may accidentally discard some reviews that are actually informative.

[0142] 4. Case Analysis To gain a deeper understanding of the effectiveness of HGD in identifying relevant reviews for prediction, we present some visualizations that illustrate the relevance scores of historical reviews during the rating estimation process. To achieve this, user-item entries from the test subset of the “Sports and Outdoors” dataset were randomly selected. Then, for each user and their historical reviews, we plotted (See formula (4)) and (See the diagram of formula (5)). Figures 4 to 8 As shown in Figure 1, these visualizations show three user-item pairs. In these figures, the x-axis represents the historical reviews, while the y-axis represents the corresponding In addition, a personalized threshold is included In the figure, the “grey” bars represent comments related to the target comment, meaning that their sentiment is consistent with the target comment; on the contrary, the “white” bars are used to represent irrelevant comments.

[0143] As can be seen from the figure, relevant comments usually receive higher relevance scores compared to irrelevant comments. 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 is able to effectively utilize the supervisory signals of the target comments and successfully distinguish relevant comments from irrelevant ones. The proposed HGD model demonstrates a strong ability to enhance prediction accuracy by effectively utilizing relevant comments. This ability is due to the adoption of a hierarchical gating module that automatically integrates informative semantics while suppressing irrelevant noise.

[0144] In this embodiment, a hierarchical gating and discriminant framework is constructed for rating prediction. Many challenges existing in traditional recommendation systems, such as text noise, lack of supervision information, and flat structure of comment features, are solved. A hierarchical gating network is used specifically for extracting features from comments. The task of the local gating module is to selectively identify comments that are useful and relevant to the target prediction; while the global gating module evaluates the effectiveness of the overall comment signal. In order 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 rating. This network generates self-supervisory signals to help learn discriminative gating at the local and global levels. The hierarchical gating and discriminant networks work together to enable the model to extract valuable comment features, thereby improving prediction accuracy. The superior effect of the HGD method is verified by detailed experiments on public datasets, showing that it outperforms existing state-of-the-art benchmarks.

[0145] Embodiment 2 The purpose of this embodiment is to provide a recommendation system rating prediction system based on hierarchical gating and discriminant network, including: 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 rating module is used to use the trained rating prediction model to complete the user's final rating prediction for the project.

[0146] Based on providing a recommendation system rating prediction system based on hierarchical gating and discriminant network, the method steps in embodiment 1 are implemented.

[0147] Embodiment 3 The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0148] Embodiment 4 The purpose of this embodiment is to provide a computer-readable storage medium.

[0149] A computer-readable storage medium stores a computer program, which executes the steps of the above method when executed by a processor.

[0150] Embodiment 5 The purpose of this embodiment is to provide a computer program product containing instructions, which, when running on a computer, enables the computer to execute the methods and functions involved in any of the above embodiments. The steps involved in the apparatus of the above embodiment correspond to the method embodiment 1, and the specific implementation method can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including 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.

[0151] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code 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 made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0152] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection 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 recommendation system rating prediction. 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