A data model driven enhanced trusted recommendation method

Abstract

The application discloses a data model driving enhanced trusted recommendation method, belonging to the technical field of artificial intelligence, data mining and intelligent recommendation. The method comprises the following steps: constructing a data-driven user-item interaction sequence; generating an enhanced user behavior sequence based on semantic perception diffusion; constructing a candidate enhanced edge set; evaluating the original interaction edge and the candidate enhanced edge by using a large language model, and performing differentiated edge discarding and potential edge adding according to the evaluation result to obtain an enhanced user-item interaction graph; combining graph contrast learning, recommendation main task and diffusion denoising constraint to optimize the recommendation model and generate a recommendation list. The method can reduce the semantic deviation caused by random sequence disturbance and the damage to the key interaction structure caused by random graph enhancement, alleviate the problems of insufficient learning of long-tail items and uncontrollable quality of candidate enhanced edges, and improve the reliability of enhanced data, the representation learning ability of the recommendation model and the credibility of the recommendation result.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, data mining and intelligent recommendation technology, and specifically relates to a reliable recommendation method based on data model-driven enhancement. Background Technology

[0002] With the rapid development of internet platforms and smart applications, a large amount of user behavior data has been generated in scenarios such as e-commerce, short videos, music, and film and television. Recommendation systems can mine user interests and preferences based on users' historical interaction behavior and item feature information, providing personalized content and serving as an important technical means to alleviate information overload. In recent years, methods such as graph neural networks, sequence recommendation models, large language models, and generative models have been widely applied in recommendation systems. These methods can learn representations of users and items from user behavior sequences and user-item interaction graphs, thereby improving recommendation performance.

[0003] In sequence recommendation scenarios, existing methods typically construct augmented behavior sequences through random masking, pruning, replacement, reordering, or diffusion perturbation to provide richer training signals for the model. However, real-world recommendation data generally exhibits a significant long-tail distribution, with a small number of popular items accounting for a large amount of interaction, while a large number of low-frequency items have fewer interactions. This makes the model more prone to biasing towards popular items, making it difficult to fully learn the representations of long-tail items. Furthermore, existing diffusion-based sequence augmentation methods often use standard Gaussian noise for perturbation. This type of noise lacks the ability to perceive the semantic structure of items and may disrupt the original semantic proximity relationships between items, causing semantic shifts in the augmented sequence and reducing the effectiveness of the augmented data.

[0004] In graph recommendation scenarios, the interaction between users and items can typically be modeled as a user-item bipartite graph. Graph neural networks can capture high-order collaborative signals between users and items through neighbor aggregation, demonstrating good representation learning capabilities for recommendation tasks. To alleviate the data sparsity problem, graph contrastive learning methods often construct different augmented graph views by discarding edges, nodes, or performing random walks, and constrain the consistency of node representations across different views through contrastive learning. However, most existing graph augmentation methods rely on random perturbations or heuristic strategies, which can easily disrupt key interaction structures in the user-item graph. This can lead to the deletion of important interactions that truly reflect user preferences, or the retention of noisy interactions with weak relevance to user interests, thus affecting the quality of the augmented graph view.

[0005] Furthermore, existing sequence augmentation and graph augmentation methods often lack effective integration. Sequence augmentation can generate potential interest items from user behavior sequences, but its results often remain at the sequence representation level, lacking further structured verification and graph structure organization. Graph augmentation methods can perturb and optimize user-item interaction structures, but the generation process of candidate interaction edges often lacks semantic drive from users' temporal behavior, easily introducing interaction relationships that do not conform to users' true interests. Large language models and pre-trained language models have strong semantic understanding and reasoning capabilities, and can determine the relevance of interaction relationships to user preferences based on users' historical interaction information and item semantic information, providing support for the semantic verification of candidate interaction edges.

[0006] Therefore, existing data augmentation methods for recommender systems still have the following shortcomings: On the one hand, random masking and Gaussian perturbation in sequence augmentation can easily ignore the learning needs of long-tail items and may destroy the semantic relationships in the sequence; on the other hand, random edge dropping and random edge addition in graph augmentation lack semantic guidance and can easily destroy key interaction structures or introduce noisy interactions; in addition, there is a lack of effective connection between the sequence augmentation results and the graph structure augmentation process, making it difficult to form a complete process from potential interest generation to interaction graph augmentation. Summary of the Invention

[0007] To address the problems of existing technologies mentioned in the background section, this invention proposes a data model-driven enhancement-based reliable recommendation method. This method establishes a data-driven foundation based on user historical interaction data, interaction time sequence, item interaction frequency, and user-item interaction relationships. First, semantic awareness diffusion is used to generate potential interest items from user behavior sequences. Then, these potential interest items are mapped to candidate user-item interaction edges. A large language model is used to evaluate the semantic importance and verify the reliability of the original interaction edges and candidate enhancement edges, forming a model-driven approach. This solves the problems in existing recommendation systems, such as semantic shift caused by random sequence enhancement, insufficient learning signals for long-tail items, damage to key interaction structures caused by random graph enhancement, and uncontrollable quality of candidate enhancement edges. Therefore, this method improves the representation learning ability of the recommendation model and the reliability of the recommendation results.

[0008] Technical Solution: To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] A data model-driven enhancement-based trusted recommendation method includes the following steps:

[0010] Step 1: Obtain historical user interaction data from the recommendation system, preprocess the historical user interaction data, construct user behavior sequences according to interaction time to obtain the original user behavior sequences, and retain the historical interaction relationships between users and items.

[0011] Step 2: Perform frequency-aware masking on the user behavior sequence based on the frequency of item interaction, and use semantic-aware noise diffusion and context-aware denoising to perturb and restore the masked position to generate an enhanced user behavior sequence.

[0012] Step 3: Generate candidate user-item interaction edges. Based on the potential interest items recovered from the enhanced user behavior sequence, form a set of candidate enhancement edges.

[0013] Step 4: Use a large language model to perform credibility verification and importance evaluation on the original interaction edges and candidate enhancement edges, and perform differential edge discarding and potential edge addition based on the evaluation results to obtain the enhanced user-item interaction graph;

[0014] Step 5: Combine graph comparison learning, recommendation main task and diffusion denoising constraint to optimize the recommendation model, predict the target user's preference score for candidate items, and generate a recommendation list based on the preference score.

[0015] Preferably, in step 1, the specific implementation details are as follows:

[0016] Obtain users' historical interaction information from target recommendation platforms or public recommendation datasets. The historical interaction information includes user identifier, item identifier, interaction behavior type, interaction time, and item semantic information.

[0017] The historical interaction information is cleaned, duplicate interactions are merged, and invalid interactions are filtered to obtain the preprocessed user historical interaction records.

[0018] Based on the preprocessed user history interaction records, the historical interaction items of each user are sorted according to the interaction time to obtain the user's... The sequence of behaviors is as follows:

[0019] ;

[0020] in, Indicates user The length of the historical interaction sequence; Indicates user In the Items that can be interacted with at a specific location; This represents the original sequence of user u's actions;

[0021] Simultaneously, the original user-item interaction edge set is preserved based on the historical interaction relationships between users and items; when there is a historical interaction between user u and item i, an interaction edge is established between user u and item i to obtain the original user-item interaction graph. ;

[0022] For user u, its set of historical interaction edges Represented as:

[0023] ;

[0024] in, This represents the set of items that user u has interacted with in the past.

[0025] As a preferred option, the specific implementation details in step 2 are as follows:

[0026] Step 2.1: Construct a representation of user behavior sequences;

[0027] Step 2.2: Perform frequency-aware masking;

[0028] Step 2.3: Construct semantically aware noise and perform diffusion perturbation;

[0029] Step 2.4: Perform context-aware denoising.

[0030] As a preferred option, the specific implementation details in step 2.1 are as follows:

[0031] For the original behavior sequence of user u Using an encoder The user's original behavior sequence is encoded to obtain the context representation of each position, where the hidden representation of the i-th position is:

[0032] ;

[0033] in, The context representation of the i-th position in the original action sequence of user u;

[0034] Then, the hidden state at the last position of the user behavior sequence is taken as the representation of the user's current preference, specifically:

[0035] ;

[0036] in, This represents the current preferences of user u. This represents the context representation of the Lth position in the original action sequence of user u.

[0037] As a preferred option, the specific implementation details in step 3 are as follows:

[0038] Step 3.1: Identify potential interest items in the enhanced user behavior sequence;

[0039] Step 3.2: Construct candidate user-item interaction edges.

[0040] As a preferred option, the specific implementation details in step 3.2 are as follows:

[0041] For the set of potential interest items obtained in step 3.1 Match user u with a set of items of potential interest Candidate connections are established for item i in the data to form a set of candidate user-item interaction edges corresponding to user u, specifically:

[0042] ;

[0043] in, This represents the set of candidate enhancement edges corresponding to user u. Represents the candidate user-item interaction edge between user u and potential interest item i; This represents the set of potentially interesting items derived from the sequence of enhanced user behavior.

[0044] For all users, the candidate enhancement edge sets corresponding to each user are merged to obtain the global candidate enhancement edge set, specifically:

[0045] ;

[0046] in, Represents a set of users. This represents the set of candidate user-item interaction edges obtained by the enhanced user behavior sequence mapping; Let represent the set of candidate enhancement edges corresponding to user u.

[0047] As a preferred option, step 4 specifically includes the following implementation details:

[0048] Step 4.1: Perform embedding enhancement based on the large language model;

[0049] Step 4.2: Perform semantic verification and importance assessment of interaction edges based on the large language model;

[0050] Step 4.3: Perform differential edge discarding;

[0051] Step 4.4: Perform semantic verification of candidate edges and add potential edges.

[0052] As a preferred option, the specific implementation details in step 4.1 are as follows:

[0053] Using a large language model, semantic embeddings of items are generated based on their textual descriptions, names, categories, and tag information. Its semantic embedding is denoted as And integrate all item semantic embeddings into an item embedding matrix A;

[0054] Principal component analysis (PCA) is performed on the item embedding matrix A to reduce its dimensionality while preserving the main semantic information, resulting in a low-dimensional item semantic embedding matrix, as follows:

[0055] ;

[0056] in, The item embedding matrix is ​​the result of mean-centered optimization. For the reason before The projection matrix composed of eigenvectors; For low-dimensional item semantic embedding matrix;

[0057] For users Based on its historical collection of interacting items The initial semantic embedding of the user is obtained by aggregating the low-dimensional semantic embeddings of the items, specifically:

[0058] ;

[0059] in, This represents the set of items that user u has interacted with in the past. This indicates the number of items that user u has interacted with in the past. Represents items semantic embedding, Indicates user Semantic embedding.

[0060] As a preferred option, step 5 specifically includes the following implementation details:

[0061] Step 5.1: Construct an enhanced graph view based on the enhanced user-item interaction graph;

[0062] Step 5.2: Perform graph comparison learning based on the enhanced graph view;

[0063] Step 5.3: Perform inter-layer comparison learning based on the representations of different layers of the graph neural network;

[0064] Step 5.4: Optimize the recommendation model based on the recommendation task and the contrastive learning task.

[0065] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0066] (1) This invention provides data-driven basis through user historical interaction data, item frequency information and user-item interaction relationship, and provides model-driven constraints through diffusion model and large language model, so as to realize collaborative optimization of candidate interest generation, candidate interaction edge verification and interaction graph enhancement.

[0067] (2) This invention reduces the semantic shift caused by random sequence perturbation and the damage to key interaction structures caused by random graph enhancement by evaluating the semantic importance of interaction edges, verifying the credibility of candidate edges and updating differentiated edges. It alleviates the problems of insufficient learning of long-tail items and uncontrollable quality of candidate enhancement edges, and improves the reliability of enhancement data, the representation learning ability of recommendation models and the credibility of recommendation results.

[0068] (3) This invention further maps the enhanced user behavior sequence generated by semantic perception diffusion into candidate user-item interaction edges, establishes a connection mechanism between user temporal behavior enhancement and user-item graph structure enhancement, so that the candidate new interaction relationship comes from the user's historical behavior and potential interests, rather than relying entirely on random edge addition or heuristic construction; at the same time, it uses a large language model to perform semantic credibility verification on the candidate enhancement edges, so that the potential interest relationship that passes the verification can participate in the subsequent graph structure enhancement, thereby reducing the possibility of irrelevant interaction and noisy interaction being introduced into the user-item interaction graph, improving the utilization efficiency of the sequence enhancement result and the semantic rationality of the enhanced graph structure.

[0069] (4) This invention integrates the recommendation main task, graph contrastive learning constraints, inter-layer contrastive learning constraints of graph neural networks, and semantic perception diffusion denoising constraints during model training. This enables the recommendation model to simultaneously utilize temporal interest information from user historical behavior, structural information in the enhanced user-item interaction graph, and semantic association information between items. Among these, graph contrastive learning enhances the consistency of user and item representations under different graph-enhanced views, inter-layer contrastive learning reduces the impact of information decay and oversmoothing during multi-layer graph propagation, and diffusion denoising constraints improve the semantic consistency of the potential interest generation process. Through these multi-level collaborative constraints, the stability and discriminative ability of user and item representations are improved, and the impact of random perturbations, noisy interactions, and unreliable candidate edges on the recommendation model training process is reduced, thereby improving the overall stability and credibility of the recommendation results. Attached Figure Description

[0070] Figure 1 This is a flowchart illustrating the data model-driven enhancement-based trusted recommendation method of this invention.

[0071] Figure 2 This is the frequency-aware mask and semantic-aware diffusion model framework in this invention;

[0072] Figure 3 This is a schematic diagram of the interaction edge enhancement strategy based on a large language model in this invention. Detailed Implementation

[0073] The present invention will be further illustrated below with reference to specific embodiments. These embodiments are implemented based on the technical solutions of the present invention, and it should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0074] This embodiment provides a data model-driven enhancement-based trusted recommendation method. It generates potential interest items based on user behavior sequences and maps them as candidate user-item interaction edges. Then, it uses a large language model to perform semantic verification of candidate edges and graph structure enhancement. By combining data-driven potential interest generation and model-driven candidate edge verification, the quality of the enhanced interaction graph and the representation learning ability of the recommendation model are improved.

[0075] like Figure 1 As shown, the specific steps include:

[0076] Step 1: Obtain historical user interaction data from the recommendation system, preprocess the historical user interaction data, and construct user behavior sequences according to interaction time to obtain the original user behavior sequences, as follows:

[0077] Historical interaction information of users is obtained from the target recommendation platform or public recommendation dataset. The historical interaction information includes user identifier, item identifier, interaction behavior type, interaction time, and item semantic information. The interaction behavior type includes at least one of click, browse, favorite, rating, and purchase. The item semantic information includes at least one of item name, category, tag, and text description.

[0078] The historical interaction information is cleaned, duplicate interactions are merged, and invalid interactions are filtered to obtain preprocessed user historical interaction records.

[0079] Data cleaning includes removing records with missing user or item identifiers, removing abnormal timestamp records, and filtering invalid interaction behaviors; duplicate interaction merging is used to merge multiple repeated interactions of the same user on the same item according to preset rules, so as to reduce the impact of noisy interactions on the subsequent modeling process.

[0080] Based on the preprocessed user history interaction records, the historical interaction items of each user are sorted according to the interaction time to obtain the original behavior sequence of user u, specifically:

[0081]

[0082] in, Indicates user The length of the historical interaction sequence, Indicates user In the Items that allow location-based interaction. ; This represents the original sequence of user u's actions.

[0083] The original user behavior sequence is used to represent the process of user interests changing over time, and serves as input for subsequent generation of enhanced user behavior sequences based on semantic-aware diffusion.

[0084] Simultaneously, the original user-item interaction edge set is preserved based on the historical interaction relationships between users and items. When there is a historical interaction between user u and item i, an interaction edge is established between user u and item i to obtain the original user-item interaction graph. Specifically:

[0085]

[0086]

[0087] in, Represents a set of nodes. This represents the set of interaction edges between a user and an item. Represents a set of users; Represents a collection of items.

[0088] For user u, its set of historical interaction edges Represented as:

[0089]

[0090] in, This represents the set of items that user u has interacted with in the past.

[0091] In this way, the user's historical behavior sequence, the interaction time sequence, the semantic information of the items, and the original user-item interaction edges together constitute the data-driven foundation for the subsequent enhancement process. Among them, the user behavior sequence is used to drive the generation of potential interest items, the original user-item interaction edges are used to drive the subsequent graph structure enhancement, and the semantic information of the items is used to support the semantic verification of candidate edges, thereby avoiding the subsequent enhancement process from relying entirely on random perturbations and improving the authenticity and reliability of the enhancement data source.

[0092] Step 2: Generate an enhanced user behavior sequence. Based on the frequency of item interactions, perform frequency-aware masking on the user behavior sequence. Then, use semantic-aware noise diffusion and context-aware denoising to perturb and restore the masked positions, generating the enhanced user behavior sequence. This process is as follows: Figure 2 As shown, the details are as follows:

[0093] Step 2.1: Construct a representation of user behavior sequences, as follows:

[0094] For the original behavior sequence of user u Using encoder The user's historical behavior sequence is encoded to obtain the context representation of each position, where the hidden representation of the i-th position is:

[0095]

[0096] in, The context representation represents the i-th position in the original behavior sequence of user u, which is used to characterize the relationship between the item at that position and the user's historical behavior context; the context representation serves as the sequence representation basis for subsequent frequency-aware masking, semantic-aware diffusion perturbation, and context-aware denoising processes.

[0097] The hidden state at the last position of the user behavior sequence is taken as the representation of the user's current preference, specifically:

[0098]

[0099] in, This represents the current preferences of user u. This represents the context representation of the Lth position in the original action sequence of user u.

[0100] Step 2.2: Perform frequency-aware masking, as follows:

[0101] Statistically calculate the interaction frequency of each item in the training set, and denote the frequency of item v as... For user behavior sequences The item corresponding to the i-th position The mask probability at a given position is determined based on its frequency of occurrence, specifically as follows:

[0102]

[0103] in, Indicates the first The probability that a position is masked. It is a function that monotonically decreases as the frequency of an item increases, so that the location corresponding to a low-frequency item has a higher mask probability.

[0104] The mask location set is obtained by sampling from the user behavior sequence based on the mask probability. Set the mask locations The positions in the text are replaced with special mask markers, and the unmasked positions are used as the context set. This yields a user behavior sequence after frequency-aware masking; in this way, long-tail items can obtain a stronger learning signal during sequence perturbation and subsequent denoising.

[0105] Step 2.3: Construct semantically aware noise and perform diffusion perturbation, as follows:

[0106] For the sequence representation to be perturbed Obtain its in the representation space Semantic neighbors, resulting in a set of semantic neighbors, specifically:

[0107]

[0108] in, Represents a semantic neighbor set; This represents the k-th semantic neighbor in the semantic neighbor set.

[0109] Based on the sequence to be disturbed The similarity between semantic neighbors is used to calculate the neighbor weights, specifically as follows:

[0110]

[0111] in, Represents the similarity function. Indicates temperature parameter, Indicates the first The weights of each semantic neighbor; The sequence to be perturbed is represented; This represents the k-th semantic neighbor in the semantic neighbor set; This represents the j-th semantic neighbor in the semantic neighbor set; This indicates the number of semantic neighbors selected.

[0112] The semantic neighbor center is calculated based on the neighbor weights, specifically as follows:

[0113]

[0114] in, This represents the weighted semantic neighbor center calculated based on each semantic neighbor and its corresponding weight; Indicates the first The weights of each semantic neighbor; This represents the k-th semantic neighbor in the semantic neighbor set.

[0115] A semantic direction is constructed based on the semantic neighbor center and the current representation, where the current representation is the sequence representation to be perturbed. Specifically:

[0116]

[0117] in, Indicates semantic direction; m represents the weighted semantic neighbor center calculated based on each semantic neighbor and its corresponding weight. This represents the sequence to be perturbed.

[0118] To avoid the impact of scale differences in semantic orientation between different samples on the diffusion process, the semantic orientation is normalized to obtain a normalized semantic orientation. Specifically:

[0119]

[0120] in, To prevent division by zero stability coefficient.

[0121] Standard Gaussian noise With normalized semantic direction Weighted fusion is performed to obtain semantically perceptual noise, specifically:

[0122]

[0123] in, This represents semantically perceptual noise; This represents the weights used to control standard Gaussian noise and semantic orientation in the diffusion perturbation, when When the noise is large, the semantic-aware noise is closer to standard Gaussian noise; when When the size is small, the diffusion perturbation is more biased towards the semantic neighborhood.

[0124] Let the noise intensity sequence of the diffusion process be... ,in, , Represents the total number of diffusion steps in the diffusion process, and is defined as follows:

[0125]

[0126] in, This represents the signal retention coefficient corresponding to the t-th diffusion step; Indicates the first The noise intensity of each diffusion step.

[0127] During the forward diffusion process, standard Gaussian noise is replaced with semantically perceptible noise, and the sequence representation at the masked position is perturbed by diffusion, specifically as follows:

[0128]

[0129] in, Indicates the first The noise intensity of each diffusion step. , Indicates the first The perturbation representation obtained from each diffusion step. Let represent the semantically perceived noise in the t-th diffusion step.

[0130] Furthermore, given a normalized semantic direction Under these conditions, the one-step forward diffusion transfer can be expressed in conditional Gaussian form:

[0131]

[0132] in:

[0133]

[0134]

[0135] in, ; Indicates a Gaussian distribution; This represents the signal retention coefficient corresponding to the t-th diffusion step; This represents the mean of the conditional Gaussian distribution corresponding to the t-th diffusion step, given the representation and semantic direction of the previous diffusion step. Represents the identity matrix that matches the dimension of the sequence representation; Let represent the covariance matrix of the conditional Gaussian distribution corresponding to the t-th diffusion step; Indicates the first The perturbation representation obtained from each diffusion step. Indicates the first The noise intensity of each diffusion step. This represents the weights used to control standard Gaussian noise and semantic orientation in the diffusion perturbation.

[0136] Through the aforementioned semantically aware noise diffusion process, sequence enhancement can introduce semantic neighborhood information between items while maintaining random perturbation capabilities, thereby reducing the destruction of semantic relationships caused by standard Gaussian noise.

[0137] Step 2.4: Perform context-aware denoising, as follows:

[0138] Let the set of mask locations be denoted as The set of context locations is denoted as During the forward diffusion process, only the set of mask locations is considered. Semantic-aware noise is added to the sequence representation to obtain the perturbed representation. and maintain the context location set Context representation in constant;

[0139] In the reverse denoising process, the representation of the unmasked positions is used as a context condition, and a conditional Gaussian distribution is used to recover the perturbed positions. The reverse denoising process is represented as follows:

[0140]

[0141] in, This represents the recovered representation of the mask position set M in the (t-1)th diffusion step. This represents the noise representation of the mask position set M in the t-th diffusion step. The representation of the context location set C. This represents the variance corresponding to the t-th diffusion step; Indicates by parameters Controlled inverse denoising conditional probability distribution; Indicates a Gaussian distribution; Indicates by parameters The mean of the conditional Gaussian distribution predicted by the context-aware denoising network under control; Indicates the current diffusion step of the sample; Represents the identity matrix that matches the dimension of the sequence representation.

[0142] Using context-aware denoising networks Prediction is based on the noise representation of the mask position set M in the t-th diffusion step. The context representation of the current diffusion step t and the context position set C. Predict the mean of the conditional Gaussian distribution for inverse denoising, specifically as follows:

[0143]

[0144] in, Indicates by parameters A context-aware denoising network for control; By modeling the sequence dependency between context locations and mask locations, the mask locations can utilize information from the context location set C during the recovery process.

[0145] During training, the output of the mean-matching target constraint denoising network is used to make the predicted mean close to the true forward and posterior mean. The true forward and posterior mean is expressed as:

[0146]

[0147] in, This represents the cumulative signal preservation coefficient from the first diffusion step to the t-th diffusion step. The original representation of the mask location set M; Let M be the noise representation of the mask position set M in the t-th diffusion step; Indicates the first The noise intensity of each diffusion step; This represents the mean of the true posterior distribution determined by the forward diffusion process.

[0148] The diffusion denoising loss is expressed as:

[0149]

[0150] in, Indicates the diffusion denoising loss; The original representation of the mask location set M; Let M be the noise representation of the mask position set M in the t-th diffusion step; The context representation of the context location set C; This indicates that sampling is performed and the expected value is calculated for different diffusion steps t; Indicates by parameters The mean of the conditional Gaussian distribution predicted by the context-aware denoising network under control; This represents the true posterior distribution mean determined by the forward diffusion process; This indicates the diffusion step of the current sample.

[0151] By minimizing the diffusion denoising loss, context-aware recovery is performed on the perturbed locations to generate enhanced user behavior sequences;

[0152] Enhanced user behavior sequence for user u Specifically:

[0153]

[0154] in, L represents the item corresponding to the i-th position in the enhanced user behavior sequence; L represents the user. The length of the historical interaction sequence.

[0155] Therefore, step 2 uses data-driven selection of the mask position based on the frequency of item interaction, and uses a semantic-aware diffusion model and a context-aware denoising network to perform model-driven recovery of the masked position.

[0156] Step 3: Generate candidate user-item interaction edges. Based on the potential interest items recovered from the enhanced user behavior sequence, form a set of candidate enhancement edges, as follows:

[0157] Step 3.1: Identify potential items of interest in the enhanced user behavior sequence, as follows:

[0158] For the original user behavior sequence and enhanced user behavior sequences Perform location correspondence analysis to identify items recovered from the enhanced user behavior sequence by semantic-aware diffusion perturbation and context-aware denoising processes, and denote them as the set of potential interest items for user u:

[0159]

[0160] in, This represents the set of potentially interesting items derived from enhanced user behavior sequences. Represents a collection of items.

[0161] Furthermore, to avoid repeatedly using historical interaction items already existing in the user's original behavior sequence as candidate enhancement items, the set of potential interest items can be limited to items added or replaced in the enhancement user behavior sequence, specifically:

[0162]

[0163] in, This represents the original sequence of user u's actions. This represents the enhanced user behavior sequence for user u. This represents the set of potentially interesting items derived from enhanced user behavior sequences. Represents a collection of items.

[0164] Step 3.2: Construct candidate user-item interaction edges, as follows:

[0165] For the set of potential interest items obtained in step 3.1 Match user u with a set of items of potential interest Candidate connections are established for item i in the data to form a set of candidate user-item interaction edges corresponding to user u, specifically:

[0166]

[0167] in, This represents the set of candidate enhancement edges corresponding to user u. Represents the candidate user-item interaction edge between user u and potential interest item i; This represents the set of potentially interesting items derived from the sequence of enhanced user behavior.

[0168] Based on the definition of the set of potential interest items in step 3.1, the set of candidate enhancement edges can also be represented as:

[0169] .

[0170] For all users, the candidate enhancement edge sets corresponding to each user are merged to obtain the global candidate enhancement edge set, specifically:

[0171]

[0172] in, Represents a set of users. This represents the set of candidate user-item interaction edges obtained by the enhanced user behavior sequence mapping; Let represent the set of candidate enhancement edges corresponding to user u.

[0173] Therefore, step 3, as a connecting link between data-driven enhancement and model-driven enhancement, transforms the potential interest items generated by user behavior data, item frequency distribution, and semantic perception diffusion model into candidate user-item interaction edges that can be verified by the large language model and utilized by the graph recommendation model.

[0174] Step 4: Generate an enhanced user-item interaction graph. Utilize a large language model to perform credibility verification and importance assessment on the original interaction edges and candidate enhanced edges. Based on the assessment results, discard differentiated edges and add potential edges, such as... Figure 3 As shown, the enhanced user-item interaction graph is obtained, as follows:

[0175] Step 4.1: Perform embedding enhancement based on the large language model, as follows:

[0176] Using a large language model, semantic embeddings of items are generated based on their textual descriptions, names, categories, and tag information. Its semantic embedding is denoted as And integrate all item semantic embeddings into an item embedding matrix A;

[0177] Since the number of items is usually large, directly using high-dimensional semantic embeddings would increase the computational complexity. Therefore, principal component analysis is performed on the item embedding matrix A to reduce its dimensionality, preserving the main semantic information while reducing the embedding dimension, resulting in a low-dimensional item semantic embedding matrix, as follows:

[0178]

[0179] in, The item embedding matrix is ​​the result of mean-centered optimization. For the reason before The projection matrix composed of eigenvectors; This is a low-dimensional semantic embedding matrix for items.

[0180] For users Based on its historical collection of interacting items The initial semantic embedding of the user is obtained by aggregating the low-dimensional semantic embeddings of the items, specifically:

[0181]

[0182] in, This represents the set of items that user u has interacted with in the past. This indicates the number of items that user u has interacted with in the past. Represents items semantic embedding, Indicates user Semantic embedding. This embedding enhancement method introduces the semantic knowledge contained in the large language model into the initial representations of users and items, thereby enhancing the representation learning ability of the subsequent graph recommendation model.

[0183] Step 4.2: Perform semantic verification and importance assessment of interaction edges based on the large language model, as detailed below:

[0184] The system converts user basic attributes, user historical interaction information, candidate enhancement edge set, and item semantic information into text representations, and constructs prompt words suitable for semantic verification and importance judgment of interaction edges. The prompt words are then input into a large language model, which evaluates the semantic importance of the original interaction edges and candidate enhancement edges based on user preference patterns, item semantic relevance, and the generation source of candidate enhancement edges.

[0185] For any interaction edge e to be evaluated, the large language model outputs the corresponding semantic verification result and interaction importance label, specifically as follows:

[0186]

[0187] in, This represents the semantic verification result of interaction edge e, used to characterize whether the interaction edge is consistent with the user's historical interests and the semantic information of the items. The label indicating the importance of the interaction edge e This represents the evaluation process of interaction edges based on a large language model. This indicates a prompt message constructed for user u, which includes the user's basic attributes, historical interaction information, and semantic information of the item.

[0188] Based on the interaction importance labels output by the large language model, the historical interaction edges of user u are divided into a high-quality interaction set. Medium-quality interaction set and low-quality interaction sets ;

[0189] Among them, high-quality interactive collection This refers to a set of medium-quality interactions that reflect core user preferences. This represents a set of low-quality interactions that have some degree of preference relevance. This indicates that the interaction may contain noise or be weakly relevant to the user's interests.

[0190] Step 4.3: Perform differential edge discarding, as follows:

[0191] Graph augmentation based on large language models includes edge dropping strategies and edge adding strategies based on large language models.

[0192] In the edge-dropping strategy, user basic attributes, user historical interaction information, and item semantic information are input into a large language model. The large language model determines the importance of user historical interaction edges and divides the interaction edges into high-quality interaction sets. Medium-quality interaction set and low-quality interaction sets .

[0193] Different edge drop rates are set based on the importance category of the interaction edges for the original user-item interaction graph. The original interaction edges in the data are discarded under a differential edge disposal operation. For the historical interaction edge set of user u... The set of interaction edges that are retained after being discarded is represented as:

[0194]

[0195] in, This represents the set of historical interaction edges for user u. This represents the set of interactive edges that are retained after an edge is discarded. Represents a random sampling function; Let e ​​represent the drop rate of the interaction edge e, and:

[0196]

[0197] in, Let represent the edge drop rates corresponding to high-quality, medium-quality, and low-quality interactions, respectively, and satisfy the following conditions:

[0198]

[0199] This results in a lower probability of discarding high-quality interaction edges and a higher probability of discarding low-quality interaction edges, thereby reducing the damage to key interaction structures caused by random edge discarding.

[0200] Step 4.4: Perform candidate edge semantic verification and potential edge addition, as detailed below:

[0201] In the edge addition strategy, for the candidate augmentation edge set generated in step 3 The historical interaction information of user u, the candidate enhancement edges, and the semantic information of the candidate items are input into the big language model, and the big language model determines whether the candidate enhancement edges are consistent with the user's historical interests and the semantic information of the items.

[0202] The candidate enhanced edges that pass semantic verification will be combined into a set of potential relevant edges, specifically:

[0203]

[0204] in, This indicates that the large language model determines whether the candidate augmentation edge between user u and item i passes semantic verification; This represents the set of candidate augmented edges corresponding to user u that have passed the semantic verification of the large language model; Let represent the set of candidate augmentation edges corresponding to user u; add the candidate augmentation edges that pass semantic verification to the user-item interaction graph to obtain the augmentation edge set:

[0205]

[0206] in, This represents the set of enhanced interaction edges corresponding to user u; This represents the set of interaction edges retained after discarding the differentiated edges from the original historical interaction edges of user u. This represents the set of candidate augmented edges corresponding to user u that have passed the semantic verification of the large language model.

[0207] The set of enhanced edges for all users is represented as:

[0208]

[0209] in, This represents the enhanced interaction edge set obtained after discarding differentiated edges and adding candidate edges; This represents the set of enhanced interaction edges corresponding to user u; This represents a set of users.

[0210] The final result is an enhanced user-item interaction graph. Specifically:

[0211]

[0212]

[0213] in, Represents a set of nodes. This represents the enhanced interaction edge set obtained after discarding differentiated edges and adding candidate edges; Represents a set of users; Represents a collection of items.

[0214] In this way, the candidate interest items generated by semantic perception diffusion are mapped as candidate enhancement edges, and after semantic verification by a large language model, they participate in the user-item interaction graph enhancement, thereby improving the credibility of the graph structure enhancement process.

[0215] Step 5: Combine graph contrast learning, the main recommendation task, and diffusion denoising constraints to optimize the recommendation model, predict the target user's preference score for candidate items, and generate a recommendation list based on the preference score, as follows:

[0216] Step 5.1: Construct an enhanced graph view based on the enhanced user-item interaction graph, as follows:

[0217] Based on the enhanced user-item interaction graph obtained in step 4 And further construct two different graph augmentation views. and ; Use a graph encoder to learn representations for the two graph augmented views to obtain the embedded representations of users and items under different augmented views;

[0218] in, and This represents two enhanced views obtained from the enhanced user-item interaction graph through different graph perturbations or different sampling methods, which are used for subsequent graph comparison learning tasks.

[0219] Step 5.2: Perform graph comparison learning based on graph-enhanced views, as detailed below:

[0220] For the embedding representation of the same node under different augmented views, it is treated as a positive sample pair; for the embedding representation of different nodes under different augmented views, it is treated as a negative sample pair; through graph contrastive learning, the representation of the same user or the same item is kept consistent under different augmented views, and the representations of different users or different items are kept distinguishable.

[0221] User side graph comparison learning loss Represented as:

[0222]

[0223] in, Representing users respectively Enhanced view in the figure and Embedded representation in Indicates user Embedded representation in another enhanced view of the graph Represents the similarity function. Indicates temperature parameter; This represents a set of users.

[0224] Similarly, the learning loss for item side image comparison. Represented as:

[0225]

[0226] in, Representing items Enhanced view in the figure and Embedded representation in Represents items Embedded representation in another enhanced view; Represents a collection of items; Represents the similarity function. This represents the temperature parameter.

[0227] Overall comparison learning loss Represented as:

[0228]

[0229] By learning constraints through graph contrast, user and item representations are kept consistent across different augmented views, reducing the impact of single graph perturbations or local noise interactions on recommendation results, thereby improving the stability and reliability of recommendation representations.

[0230] Step 5.3: Perform inter-layer comparative learning based on the representations of different layers of the graph neural network, as detailed below:

[0231] Utilizing graph neural networks to perform multi-layer message propagation on the user-item interaction graph and obtain nodes. In the Layer and first The embedding representation of the layers is denoted as follows: and ; where node v can represent a user node or an item node;

[0232] Consistency constraints are applied to the representation of the same node in adjacent layers to reduce information decay and oversmoothing problems during multi-layer graph propagation.

[0233] Inter-layer contrastive learning loss Represented as:

[0234]

[0235] in, Represents a node In the Layer embedding representation, Indicates that node v is at the th Layer embedding representation, Indicates that node k is at the 1st rank. Layer embedding representation; Represents a set of nodes; Represents the similarity function. This represents the temperature parameter.

[0236] By using inter-layer contrastive learning constraints, the graph neural network retains structural information at different levels during propagation, thereby improving the stability and discriminative ability of user and item representations.

[0237] Step 5.4: Optimize the recommendation model based on the recommendation task and the contrastive learning task, as follows:

[0238] Calculate users based on user embedding and item embedding. For items Preference score:

[0239]

[0240] in, This represents the final embedded representation of user u. The final embedded representation of item i;

[0241] Bayesian personalized ranking loss is used as the main loss for recommendation. :

[0242]

[0243] in, Let represent a training sample triple, where i represents a positive sample item that user u has interacted with, and j represents a negative sample item that user u has not interacted with. Represents the Sigmoid function; This represents the set of sample triples used for training the main recommendation task; This represents the preference score of user u for positive sample item i, as predicted by the recommendation model; This represents the preference score of user u for negative sample item j predicted by the recommendation model.

[0244] The model parameters are jointly optimized using recommendation main task loss, graph contrastive learning loss, inter-layer contrastive learning loss, and regularization term. The total loss function is... Represented as:

[0245]

[0246] in, This indicates the recommended main task loss. This represents the learning loss based on the graph contrast. Indicates interlayer contrast loss. This represents the diffusion denoising loss of the semantic-aware diffusion module. , and Let represent the weight coefficients of the graph contrast learning loss, inter-layer contrast learning loss, diffusion denoising loss, and regularization term, respectively. This represents the set of model parameters.

[0247] After training, the optimized recommendation model is used to calculate the target user's preference score for candidate items, and the candidate items are sorted according to the preference score. The top K items with the highest scores are selected to generate a recommendation list.

[0248] In this way, the generation of the recommendation list not only depends on the prediction score between the user and the item embedding, but is also affected by the combined effects of candidate edge credibility verification, enhanced graph structure constraints, graph comparison consistency constraints, and diffusion denoising constraints. This reduces the recommendation bias caused by noisy interactions, unreasonable candidate edges, or random enhancement perturbations, and improves the stability and credibility of the recommendation results.

[0249] like Figure 2 As shown, the generation process of enhanced user behavior sequences mainly includes three stages: frequency-aware masking, semantic-aware noise construction, and context-aware denoising. First, the masking probability corresponding to each position in the user behavior sequence is determined based on the interaction frequency of different items in the training data. Items with lower interaction frequencies have higher masking probabilities, thereby increasing the chances of long-tail items participating in perturbation and recovery learning. Then, semantic-aware noise is constructed to perform diffusion perturbation on the masked positions, ensuring that the perturbation process is guided by the semantic relationships of the items while retaining randomness. Finally, in the reverse denoising process, the context representation of the unmasked positions is kept unchanged and used as conditional information to recover the perturbed positions, thus generating an enhanced user behavior sequence that maintains semantic consistency with the original user interests.

[0250] like Figure 3 As shown, the interaction edge enhancement process based on a large language model mainly includes two aspects: edge discarding and edge enhancement. In the edge discarding process, user historical interaction information and the semantic information of corresponding items are organized into prompts and input into the large language model. The large language model combines user historical interests and item semantic features to perform relevance analysis and importance judgment on existing interaction relationships, assigning different discard probabilities to different edges. In the edge enhancement process, user historical interaction information, user attribute information, and candidate item information corresponding to the candidate enhancement edges generated in step 3 are input into the large language model. The large language model judges the semantic consistency between candidate items and user historical interests, and adds candidate enhancement edges that pass semantic verification as potential new interaction relationships to the user-item interaction graph. Through these two strategies, on the one hand, original interaction relationships are selectively filtered and retained; on the other hand, potential interest relationships are semantically verified and supplemented, thereby achieving reliable graph structure enhancement based on user interests and item semantic information.

[0251] In this embodiment, the MovieLens movie recommendation dataset is used as an example. This dataset includes user identifiers, movie identifiers, user rating behavior for movies, interaction timestamps, and semantic information such as movie names and genres. For ease of explanation, this embodiment selects a subset of users and movies to construct illustrative examples. Let the user set be: The collection of movie items includes: .

[0252] The semantic information of each movie item is shown in the table below:

[0253] Table 1 Semantic Information of Various Movie Items

[0254]

[0255] Based on user rating records and interaction timestamps, the historical interaction sequences of some users are shown in the table below:

[0256] Table 2 Historical interaction sequences of some users

[0257]

[0258] Based on the above interaction records, a user-item interaction graph can be constructed. ,in:

[0259]

[0260]

[0261] At the same time, user behavior sequences can also be constructed based on interaction time.

[0262] The following is based on the user Taking an example, this invention illustrates how to generate potential interest items through user behavior sequences, map them as candidate user-item interaction edges, and then use a large language model to perform semantic verification of candidate edges and graph structure enhancement.

[0263] user The original sequence of behaviors is:

[0264]

[0265] in, "The Silence of the Lambs" belongs to the drama / thriller genre. 2001: A Space Odyssey belongs to the science fiction and suspense genre. Apollo 13 belongs to the drama and adventure genre.

[0266] Statistical analysis was performed based on the frequency of movies in the training set. To simplify the implementation, the statistics were conducted on a subset of the dataset. The frequency of some movies is shown in the table below:

[0267] Table 3. Frequency of appearance in some movies

[0268]

[0269] According to the frequency-aware masking strategy, the lower the frequency of an item's occurrence, the higher the probability that its corresponding position in the user's behavior sequence will be masked. For users... sequence ,because The occurrence frequency of is low, therefore its corresponding position has a high mask probability.

[0270] Based on the mask sampling results, the frequency-aware masked sequence is obtained:

[0271]

[0272] in, Indicates user The sequence after masking; This indicates the masked item.

[0273] The representation corresponding to the masked position It retrieves its semantic neighbors in the representation space. In this embodiment, The semantic neighbors of "2001: A Space Odyssey" include science fiction or space-themed films, thus yielding a semantic neighbor set:

[0274]

[0275] in, Star Wars: Episode IV – A New Hope The Matrix and All of *Terminator 2: Judgment Day* are related to science fiction, action, or space themes. Semantic neighbor centers and semantic directions are constructed based on semantic neighbor sets, and standard Gaussian noise is weighted and fused with normalized semantic directions to obtain semantic perceptual noise. Thus, the semantic perception diffusion process no longer relies solely on random Gaussian perturbations, but is driven by both the semantic neighborhood of the item and the diffusion model, enabling the enhanced user behavior sequence to maintain semantic consistency while generating items of potential interest.

[0276] Then, semantically aware noise is used to perturb the masked positions during the forward diffusion process, and unmasked positions are utilized during the reverse diffusion process. and The following text indicates that context-aware denoising is performed to restore the representation of the perturbed position and generate an enhanced user behavior sequence. In this embodiment, after semantic-aware diffusion perturbation and context-aware denoising, the user behavior sequence can be obtained. Enhanced user behavior sequence :

[0277]

[0278] in, Terminator 2: Judgment Day is a potential interest item obtained through semantic perceptual diffusion and contextual denoising processes. This film is related to the original sequence. All of "2001: A Space Odyssey" have a science fiction semantic relevance, and are also related to users' historical behavior in terms of plot, adventure, and science fiction interests.

[0279] Next, the enhanced user behavior sequences are mapped to candidate user-item interaction edges. Because users... The original behavioral sequence is Enhance user behavior sequence ,in, To enhance the potential interest items obtained by replacing or generating in the sequence, and Therefore, the user With movies The connections between them are used as candidate user-item interaction edges, resulting in a set of candidate enhancement edges:

[0280]

[0281] in, For users The set of candidate augmenting edges.

[0282] Therefore, a connection was established between the data-driven sequence augmentation results and the model-driven graph structure verification process. Subsequently, a large language model was used to evaluate the semantic importance of the original interaction edges and candidate augmentation edges. For users... Its original set of interaction edges is:

[0283]

[0284] User The historical interactive movies, candidate enhancement edges, movie names, movie genres, and movie synopses are input into the large language model. The large language model determines the importance of each interactive edge based on user preference patterns and semantic relevance of the movies.

[0285] In this embodiment, the interaction importance judgment results output by the large language model are shown in the table below:

[0286] Table 4. Results of Interaction Importance Judgment from the Large Language Model

[0287]

[0288] in, Represents a high-quality collection of interactions. Represents a set of interactions of medium quality. This represents a low-quality set of interactions.

[0289] In this embodiment, the discard rate for different types of interaction edges is set as follows:

[0290]

[0291] According to the differentiated edge discarding strategy, high-quality interaction edges have a low discard probability, medium-quality interaction edges have a moderate discard probability, and low-quality interaction edges have a high discard probability. After one sampling, let the user... The set of retained interaction edges for:

[0292]

[0293] For candidate enhanced edges Because the large language model determines its relationship with the user The historical interests and semantic information of the film are consistent, so it is used as a candidate enhancement edge that passes the credibility verification, resulting in a set of potential related edges. :

[0294]

[0295] This leads to users Enhanced edge set :

[0296]

[0297] Right now:

[0298]

[0299] Incorporating the aforementioned enhanced edge set into the overall user-item interaction graph yields the enhanced user-item interaction graph. :

[0300]

[0301] in, This represents the set of interactive edges after discarding differentiated edges and adding candidate edges.

[0302] Based on the enhanced user-item interaction graph Construct two different graph augmentation views and The model learns the embedding representations of users and movies under different augmented views using a graph encoder. For the same node's embedding representations under different augmented views, they are treated as positive sample pairs; for different nodes' embedding representations under different augmented views, they are treated as negative sample pairs. A graph contrastive learning loss is used to constrain the consistency of representations across different views. Simultaneously, the model is optimized by combining the diffusion denoising loss from step 2 and the recommendation main task loss, enabling the model to utilize both the latent interest information generated by semantically aware diffusion and the structured interaction relationships in the augmented user-item graph.

[0303] After training, the recommendation model will recommend to users. Preference scores are predicted for candidate movies.

[0304] In this embodiment, the prediction results are shown in the following table:

[0305] Table 5 Prediction Results

[0306]

[0307] Therefore, based on the data model-driven semantic enhancement method for recommendation system interaction graphs of this invention, a user-oriented recommendation system is obtained. The recommended list is as follows:

[0308]

[0309] In this embodiment, user historical behavior sequences, movie appearance frequency, and user-movie interaction relationships provide data-driven basis for potential interest generation and interaction graph construction. The semantic-aware diffusion model is used to generate enhanced user behavior sequences, and the large language model is used to perform credible semantic verification on candidate enhanced edges and evaluate the importance of original interaction edges, thereby forming model-driven constraints. Finally, graph comparison learning and recommendation task optimization are performed based on the enhanced user-item interaction graph.

[0310] In this embodiment, taking the long-tail product recommendation scenario in an e-commerce platform as an example, the application of the data model-driven enhanced trusted recommendation method of the present invention in recommendation scenarios with obvious long-tail distribution and rich product semantic information is further explained.

[0311] User historical behavior data in e-commerce recommendation platforms includes user identifiers, product identifiers, browsing, favorites, adding to cart or purchasing interactions, interaction times, and product semantic information such as product name, product category, brand, tags, and text descriptions. Because the frequency of user interaction varies significantly among different products—some popular products have a large number of user interactions, while many long-tail products have only a few interaction records—the frequency-aware masking and semantic-aware enhancement method of this invention is suitable.

[0312] For ease of explanation, this embodiment denotes the user as u, and uses some of their historical interaction behaviors for illustration. Assume the product set includes the following products:

[0313] Table 6 Semantic Information of Some Products

[0314]

[0315] Based on the user's browsing, favorites, and purchase records on the e-commerce platform, the user's historical behavior sequence is obtained by sorting the interaction time:

[0316]

[0317] Among them, user u's main historical behavior focused on skin care products such as cleansing, moisturizing and hydrating, and also included one interaction with a perfume product that was relatively less related to its main skin care interests.

[0318] The interaction frequency of each product in the statistical training data is shown in the table below, with the frequency of some products appearing as shown in the table below.

[0319] Table 7. Interaction frequency of some products

[0320]

[0321] According to the frequency-aware masking strategy of this invention, the lower the interaction frequency of items, the higher the masking probability of their corresponding sequence positions. Because... The interaction frequency is relatively low, therefore, in this mask sampling process, its corresponding position is selected as the position to be recovered, resulting in the user behavior sequence after frequency-aware masking:

[0322]

[0323] For the product representation corresponding to the masked location, its semantic neighbors are obtained in the product representation space. Based on semantic information such as product name, category, tags, and text description, semantic neighbors related to the topic of hydration and moisturizing can be obtained, such as products like hydrating serums, moisturizing creams, and body lotions.

[0324] The semantic neighbor center and semantic direction are calculated based on semantic neighbors. The normalized semantic direction is then weighted and fused with standard Gaussian noise to form semantically perceptual noise. Subsequently, the masked positions are perturbed during the forward diffusion process, and conditional recovery is performed during the reverse denoising process using contextual information from unmasked positions such as facial cleanser, wood fragrance, and moisturizing cream.

[0325] After semantic-aware diffusion and context-aware denoising, the enhanced user behavior sequence for user u is obtained:

[0326]

[0327] in, This product, a hydrating essence, is a potential interest item identified through an enhancement process. Its semantic attributes, such as hydration and moisturizing, show a high relevance to hydrating masks and moisturizing creams in the user's historical interactions.

[0328] Due to the goods Items that do not belong to user u's original historical interaction items are therefore identified as potential interest items for user u, and candidate user-item interaction edges are constructed, thus obtaining the candidate enhancement edge set corresponding to user u:

[0329]

[0330] Subsequently, a large language model is used to perform semantic verification and importance evaluation on the interaction edges to be processed related to user u. The interaction edges to be processed include the user's original historical interaction edges and candidate enhancement edges generated by enhancing user behavior sequences.

[0331] The historical interaction information of user u1, product name, product category, product tag, product text description, and candidate enhancement edge information are used to construct prompt information and input into the large language model. The large language model evaluates each interaction edge based on the user's historical interest pattern and the semantic relationship between products.

[0332] In this embodiment, the evaluation results of the large language model for some interaction edges are shown in the table below.

[0333] Table 8. Results of Interaction Importance Judgment from Large Language Models

[0334]

[0335] Based on the above evaluation results, differentiated edge dropping is implemented for interaction edges that need to be dropped. High-importance interaction edges have a lower drop probability, while low-importance interaction edges have a higher drop probability.

[0336] In this random sampling process, let user u and product u be... If low-importance interaction edges are discarded, while the remaining high-importance interaction edges are retained, then the result after discarding the edges is:

[0337]

[0338] For candidate enhanced edges Because it passes the semantic verification of the large language model and has a high consistency with the user's interest in hydration, moisturizing and skin care reflected in the user's historical behavior, this candidate enhancement edge is added to the enhanced user-item interaction graph as a potential new interaction edge.

[0339] This yields the set of enhanced interaction edges for user u:

[0340]

[0341] Through the above process, on the one hand, interaction edges in the original interaction graph that are weakly related to the user's main interests are likely to be discarded; on the other hand, potential interest items generated by frequency-aware masking, semantic-aware diffusion, and context-aware denoising are added as new user-item interaction relationships after semantic verification by a large language model, thereby enhancing the credibility of the original user-item interaction graph.

[0342] Furthermore, based on the obtained enhanced user-item interaction graph, different graph augmentation views are constructed, and graph contrastive learning, inter-layer contrastive learning, and joint optimization of the recommendation task are performed according to the method in step 5 above. After training, the preference score of user u for candidate items is calculated using the recommendation model.

[0343] The prediction results for some candidate products are shown in the table below.

[0344] Table 9. Partial Product Forecast Results for User u

[0345]

[0346] By sorting the candidate products according to their predicted scores, a product recommendation list can be generated for user u.

[0347] In this embodiment, user history behavior, product interaction frequency, and user-product interaction relationship in e-commerce product interaction data constitute the data-driven foundation. Frequency-aware masking enables products with low interaction frequency to obtain more opportunities for enhancement and recovery. Semantic-aware diffusion and context-aware denoising can generate potential interest products that are semantically related to users' historical interests. The large language model further performs unified semantic verification and importance evaluation on the original interaction edges and candidate enhancement edges, and performs differentiated edge discarding and potential edge addition accordingly.

[0348] Therefore, this invention combines data-driven candidate interest generation with model-driven candidate edge reliability verification to transform potential interest items in augmented user behavior sequences into reliable user-item interaction edges. Finally, the recommendation task is optimized based on the augmented user-item interaction graph, which improves the reliability of augmented data and the credibility of recommendation results.

[0349] This invention provides data-driven support through user historical interaction data, item frequency information, and user-item interaction relationships. It provides model-driven constraints through diffusion models and large language models, achieving collaborative optimization of candidate interest generation, candidate interaction edge verification, and interaction graph enhancement. Simultaneously, by evaluating the semantic importance of interaction edges, verifying the credibility of candidate edges, and updating differentiated edges, it reduces semantic shifts caused by random sequence perturbations and the damage to key interaction structures caused by random graph enhancement. This alleviates the problems of insufficient learning of long-tail items and uncontrollable quality of candidate enhancement edges, improving the reliability of enhanced data, the representation learning ability of recommendation models, and the credibility of recommendation results.

[0350] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A reliable recommendation method based on data model-driven enhancement, characterized in that: Includes the following steps: Step 1: Obtain historical user interaction data from the recommendation system, preprocess the historical user interaction data, construct user behavior sequences according to interaction time to obtain the original user behavior sequences, and retain the historical interaction relationships between users and items. Step 2: Perform frequency-aware masking on the user behavior sequence based on the frequency of item interaction, and use semantic-aware noise diffusion and context-aware denoising to perturb and restore the masked position to generate an enhanced user behavior sequence. Step 3: Generate candidate user-item interaction edges. Based on the potential interest items recovered from the enhanced user behavior sequence, form a set of candidate enhancement edges. Step 4: Use a large language model to perform credibility verification and importance evaluation on the original interaction edges and candidate enhancement edges, and perform differential edge discarding and potential edge addition based on the evaluation results to obtain the enhanced user-item interaction graph; Step 5: Combine graph comparison learning, recommendation main task and diffusion denoising constraint to optimize the recommendation model, predict the target user's preference score for candidate items, and generate a recommendation list based on the preference score.

2. The trusted recommendation method based on data model-driven enhancement according to claim 1, characterized in that, In step 1, the specific implementation details are as follows: Obtain users' historical interaction information from target recommendation platforms or public recommendation datasets. The historical interaction information includes user identifier, item identifier, interaction behavior type, interaction time, and item semantic information. The historical interaction information is cleaned, duplicate interactions are merged, and invalid interactions are filtered to obtain the preprocessed user historical interaction records. Based on the preprocessed user history interaction records, the historical interaction items of each user are sorted according to the interaction time to obtain the user's... The behavioral sequence is as follows: ； in, Indicates user The length of the historical interaction sequence; Indicates user In the Items that can be interacted with at a specific location; This represents the original sequence of user u's actions; Simultaneously, the original user-item interaction edge set is preserved based on the historical interaction relationships between users and items; when there is a historical interaction between user u and item i, an interaction edge is established between user u and item i to obtain the original user-item interaction graph. ; For user u, its set of historical interaction edges Represented as: ； in, This represents the set of items that user u has interacted with in the past.

3. The trusted recommendation method based on data model-driven enhancement according to claim 1, characterized in that, Step 2 includes the following specific implementation details: Step 2.1: Construct a representation of user behavior sequences; Step 2.2: Perform frequency-aware masking; Step 2.3: Construct semantically aware noise and perform diffusion perturbation; Step 2.4: Perform context-aware denoising.

4. The trusted recommendation method based on data model-driven enhancement according to claim 3, characterized in that, Step 2.1 specifically involves the following implementation details: For the original behavior sequence of user u Using an encoder The user's original behavior sequence is encoded to obtain the context representation of each position, where the hidden representation of the i-th position is: ； in, The context representation of the i-th position in the original action sequence of user u; Then, the hidden state at the last position of the user behavior sequence is taken as the representation of the user's current preference, specifically: ； in, This represents the current preferences of user u. This represents the context representation of the Lth position in the original action sequence of user u.

5. The trusted recommendation method based on data model-driven enhancement according to claim 1, characterized in that, Step 3 includes the following specific implementation details: Step 3.1: Identify potential interest items in the enhanced user behavior sequence; Step 3.2: Construct candidate user-item interaction edges.

6. The trusted recommendation method based on data model-driven enhancement according to claim 5, characterized in that, Step 3.2 specifically involves the following implementation details: For the set of potential interest items obtained in step 3.1 Match user u with a set of items of potential interest Candidate connections are established for item i in the data to form a set of candidate user-item interaction edges corresponding to user u, specifically: ； in, This represents the set of candidate enhancement edges corresponding to user u. Represents the candidate user-item interaction edge between user u and potential interest item i; This represents the set of potentially interesting items derived from the sequence of enhanced user behavior. For all users, the candidate enhancement edge sets corresponding to each user are merged to obtain the global candidate enhancement edge set, specifically: ； in, Represents a set of users. This represents the set of candidate user-item interaction edges obtained by the enhanced user behavior sequence mapping; Let represent the set of candidate enhancement edges corresponding to user u.

7. The trusted recommendation method based on data model-driven enhancement according to claim 1, characterized in that, Step 4 includes the following specific implementation details: Step 4.1: Perform embedding enhancement based on the large language model; Step 4.2: Perform semantic verification and importance assessment of interaction edges based on the large language model; Step 4.3: Perform differential edge discarding; Step 4.4: Perform semantic verification of candidate edges and add potential edges.

8. The trusted recommendation method based on data model-driven enhancement according to claim 7, characterized in that, Step 4.1 specifically involves the following implementation details: Using a large language model, semantic embeddings of items are generated based on their textual descriptions, names, categories, and tag information. Its semantic embedding is denoted as And integrate all item semantic embeddings into an item embedding matrix A; Principal component analysis (PCA) is performed on the item embedding matrix A to reduce its dimensionality while preserving the main semantic information, resulting in a low-dimensional item semantic embedding matrix, as follows: ； in, The item embedding matrix is ​​the result of mean-centered optimization. For the reason before The projection matrix composed of eigenvectors; For low-dimensional item semantic embedding matrix; For users Based on its historical collection of interacting items The initial semantic embedding of the user is obtained by aggregating the low-dimensional semantic embeddings of the items, specifically: ； in, This represents the set of items that user u has interacted with historically. This indicates the number of items that user u has interacted with in the past. Represents items semantic embedding, Indicates user Semantic embedding.

9. The trusted recommendation method based on data model-driven enhancement according to claim 1, characterized in that, Step 5 includes the following specific implementation details: Step 5.1: Construct an enhanced graph view based on the enhanced user-item interaction graph; Step 5.2: Perform graph comparison learning based on the enhanced graph view; Step 5.3: Perform inter-layer comparison learning based on the representations of different layers of the graph neural network; Step 5.4: Optimize the recommendation model based on the recommendation task and the contrastive learning task.

Images