A personalized recommendation-oriented potential positive sample mining method
By identifying potential positive samples through semantic factor routing and alignment modules, the problem of mislabeling negative samples in sparse data in recommender systems is solved, improving recall and computational efficiency, and realizing efficient personalized recommendations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUILIN UNIV OF ELECTRONIC TECH
- Filing Date
- 2026-03-11
- Publication Date
- 2026-06-09
AI Technical Summary
Existing recommendation systems, when processing sparse data, incorrectly label items that users have not interacted with as negative samples, exacerbating the data sparsity problem and reducing the model's generalization ability. Furthermore, graph convolution-based methods are computationally expensive and overly smooth, making it difficult to effectively capture high-order connectivity.
The semantic factor routing module deconstructs item representations into independent semantic factors, constructs a semantic factor matching module to identify potential positive samples, optimizes the distribution of user and item representations through a semantic pair alignment module, and improves recommendation accuracy and efficiency by using a joint loss function.
It effectively identifies potential positive samples, improves recall and normalized loss cumulative gain, reduces computational complexity, achieves efficient training and inference, and enhances the generalization ability of recommendation systems in sparse data scenarios.
Smart Images
Figure CN122175672A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of neural networks, recommender systems, and unentangled representation learning, specifically to the design of a latent positive sample mining method for personalized recommendations. Background Technology
[0002] Recommendation systems are a crucial tool for addressing information overload, widely applied in e-commerce platforms, online video websites, and social media, profoundly impacting various stakeholders including platforms, users, and content providers. By analyzing users' historical purchasing behavior, recommendation systems predict their potential interests and accurately recommend products or services that match their preferences. This not only helps users quickly find the content they need amidst massive amounts of information, improving the user experience, but also enhances the platform's competitiveness in the market. Existing recommender systems are mostly based on collaborative filtering, typically treating items not interacted with by the user as negative samples for model training. However, in practical applications, due to the extreme sparsity of user behavior data, a large number of items that users are interested in but have not yet encountered (i.e., potential positive samples) are incorrectly labeled as negative samples. This erroneous negative sampling assumption not only wastes valuable data information and exacerbates the data sparsity problem, but also severely limits the model's generalization ability. To alleviate this problem, some existing technologies have introduced graph neural networks, attempting to incorporate high-order neighborhood nodes into a unified modeling process through message passing mechanisms. However, while graph convolution-based methods attempt to capture high-order signals, they are limited by high computational costs and the over-smoothing problem of deep networks, making it difficult to effectively capture high-order connectivity under sparse supervised signals, ultimately leading to failure or inefficiency in large-scale sparse data scenarios. As a cutting-edge direction in the field of artificial intelligence, unentangled representation learning aims to separate independent semantic factors with clear semantic meaning from observable data. Although unentangled representation learning can separate the mixed features of objects into independent semantic factors, making fine-grained analysis possible, existing technologies mostly stop at factor separation and fail to effectively utilize these semantic factors to re-examine non-interactive data, lacking a method to accurately identify potential positive samples from non-interactive data. Summary of the Invention
[0003] I. Technical problems to be solved Traditional collaborative filtering methods typically rely on the strong assumption of "unobserved is negative feedback" when constructing training objectives, treating all items that users haven't interacted with as negative samples. However, this approach ignores the data sparsity problem prevalent in recommendation scenarios. This leads to a large number of potential positive samples that users are actually interested in but haven't yet encountered being incorrectly labeled as negative examples. The model is forced to learn incorrect discrimination boundaries during optimization, making it difficult to effectively extract true user preferences from massive amounts of uninterrupted data, ultimately resulting in decreased recommendation generalization ability. While graph convolution-based methods attempt to capture higher-order signals, they are limited by high computational costs and the over-smoothing problem of deep networks, making it difficult to effectively capture higher-order connectivity under sparse supervised signals. Ultimately, they fail or become inefficient in large-scale sparse data scenarios. To address the shortcomings of traditional recommendation models, this invention proposes a latent positive sample mining method for personalized recommendations. This method involves using a semantic factor routing module to dewrap item representations into independent semantic factors, constructing a semantic factor matching module to identify non-interacting items with the same factors as latent positive samples, and then using a semantic pair alignment module to simultaneously optimize the representation distribution of observed interaction pairs and latent positive sample pairs, promoting the uniformity of user and item representations. This invention achieves high-order signal capture without complex graph neighborhood aggregation, effectively mitigating the false negative example labeling problem and improving recommendation accuracy and computational efficiency in sparse scenarios. II. Technical Solution Step 1: Download a publicly available recommendation system dataset containing user-item interaction records. Perform preprocessing: First, remove duplicate interaction records. Then, retain user and item entities with at least 5 interactions to filter noise and ensure data sparsity is within a reasonable range. Re-index and re-encode the retained user and item entities to construct a sparse interaction matrix. Finally, randomly divide each user's interaction records into training, validation, and test sets in an 80 / 10 / 10 ratio. Step 2: Import the preprocessed interaction matrix from Step 1, initialize the user and item embedding vectors, and then perform L2 normalization. This process does not rely on complex graph neighborhood aggregation; it generates initial user representation vectors and item embedding vectors based solely on observed interaction signals, avoiding oversmoothing and reducing computational costs. Step 3: Construct a Semantic Factor Routing (SFR) module. Input the item embedding vector obtained in Step 2 into this module. Through multiple iterations, the item representation is decoupled into multiple independent and globally shared semantic factor vectors. Each factor represents the feature of the item in a specific semantic dimension (such as category, style, function, etc.), thereby obtaining a set of fine-grained semantic factors for subsequent matching. Step 4: Based on the semantic factor set obtained in Step 3, construct a semantic factor matching module (SFM). For each item that a user has interacted with, calculate the similarity between its semantic factors and the semantic factors of uninteracted items; set a similarity threshold to identify those uninteracted items that share high similarity semantic factors, and mark them as "potential positive sample pairs," thereby enriching the sparse supervision signal and alleviating the problem of false negatives. Step 5: Construct a Semantic Pairs Alignment (SPA) module, using the latent positive sample pairs mined in Step 4 and the observed interaction pairs as input for positive sample pairs. Design a joint loss function that, on the one hand, brings the user representation closer to the observed / potential positive sample item representation (alignment), and on the other hand, widens the distance between different user and different item representations (uniformity), thereby promoting a uniform distribution of user and item representations in the feature space. Step 6: Iteratively update the model parameters based on the gradient information from Step 5 until convergence. Explore model performance under different numbers of semantic factors, similarity thresholds, and loss function weights. Combine multi-dimensional evaluation metrics such as recall, Normalized Discounted Cumulative Gain (NDCG), and training time to select the optimal parameter combination. Finally, use the trained model to calculate the matching score between users and candidate items, generating a recommendation list that meets users' personalized needs and has high generalization ability. III. Beneficial Effects Compared with existing technologies, the present invention has the following advantages: 1. Current mainstream collaborative filtering methods often simply treat uninteracted items as negative samples, leading to a large number of potentially interesting items being misclassified as negative examples, severely limiting the model's generalization performance in sparse data scenarios. This invention addresses this limitation by innovatively introducing a semantic factor matching module. By unwrapping the semantic factors of items and identifying uninteracted items sharing these factors as potential positive samples, it effectively enriches the sparse supervision signal. Experiments show that this method significantly improves core recommendation metrics such as recall and normalized loss cumulative gain, more accurately capturing users' true potential preferences and meeting the actual business requirements for high-coverage recommendations. 2. Existing technologies often rely on graph convolutional networks to capture high-order connectivity, but they suffer from high computational costs and oversmoothing issues caused by deep networks, resulting in reduced distinguishability between user and item features and making them difficult to apply effectively to large-scale data. This invention aligns the mined potential positive sample pairs with observed interaction pairs, eliminating the need for complex graph neighborhood aggregation operations, and implicitly transmits high-order cooperative signals using globally shared semantic factors. This mechanism breaks away from the traditional graph methods' reliance on stacked multi-layer networks, fundamentally avoiding oversmoothing, while significantly reducing computational complexity and memory consumption, enabling efficient training and inference in resource-constrained environments. Attached Figure Description
[0004] Figure 1 Overall flowchart of the present invention Figure 2 Overall model framework diagram of the present invention Detailed Implementation
[0005] To clearly illustrate the purpose, technical solution, and advantages of this invention, the invention will be further described in detail below with reference to specific examples and accompanying drawings. This invention implements project recommendations based on users' historical interaction data. The overall process is shown in Figure 1, and the specific steps are as follows: Step 1: Obtain publicly available Gowalla, Yelp 2018, Toys-and-Games, and Beauty recommendation datasets. Perform preprocessing operations on the datasets: First, remove duplicate interaction records; then, retain user and item entities with at least 5 interactions to filter noise and ensure data sparsity is within a reasonable range; re-index and re-encode the retained user and item entities to construct a sparse interaction matrix; finally, randomly divide each user's interaction records into training, validation, and test sets in an 80 / 10 / 10 ratio. Step 2: Import the preprocessed interaction matrix from Step 1 and perform user embedding vector processing. With item embedding vector Initialization is performed. Then, the embedding vector is L2 normalized and mapped onto the unit hypersphere to obtain the normalized representation. and : in, Representing vectors Norm: This process does not rely on complex graph neighborhood aggregation, but generates initial user representation vectors and item embedding vectors based solely on observed interaction signals, avoiding oversmoothing problems and reducing computational costs. Step 3: Construct a semantic factor routing module by inputting the item embedding vector obtained in Step 2 into this module. Introduce learnable routing logic values. To control items To semantic factors The adaptive route allocation is initialized with small random values that follow a normal distribution. This provides an unbiased starting point for subsequent routing iterations. In each routing iteration, Softmax normalization is first applied to the routing logical values to update the semantic allocation weights. Then, independent semantic factors are generated through weighted summation and normalization operations. Finally, the routing logic value is updated using the consistency feedback between items and factors. This semantic factor routing process can be represented as: in, This indicates the operation of the natural exponent. This represents the total number of preset semantic factors. Indicates the first The intermediate representation vector of each semantic factor This represents the unit vector representation after L2 normalization. Step 4: Based on the semantic factor set obtained in Step 3, construct a semantic factor matching module. This module aims to establish semantic associations between items by sharing semantic factors, thereby mining potential positive samples. Specifically, firstly, for each item... Weights are assigned based on their semantics. Select the top contributors A semantic factor is used as its dominant feature; then a semantic threshold is introduced. The correlation strength between items is dynamically controlled, meaning that only when two items share at least one dominant semantic factor will the correlation strength be determined. Only when a semantic relationship is established between two items is a semantic association determined. Finally, based on the above association rules, a semantically enhanced set is constructed for each interacted item, and non-interacted items are considered semantically consistent extensions, generating augmented positive sample pairs for use by subsequent modules. This semantic factor matching and potential positive sample mining process can be represented as: in Represents items The dominant semantic factor index set, containing the largest weights The number of each factor; It represents the collection of all items. Indicates items All other items; This indicates the number of elements in the set (i.e., the size of the intersection); The preset semantic similarity threshold is used to determine whether two items share enough semantic features. Represents items A semantically enhanced set, in which any item All are considered Potential positive samples for users. With items Each observation interaction instance The semantic factor matching module will collect All items Constructing augmented positive sample pairs This method effectively identifies items that, although they do not interact, share highly similar semantic factors and marks them as potential positive sample pairs, thereby enriching the sparse supervision signal and alleviating the problem of false negatives. Step 5: Construct a semantic pair alignment module, taking the potential positive sample pairs mined in Step 4 and the observed interaction pairs as input positive sample pairs. Design a joint loss function to, on the one hand, shorten the distance between user representations and observed / potential interaction item representations (alignment), and on the other hand, widen the distance between different user and different item representations (uniformity), thereby promoting a uniform distribution of user and item representations in the feature space. The alignment loss between interacting users and items is: in, This represents the observed positive sample interaction distribution. This represents a combination of users and items randomly selected from all observed interaction pairs. Represents the distribution of positive samples Mathematical expectation operation, This represents the square of the Euclidean distance. The homogenization loss among all user / item nodes is: in, Represents the natural logarithm operation. This indicates the operation of the natural exponent. and These represent the distribution of users and projects, respectively. and These represent node pairs randomly sampled from the user set and the item set, respectively. and These represent the pairs of nodes that are independently and randomly sampled from the set of all users and the set of all items, respectively. and The calculated index similarity values are averaged. The alignment loss between users and potential interactive items is: The final total loss function is defined as: in, and These are hyperparameters used to control the relative importance of the homogenization loss among all user / item nodes and the alignment loss between users and potential interactive items. Step 6: Iteratively update the model parameters based on the gradient information from Step 5, using the Adam optimizer to optimize the model parameters iteratively until convergence. By exploring the model performance under different numbers of semantic factors, similarity thresholds, and different loss function weight settings, and combining multi-dimensional evaluation metrics such as recall, normalized loss cumulative gain, and training time, the optimal parameter combination is selected. Finally, the trained model is used to calculate the matching score between users and candidate items, generating a recommendation list that meets users' personalized needs and has high generalization ability. It should be noted that although the embodiments described above are illustrative, they are not intended to limit the invention. Therefore, the invention is not limited to the specific embodiments described above. Any other embodiments obtained by those skilled in the art under the guidance of this invention without departing from its principles are considered to be within the protection scope of this invention.
Claims
1. A method for mining potential positive samples for personalized recommendation, mainly including the following steps: Step 1: Download the publicly available recommendation system dataset that includes user and item interaction records; perform preprocessing operations: first, remove duplicate interaction records, and then retain user and item entities with at least 5 interactions to filter noise and ensure that data sparsity is within a reasonable range. The retained user and item entities are indexed and recoded to construct a sparse interaction matrix; finally, the interaction records of each user are randomly divided into training set, validation set and test set in a ratio of 80 / 10 / 10. Step 2: Import the preprocessed interaction matrix from Step 1, initialize the embedding vectors of users and items using lightweight matrix decomposition, and perform L2 normalization. This process does not rely on complex graph neighborhood aggregation, but generates initial user representation vectors and item embedding vectors based solely on observed interaction signals, thus avoiding oversmoothing problems and reducing computational costs. Step 3: Construct a semantic factor routing module and input the item embedding vector obtained in Step 2 into this module; through multiple iterations, decouple the item representation into multiple independent and globally shared semantic factor vectors, each factor representing the feature of the item in a specific semantic dimension, thereby obtaining a fine-grained set of semantic factors for the item. Step 4: Based on the semantic factor set obtained in Step 3, construct a semantic factor matching module; for each item that a user has interacted with, calculate the similarity between its semantic factors and the semantic factors of items that have not been interacted with. By setting a similarity threshold, we can identify non-interactive items that share high similarity semantic factors but do not interact with each other, and mark them as "potential positive sample pairs" to enrich the sparse supervision signal and alleviate the problem of false negatives. Step 5: Construct a semantic pair alignment module, taking the potential positive sample pairs mined in Step 4 and the original observed interaction pairs as positive sample pair inputs; design a joint loss function, on the one hand, to bring the user representation closer to the observed / potential positive sample item representation (alignment), and on the other hand, to widen the distance between different user and different item representations (uniformity), thereby promoting the uniform distribution of user and item representations in the feature space; Step 6: Iteratively update the model parameters based on the gradient information from Step 5 until convergence; explore the model performance under different semantic factor numbers, similarity thresholds, and different loss function weight settings, and select the optimal parameter combination by combining multi-dimensional evaluation indicators such as recall, normalized loss cumulative gain, and training time; finally, use the trained model to calculate the matching score between users and candidate items to generate a recommendation list that meets users' personalized needs and has high generalization ability.
2. The latent positive sample mining method for personalized recommendation according to claim 1, characterized in that, In step 2, the specific process of initializing and normalizing the user embedding vector and the item embedding vector is as follows: in, and Represents the embedding vectors of users and items; Representing vectors Norm.
3. The latent positive sample mining method for personalized recommendation according to claim 1, characterized in that, In each round of routing iteration, Softmax normalization is first applied to the routing logical values to update the semantic assignment weights. Then, independent semantic factors are generated through weighted summation and normalization operations. Finally, the routing logic value is updated using the consistency feedback between items and factors. This semantic factor routing process can be represented as: in, This represents the total number of preset semantic factors. Represents a collection of items. Indicates the first An intermediate representation vector of semantic factors.
4. The latent positive sample mining method for personalized recommendation according to claim 1, characterized in that, For each item Weights are assigned based on their semantics. Select the top contributors Each semantic factor serves as its dominant feature; Subsequently, a semantic threshold was introduced. By dynamically controlling the correlation strength between items, this semantic factor matching and potential positive sample mining process can be represented as: in, Represents items The dominant semantic factor index set, Indicates items All other items; This indicates the number of elements in the set (i.e., the size of the intersection). The preset semantic similarity threshold, Represents items A semantically enhanced set, in which any item All are considered Potential positive samples.
5. The latent positive sample mining method for personalized recommendation according to claim 4, characterized in that: The discovered potential positive sample pairs and the original observed interaction pairs are used together as positive sample pair inputs; the semantic pair alignment and model optimization process can be represented as follows: The final loss function is expressed as: in, This represents the observed positive sample interaction distribution. This represents a combination of users and items randomly selected from all observed interaction pairs. Represents the distribution of positive samples Mathematical expectation operation, Represents the square of the Euclidean distance. Represents the natural logarithm operation. This indicates the operation of natural exponents. and These represent the distribution of users and projects, respectively. and These represent node pairs randomly sampled from the user set and the item set, respectively. and These represent the pairs of nodes that are independently and randomly sampled from the set of all users and the set of all items, respectively. and The mean of the calculated exponential similarity, and These are hyperparameters used to control the relative importance of the homogenization loss among all user / item nodes and the homogenization loss among semantic factors, respectively.