A session recommendation denoising method based on mixed prototype internal and external flow network

Abstract

The application discloses a conversation recommendation denoising method based on an internal and external flow network of a hybrid prototype, and belongs to the field of computer science.A new hybrid prototype enhanced conversation recommendation framework is provided, including a hybrid prototype learning module, which learns the hybrid prototype representation of items from all conversations through semantic clustering and topological clustering, so as to capture the global semantics and interaction patterns between items.A global-local enhancement module is also provided to integrate global prototype information into local conversation representation, so as to enrich the intention information of the local conversation.In addition, in order to enhance the understanding of the item neighborhood of the model, a neighborhood enhancement task is introduced, which maximizes the mutual information between the neighborhoods through contrast learning, so as to improve the recommendation performance.The method can generate better conversation representation, so as to realize more accurate recommendation.

Description

Technical Field

[0001] This invention belongs to the field of computer science, and in particular relates to a method for denoising session recommendation in a hybrid prototype-based inbound and outbound flow network. Background Technology

[0002] In the information age, the ever-increasing flow of information has made people's lives more convenient, but it has also brought about the problem of information overload. The sheer volume of information often exceeds the limits of what individuals or systems can accept and process. Therefore, how to effectively filter out the information that users truly need from this massive amount of data has always been a hot topic of discussion in academia and industry. Based on this, recommender systems have emerged. They are mainly based on modeling user profiles and historical preferences to recommend information that is truly valuable and of interest to users. Currently, recommender systems are widely used in e-commerce, news feeds, streaming media platforms, and other fields.

[0003] However, in real-world scenarios, users are often anonymous (e.g., unlogged-in guest users) and may only generate a limited amount of interaction information on the platform. To address these issues, the concept of conversational recommendation has been proposed. A major challenge in building a recommendation system lies in effectively modeling the complex relationships between individual items within a conversation. Past work has often focused on building different model architectures to better capture the complex dependencies between items. Broadly speaking, these can be categorized into sequence-based and graph-based methods. Sequence-based methods include those based on recurrent neural networks (RNNs) and attention mechanisms, while graph-based methods include those based on single-channel and two-channel graphs. These methods have made significant contributions to extracting better item features, but they often only consider individual items as the basic unit of consideration, limiting the amount of information the model can utilize. Recently, researchers have discovered that items with high dependencies often exhibit similar information, and considering them together may be able to model general user behavior patterns. These behavioral patterns essentially reflect the user's intent. By mining user intent information from conversational data, recommendation models can gain two advantages: accuracy and generalization. Based on these advantages, many recent studies have focused on exploring how intent information helps in conversational recommendation.

[0004] The ignorance of interactive noise in the data by intent-based mining methods hinders further improvements in recommendation performance. After recognizing the impact of noise on recommendation system performance, numerous studies have introduced denoising modules into recommendation systems. Although existing work has made progress in eliminating the influence of noise items, their strategies for identifying noisy items are based solely on information contained in a single session, which may pose a challenge in short-session-dominated scenarios. Furthermore, they primarily focus on the outflow of irrelevant information within each session, ignoring the inflow of relevant information, which limits the available information in the session-based recommendation model. Summary of the Invention

[0005] This invention proposes a session recommendation denoising method based on a hybrid prototype-based in-and-out flow network (HyPro). It mainly includes the following three core technical points:

[0006] Part 1: Hybrid Prototype Learning Module

[0007] Previous work has largely focused on acquiring semantic representations of items, considering relationships between items only within the current session. However, global information is also crucial, limiting the amount of information a model can capture. Therefore, we propose a technique called hybrid clustering. First, semantic clustering is used to obtain the semantic prototype of each item. Then, a global graph is constructed, and topological clustering is performed to obtain the topological prototype of each item. The representation of the topological prototype is obtained by average pooling the representations of the semantic prototypes of the items belonging to it. Thus, the final topological prototype representation combines semantic and topological information, enriching the information content of the intent representation.

[0008] II. Information Outflow Module

[0009] To avoid the negative impact of irrelevant information on the representation learning process, we designed an information outflow channel consisting of two stages of denoising: data-level denoising and feature-level denoising. Data-level denoising is implemented using a Prototype Denoising Module (PDM), which defines a representation boundary for each session and filters out items exceeding this boundary. Feature-level denoising is implemented using a Sparse Edge Graph Neural Network (SEG), which operates on a local graph and dynamically learns the importance of connections between item pairs using an attention mechanism. Finally, we introduce... The function rescales the similarity scores between items to a sparse feature space to further eliminate the influence of irrelevant neighbor embeddings on the central node during feature aggregation.

[0010] III. Information Inflow Module

[0011] To avoid excessively sparse session information due to denoising in short-session-dominated scenarios, we introduce an information inflow module, setting up a Global Enhancement Module (GEM) and a Domain Enhancement Module (NET). The GEM refines each item in the local graph by using the difference between the item representation and its hybrid prototype representation, enabling the model to capture interaction features at different levels and thus enhancing its ability to model complex relationships between items. In the Domain Enhancement Module, we use cosine similarity to find the neighborhood of each session. Then, we use a contrastive learning task to maximize mutual information between domains. To prevent the influence of noisy sessions, we introduce a penalty term to reduce the impact of neighborhood sets with large internal differences. This module expands the information domain that can be used to learn individual representations at the session level.

[0012] Thanks to these three design features, HyPro has achieved the best recommendation performance in the field of conversational recommendation.

[0013] The technical solution provided by this invention is as follows:

[0014] Step 1: Based on the hybrid prototype learning module, perform semantic clustering and topological clustering on the given set of item embedding vectors. First, perform semantic clustering on the given item dataset. The resulting semantic prototype embeddings are used to construct a global graph based on the co-occurrence frequency among items. Then, cluster the global graph, and the nodes in the global graph are clustered into k groups. t A hybrid prototype;

[0015] Step 2: Perform data-level denoising and feature-level denoising based on the information outflow module; data-level denoising is achieved by defining a representation boundary for each session and filtering out items that exceed this boundary; feature-level denoising is achieved by constructing a session local graph and eliminating the influence of irrelevant neighbors on the central node based on a sparse edge graph neural network.

[0016] Step 3: Increase the information content of item and session representations based on the information inflow module; use the global-local enhancement module to refine each item in the local graph through element-level offsets between item representations and their hybrid prototype representations, and then learn session representations through a soft attention mechanism;

[0017] Step 4: Integrate neighborhood information through the neighborhood enhancement task to enhance local session representation; avoid long-tail problems in the recommendation generation process through the next item prediction task.

[0018] Furthermore, step 1 is detailed as follows:

[0019] Step 101: Semantic clustering; given a set of items The item embeddings are initialized, and a clustering algorithm is applied to all item embeddings, with the cluster centroids serving as their semantic prototypes. ,in It is a set of semantic prototype embeddings of objects;

[0020] Step 102: Global Graph Construction; Specifically, iterate through all items in each session and count the occurrences of each item's 1-hop neighbors; for each item, filter out its less frequent neighbors, and the remaining items are considered its global graph neighbors; the global graph is represented as... ,in, For the i-th item up to the j-th item The set of edges, where For items The neighborhood group, Let be the set of weights, and the edges ( , The weight of ) It is an item With items The co-occurrence frequency;

[0021] Step 103: Topological Clustering; Calculate the Laplacian matrix based on the global graph, obtain the q smallest eigenvalues ​​of the normalized Laplacian matrix, and form matrix V with the corresponding eigenvector set. Each row of matrix V represents the embedding of an item in the q-dimensional space. Apply the k-means clustering algorithm to the rows of matrix V to obtain k t A hybrid prototype.

[0022] Furthermore, the specific process of data-level denoising is as follows:

[0023] Given input session and representation space , where d represents the dimension of the embedding vector. Representative conversation The m-th item, where m is the session length; define a representation space. The formula for the center point within is as follows:

[0024]

[0025] in, The symbol represents the center point of the space; It is a conversation Chinese items Hybrid prototype;

[0026] For the conversation For each item in the list, calculate its L2 normalized value. and L2 normalization The cosine similarity between them, at the threshold Under constraints, the cosine similarity is higher than that of other similarities. The items, and the cosine similarity is lower than that of the items. The item is set to 0, and this process is described as follows:

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] Where L2Norm represents L2 normalization. yes The normalized representation, yes The normalized representation, It is a mask sequence. For elements in the mask sequence, It is a conversation The mask representation, This represents element-wise multiplication;

[0032] Design a backward structure that allows the model to adaptively decide whether to reconsider filtered items:

[0033] ;

[0034] ;

[0035] ;

[0036] in, For the session The backmask representation, It is a trade-off coefficient. It is an activation function. and These are learnable parameters. yes The average pooling result, It is a session mask representation The Middle The representation of an item, It is a session backward mask representation The Middle The representation of an item, For the first The final embedding vector of each item.

[0037] Furthermore, the specific process of feature-level denoising is as follows:

[0038] First, the conversation Two items and The transfer patterns between them are further refined into four types: , , and ;in, Indicates from arrive The transfer, Indicates from arrive The transfer, Indicates from arrive The transfer, Indicates from arrive And from arrive Two transfers; session The session local graph is represented as ,in It is a conversation Items in There are two items. and The edges between them; construct a prototype local graph ,in It is a hybrid prototype embedding that contains globally relevant item information from each item;

[0039] Then, feature-level denoising of the session is performed based on an L-layer sparse edge graph neural network, and information aggregation is performed on the session local graph. For the session local graph, an attention mechanism is introduced at each layer to calculate the attention score, as shown in the following formula:

[0040]

[0041] in It is a weight vector. It is the type of edge. Indicates from items to items Importance weight, For learnable parameters are of function;

[0042] After obtaining the attention score, for each item Embedding for neighbors Linear combination:

[0043]

[0044] Therefore, the first sparse edge graph neural network The layers are represented as follows:

[0045]

[0046] in It is the first Layer SEG is used for learnable parameters of the session local graph;

[0047] In addition, information aggregation is performed on the prototype local graph. The aggregation process is the same as that of the session local graph, but different learnable parameters are used:

[0048]

[0049] in It is the first Layer SEG is used for learnable parameters of the prototype local graph;

[0050] Finally, add a residual connection to each layer for the session local graph and the prototype local graph:

[0051] ;

[0052] .

[0053] Furthermore, step 3 is detailed as follows:

[0054] After SEG, based on the N-layer global-local enhancement module, the local update of the nth-layer global-local enhancement module is formulated as follows:

[0055]

[0056]

[0057] in To locally update the weight coefficients, These are learnable parameters. This represents a global hybrid representation of the (n-1)-layer global-local enhancement modules. For the local item representation of the (n-1)th layer global-local enhancement module; the global update representation is:

[0058]

[0059]

[0060] in To update the weight coefficients globally, These are learnable parameters;

[0061] Then, a soft attention mechanism is used to learn the conversation representation; first, the item embeddings are connected with the learnable location embeddings:

[0062]

[0063] in For items The embedding vector output by the Nth layer global-local enhancement module, where || represents concatenation. and These are learnable parameters. It is the first The location embedding of each item; then, the weights learned from the attention mechanism are used to aggregate the item representation:

[0064]

[0065]

[0066] in These are the weighting coefficients. and These are learnable parameters. It is a conversation The average polling result represented by the items in the middle, It is a conversation The conversation is represented by [the subject / method].

[0067] Furthermore, step 4 is detailed as follows:

[0068] For the conversation Its neighborhood is defined as:

[0069]

[0070] in It's a small batch of data. and Is assigned to a session and conversation The representation of the hybrid prototype of the next interactive item; using a contrastive learning task. To maximize mutual information between collaborative sessions:

[0071]

[0072] in, yes The average pooling result, For the session The conversation indicates that, It is a temperature parameter that controls the discrimination of negative samples. This is a penalty item;

[0073] The next item prediction task uses cross-entropy loss as the learning objective:

[0074]

[0075]

[0076] in, This is the one-hot encoded vector of the real label. The probability of each item being the next item. For the representation of an item, i.e., an item Embedded representation, Session representation, item representation, and session representation are achieved through... Normalization smooths the distribution of items;

[0077] The total loss is expressed as:

[0078]

[0079] Where λ is a hyperparameter.

[0080] Furthermore, the aforementioned The function is as follows:

[0081]

[0082] in To determine the set of variables, To determine the variable, It is a single threshold. It is a vector of all 1s, and for the score it satisfies The probability of the entry is set to zero; For the session The weights are calculated as follows:

[0083]

[0084] in , Both are parameters that the model can learn. It is an activation function. This represents the session length.

[0085] The beneficial effects of this invention are as follows:

[0086] Compared to traditional conversation recommendation models, this invention innovatively re-examines the impact of noisy items in conversation recommendation, proposing that in scenarios dominated by short conversations, removing noise does not improve recommendation performance.

[0087] Based on this theoretical foundation, this invention proposes a novel hybrid prototype-enhanced conversation recommendation framework, abbreviated as HyPro. Specifically, HyPro proposes a hybrid prototype learning module that learns hybrid prototype representations of items from all conversations through semantic clustering and topological clustering, thereby capturing global semantics and interaction patterns between items. Secondly, to leverage the compatibility of global and local intents, HyPro also proposes a Global-Local Enhancement Module (GEM) to integrate global prototype information into local conversation representations, thereby enriching the intent information of local conversations. Furthermore, to enhance the model's understanding of item neighborhoods, HyPro introduces a Neighborhood Enhancement Task (NET), which maximizes mutual information between neighborhoods through contrastive learning, thereby improving recommendation performance. Attached Figure Description

[0088] Figure 1 This is a schematic diagram of the structure of the method of the present invention; Detailed Implementation

[0089] To better understand the purpose, structure, and function of this invention, the following detailed description, in conjunction with the accompanying drawings, provides a method for denoising session recommendations in hybrid prototype-based inbound and outbound flow networks.

[0090] A denoising method for session recommendation based on a hybrid prototype inbound / outbound flow network, the overall process is as follows: Figure 1 As shown in (b), a prototype local map is first obtained based on hybrid prototype learning. Then, the prototype denoising module in the information outflow module and the sparse edge map neural network SEG are used to avoid the negative impact of irrelevant information on the representation learning process. Finally, the information domain that can be used to learn individual representations at the session level is expanded through the global-local enhancement module in the information inflow module and the domain enhancement task. The stability of the training process is ensured based on the next item prediction task, and the long-tail problem in the recommendation generation process is avoided. Specifically, the following steps are included:

[0091] Step 1: Based on the Hybrid Prototype Learning module, perform semantic clustering and topological clustering on the given set of item embedding vectors; for example... Figure 1 As shown in (a), in order to learn a hybrid prototype carrying global context information for each item, this embodiment proposes a method called hybrid clustering, which uses semantic and topological clustering sequentially on all items.

[0092] Step 101: Semantic clustering; given a set of items The item embeddings are initialized using a Gaussian distribution with a mean of 0 and a standard deviation of 0.1, and this distribution is optimized during training. Then, a clustering algorithm is applied to all item embeddings to obtain their semantic prototypes, i.e., cluster centroids. Specifically, for an item set... The i-th item Its semantic prototype embedding is defined as ,in It is the semantic prototype embedding set of items. If = Then the item and items Belonging to the same semantic category, among which For item collection The j-th item Semantic prototype embedding.

[0093] Step 102: Global Graph Construction; To better explore interaction patterns from a global perspective, a local graph is generated based on the current session. This local graph is then expanded into an undirected global graph encompassing all interaction sessions. Specifically, all items in each session are traversed, and the occurrence count of each item's one-hop neighbors is counted. Then, for each item, neighbors appearing less than three times are filtered out, and the remaining items are considered its global graph neighbors. The global graph can be represented as follows: ,in, A collection of semantic prototype embeddings. For items arrive The set of edges, where For items The neighborhood group, Let be the set of weights, and the edges ( , The weight of ) It is an item With items The co-occurrence frequency.

[0094] Step 103: Topological clustering; in obtaining the global graph Subsequently, in order to uncover the global transformation patterns of items, in the global graph Then, spectral clustering is used to further learn the topological category of each item. Given a graph First, calculate its Laplace matrix:

[0095]

[0096]

[0097] in Representativeness matrix The i-th row and i-th column, The identity matrix is ​​used, and then the normalized Laplacian matrix is ​​calculated. The eigenvalues ​​and eigenvectors of . Let . To minimize the normalized Laplace matrix The set of eigenvalues ​​and their corresponding eigenvectors Forming a matrix ,matrix Each row represents the item in Embedding in 3D space. Then, the k-means clustering algorithm is applied to the matrix. The number of clusters in the predefined clustering is... ,matrix The i-th row corresponds to the item set The i-th item also corresponds to a node in the global graph, after which the nodes of the global graph are reclassified as k. t A topological prototype. For k t A topological prototype, whose embedding vector is obtained by average pooling the embedding vectors of all semantic prototypes belonging to the corresponding cluster. In this case, since the representation of the topological prototype combines semantic and topologically related item information, we name it a hybrid prototype and use... Represents items A hybrid prototype.

[0098] Step 2: Perform data-level and feature-level denoising based on the out-flow of information module. To eliminate the impact of irrelevant information on the representation learning process, this invention designs an out-flow channel containing two stages of denoising. At the data level, since noisy items often exhibit characteristics inconsistent with the overall features of their respective sessions, the goal is to define a representation boundary for each session in the feature space to exclude items that do not belong to that boundary. At the feature level, the goal is to eliminate the influence of irrelevant neighbors on the central node during the propagation of the GNN. Figure 1 (c) in the figure provides a brief demonstration of data and feature-level denoising.

[0099] Step 201: Construct a Prototypical Denoise Module to perform data-level denoising;

[0100] Given input session and representation space , where d represents the dimension of the embedding vector. Representative conversation Let m be the m-th item, and m be the session length. Define a representation space. The formula for the center point within is as follows:

[0101]

[0102] Where m is the session length. Represents the center point of the representation space; It is a conversation Chinese items A hybrid prototype. (Conversation) The boundary of the representation is defined as:

[0103]

[0104] Where x is any point in space, Represents the L2 norm. It is a representation The range of hyperparameters is limited. Given the vast range of Euclidean space, it is difficult to set hyperparameters to determine the Euclidean threshold. Therefore, the Euclidean distance metric is approximated using cosine similarity through the following relationship, thereby scaling the Euclidean space to a unit hypersphere:

[0105]

[0106] Where x and y are arbitrary variables. yes and L2 norm, It's a calculation using Euclidean distance. This is a cosine similarity calculation. Since the above formula is only applicable to... and It is valid when they have the same L2 norm, therefore by applying... The norm-normalized input is approximated by cosine similarity as a Euclidean distance metric. Based on the above formula, we use another hyperparameter. It is the threshold of cosine similarity used to replace .

[0107] Based on the above analysis, for the session For each item in the list, calculate its L2 normalized value. and L2 normalization The cosine similarity between them. Then, at a threshold... Under constraints, the cosine similarity is higher than that of other similarities. The items, and the cosine similarity is lower than that of the items. The items are set to 0. This process can be described as:

[0108]

[0109]

[0110]

[0111]

[0112] Where L2Norm represents L2 normalization. yes The normalized representation, yes The normalized representation, It is a mask sequence. For elements in the mask sequence, It is a conversation The mask representation, This represents element-wise multiplication. A mixed prototype is used here. To determine the boundary rather than the item The reason for this embedding vector is that even for items of the same category, their embeddings may still be relatively distant in the feature space. If the hyperparameters... If not properly tuned, this can cause the model to discard items belonging to the same category. In contrast, hybrid prototype representations can assign similar representations to nodes with similar semantics and topological features, thus avoiding this problem.

[0113] Furthermore, by designing a backward structure, the model is allowed to adaptively decide whether to reconsider filtered items:

[0114]

[0115]

[0116]

[0117] in, For the session The backmask representation, It is a trade-off coefficient. It is an activation function. yes The average pooling result, and These are learnable parameters. It is a session mask representation The Middle The representation of an item, It is a session backward mask representation The Middle The representation of an item, For the first The final embedding vector of each item.

[0118] After removing noisy items from the input session, the session is reconstructed into a local graph to leverage higher-order correlations between items. Specifically, the session is... Two items and The transfer patterns between them are further refined into four categories: , , and .in, Indicates from arrive The transfer, Indicates from arrive The transfer, Indicates from arrive The transfer, Indicates from arrive And from arrive The two transitions. This division aims to give the model the ability to recognize different transition patterns between items, thus laying the foundation for learning higher-quality conversation representations later. Conversation The session local graph is represented as ,in It is a conversation Items in There are two items. and The edges between them. In addition, we constructed a prototype local graph. ,in It is a hybrid prototype embedding that contains global relevant item information from each item.

[0119] Step 202: Feature-level denoising of the session based on Sparse-Edge GNNs (SEG). Specifically, after constructing the local session graph, different learnable parameters are assigned to each type of edge to enhance information aggregation in the SEG. To effectively capture the different importance of neighboring items, an attention mechanism is introduced, which dynamically learns the importance of connections between item pairs. These learned attention weights allow the model to prioritize more relevant neighbors. The attention score is calculated as follows:

[0120]

[0121] in It is a weight vector. It is the type of edge. Indicates from items to items Importance weights. The formula for calculating the function is:

[0122]

[0123] in To determine the set of variables, To determine the variable, It is a single threshold. It is a vector of all 1s. For the score, it satisfies... The entries are set to have a probability of zero. To enable the attention score to learn adaptively, we calculate weights based on the overall characteristics exhibited by each session. Instead of treating it as a hyperparameter, it is used in the context of a session. weight The calculation process is as follows:

[0124]

[0125] in , Both are parameters that the model can learn. It is an activation function.

[0126] After obtaining the attention score, perform a linear combination of the neighbors for each item embedding:

[0127]

[0128] By introducing The function rescales the similarity scores between items to a sparse feature space, which can further eliminate the influence of the representations of irrelevant neighbors on the central node during feature aggregation.

[0129] Therefore, SEG's first The layers are:

[0130]

[0131] in It is the first Layer SEG is used for learnable parameters of the session local graph.

[0132] In addition, information aggregation is performed on the prototype local graph. The aggregation process is the same as that of the session local graph, but different learnable parameters are used:

[0133]

[0134] in It is the first Layer SEG is used for learnable parameters of the prototype local graph.

[0135] Finally, to address the smoothing issue of GNNs and prevent overfitting, a residual connection is added to both the session local map and the prototype local map at each layer:

[0136]

[0137]

[0138] Step 3: Increase the information content of item and session representations based on the In-flow of Information module. An information inflow channel is designed within the module, which includes a global-local enhancement module and a neighborhood enhancement task.

[0139] Step 301: Construct the Global-local Enhancement Module; the Global-local Enhancement Module is constructed in the... After layer SEG, each item in the local graph is refined by element-level offsets between its item representation and its hybrid prototype representation. For simplicity, the superscripts are omitted in the following formulas. The nth layer of this module consists of two parts: local updates and global updates. For the local update at the nth layer, we formulate it as:

[0140]

[0141]

[0142] in To locally update the weight coefficients, These are learnable parameters. This represents a global hybrid representation of the (n-1)-layer global-local enhancement modules. For the local item representation of the n-1 layer global-local enhancement module; global update can be expressed as:

[0143]

[0144]

[0145] in To update the weight coefficients globally, These are learnable parameters. The fusion of elements from global and local representations enables the model to capture interaction features at different levels, thereby enhancing its ability to model complex relationships between items.

[0146] go through After layer enhancement, to distinguish and aggregate the contribution of each item to the current session, we employ a soft attention mechanism to learn session representations. First, we concatenate item embeddings with learnable location embeddings to capture the temporal dependencies between items:

[0147]

[0148] in For items The embedding vector output by the Nth layer global-local enhancement module, where || represents concatenation. and These are learnable parameters. It is the first The location of each item is embedded.

[0149] Next, the weights learned from the attention mechanism are used to polymerize the product representation:

[0150]

[0151]

[0152] in These are the weighting coefficients. and These are learnable parameters. It is a conversation The average polling result represented by the items in the middle, It is a conversation The conversation is represented by [the subject / method].

[0153] Step 302: Neighborhood Enhancement Task (NET) – Enhances local session representations. In real-world scenarios, different interactive sessions may point to similar next-interaction item categories, making it possible to consider the mutual enhancement of these collaborative sessions, thereby expanding the information domain used for learning individual representations at the session level. We define these collaborative sessions as neighborhoods, for each session... Its neighborhood is defined as:

[0154]

[0155] in It's a small batch of data. and Is assigned to a session and conversation The next interactive item is represented by a hybrid prototype. Following this, a contrastive learning task is used to maximize mutual information between collaborative sessions:

[0156]

[0157] in, yes The average pooling result, For the session The conversation indicates that, This is a temperature parameter that controls the discrimination of negative samples. In addition, another penalty term is introduced. To limit the variance of neighborhood sessions:

[0158]

[0159] in, It is a hyperparameter that controls the degree of punishment. In small batches of data The normalization operation performed on top, Calculation belongs to The variance of the session representation is used to combine the above formula to obtain the learning loss for the neighborhood augmentation task (NET):

[0160]

[0161] Using the above formula, the session and its neighborhood set Mutual information is maximized. Furthermore, to reduce the influence of neighborhood sets with large internal differences, a penalty term is employed, implying the potential inclusion of noisy sessions. Therefore, the Neighborhood Enhancement Task (NET) enables HyPro to integrate additional relevant information to enhance local session representations.

[0162] Step 4: Based on the Next-item Prediction Task, ensure the stability of the training process and avoid long-tail problems in the recommendation generation process.

[0163] Perform on item representation matrix and session representation Normalization to smooth the distribution of items:

[0164]

[0165] in These are the normalized weights. Then, the normalized session representation is calculated. Representation of items The product of, and apply The function determines the probability of each item being the next item. :

[0166]

[0167] in For items The embedding representation is then used. Cross-entropy loss is then employed as the learning objective.

[0168]

[0169] in Let be the one-hot encoded vector of the real label. To combine the neighborhood enhancement task and the next item prediction task, a multi-task learning loss is introduced. Specifically, a hyperparameter λ is introduced to control the scale of NET. The total loss can be expressed as:

[0170]

[0171] Extensive experiments were conducted on three public datasets, and the results show that our method achieves state-of-the-art performance.

[0172] Datasets: Tmall, from the IJCAI-15 competition, consists of shopping logs from numerous unnamed users on the Tmall online shopping platform. Diginetica, from the 2016 CIKM Cup, describes users' music listening behavior. RetailRocket, released by an e-commerce company for a Kaggle competition, contains users' browsing activity.

[0173] The data format is shown in Table 1:

[0174] Table 1. Dataset Parameters

[0175] Dataset Tmall Diginetica Retailrocket Training Session 351,268 433,643 719,470 Test session 25,898 15,132 60,858 thing 40,728 36,968 43,097 average length 6.69 5.43 5.12

[0176] Evaluation indicators:

[0177] P@K: The P@K score is widely used to measure the prediction accuracy in the context of conversational recommendation. P@K represents the proportion of correctly recommended items in the test case that rank within the top K positions in the ranking table.

[0178] MRR@K: The MRR@K score emphasizes the position of the user's desired item in the model's recommendation list; the higher the position, the better the value.

[0179] Comparison method:

[0180] NARM: Uses RNN and attention mechanisms to extract users' general interests.

[0181] STAMP: Emphasizes short-term memory by utilizing attention mechanisms.

[0182] SR-GNN: Uses GNN to learn the embeddings of items and learns the representation of the conversation through a soft attention mechanism.

[0183] GCE-GNN: Considers extracting spatial information from global and local views of session data.

[0184] DHCN uses a hypergraph to generate two distinct views from the original session, treating them as self-supervised signals to enhance the session representation.

[0185] AttenMixer: Treats combinations of the last few items in a session as different intentions and implements a multi-level inference component for GNNs.

[0186] SRGI: Integrates global transition patterns into local sessions through fusion-based and constraint-based models.

[0187] Table 2. Overall HyPro Results Comparison

[0188]

[0189] As shown in Table 2, the HyPro method provided in this invention exhibits the best performance on all three datasets. This demonstrates the effectiveness of the inbound and outbound flow network composed of the information outflow module and the information inflow module. HyPro's advantages come from two sources: (i) it incorporates global semantic and topological information into the denoising process, alleviating the biased denoising problem in information-scarce scenarios. (ii) in addition to denoising, it introduces global information to enhance the amount of relevant information contained in each session. These advantages ensure that HyPro's inbound and outbound flow network increases relevant information while reducing irrelevant information. According to the information bottleneck theory, HyPro can produce better session representations compared to the baseline model, thereby achieving more accurate recommendations.

[0190] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for denoising session recommendation in a hybrid prototype-based inbound / outbound flow network, characterized in that, Includes the following steps: Step 1: Based on the hybrid prototype learning module, perform semantic clustering and topological clustering on the given set of item embedding vectors. First, perform semantic clustering on the given item dataset. The resulting semantic prototype embeddings are used to construct a global graph based on the co-occurrence frequency among items. Then, cluster the global graph, and the nodes in the global graph are clustered into k groups. t A hybrid prototype; Step 2: Perform data-level denoising and feature-level denoising based on the information outflow module; Data-level denoising works by defining a representation boundary for each session and filtering out items that exceed this boundary. Feature-level denoising eliminates the influence of irrelevant neighbors on the central node by constructing a session local graph and using a sparse edge graph neural network (SEG). Step 3: Increase the information content of item and session representations based on the information inflow module; use the global-local enhancement module to refine each item in the local graph through element-level offsets between item representations and their hybrid prototype representations, and then learn session representations through a soft attention mechanism; Step 4: Integrate neighborhood information through the neighborhood enhancement task to enhance local session representation; avoid long-tail problems in the recommendation generation process through the next item prediction task.

2. The session recommendation denoising method for inbound and outbound flow networks based on a hybrid prototype as described in claim 1, characterized in that, Step 1 is described in detail as follows: Step 101: Semantic clustering; given a set of items The item embeddings are initialized, and a clustering algorithm is applied to all item embeddings, with the cluster centroids serving as their semantic prototypes. ,in It is a set of semantic prototype embeddings of objects; Step 102: Global graph construction; specifically, traverse all items in each session and count the occurrences of each item's 1-hop neighbors; For each item, its least frequent neighbors are filtered out, and the remaining items are considered its global graph neighbors; the global graph is represented as follows: ,in, For the i-th item up to the j-th item The set of edges, where For items The neighborhood group, Let be the set of weights, and the edges ( , The weight of ) It is an item With items The co-occurrence frequency; Step 103: Topological clustering; The Laplacian matrix is ​​calculated based on the global graph. The q smallest eigenvalues ​​of the normalized Laplacian matrix are obtained, and the set of their corresponding eigenvectors forms matrix V. Each row of matrix V represents the embedding of an item in q-dimensional space. The k-means clustering algorithm is applied to the rows of matrix V to obtain k t A hybrid prototype.

3. The session recommendation denoising method for inbound and outbound flow networks based on a hybrid prototype according to claim 2, characterized in that, The specific process of data-level denoising is as follows: Given input session and representation space , where d represents the dimension of the embedding vector. Representative conversation The m-th item, where m is the session length; define a representation space. The formula for the center point within is as follows: ； in, The symbol represents the center point of the space; It is a conversation Chinese items Hybrid prototype; For the conversation For each item in the list, calculate its L2 normalized value. and L2 normalization The cosine similarity between them, at the threshold Under constraints, the cosine similarity is higher than that of other similarities. The items, and the cosine similarity is lower than that of the items. The item is set to 0, and this process is described as follows: ； ； ； ； Where L2Norm represents L2 normalization. yes The normalized representation, yes The normalized representation, It is a mask sequence. For elements in the mask sequence, It is a conversation The mask representation, This represents element-wise multiplication; Design a backward structure that allows the model to adaptively decide whether to reconsider filtered items: ； ； ； in, For the session The backmask representation, It is a trade-off coefficient. It is an activation function. and These are learnable parameters. yes The average pooling result, It is a session mask representation The Middle The representation of an item, It is a session backward mask representation The Middle The representation of an item, For the first The final embedding vector of each item.

4. The session recommendation denoising method for hybrid prototype-based inbound / outbound flow networks according to claim 3, characterized in that, The specific process of feature-level denoising is as follows: First, the conversation Two items and The transfer patterns between them are further refined into four types: , , and ;in, Indicates from arrive The transfer, Indicates from arrive The transfer, Indicates from arrive The transfer, Indicates from arrive And from arrive Two transfers; session The session local graph is represented as ,in It is a conversation Items in There are two items. and The edges between them; construct a prototype local graph ,in It is a hybrid prototype embedding that contains globally relevant item information from each item; Then, feature-level denoising of the session is performed based on an L-layer sparse edge graph neural network, and information aggregation is performed on the session local graph. For the session local graph, an attention mechanism is introduced at each layer to calculate the attention score, as shown in the following formula: ； in It is a weight vector. It is the type of edge. Indicates from items to items Importance weight, The function is a learnable parameter. of function; After obtaining the attention score, for each item Embedding for neighbors Linear combination: ； Therefore, the first sparse edge graph neural network The layers are represented as follows: ； in It is the first Learnable parameters of the Sparse Edge Graph Neural Network (SEG) for session local graphs; In addition, information aggregation is performed on the prototype local graph. The aggregation process is the same as that for the session local graph, but different learnable parameters are used: ； in It is the first Layer SEG is used for learnable parameters of the prototype local graph; Finally, add a residual connection to each layer for the session local graph and the prototype local graph: ； 。 5. The session recommendation denoising method for inbound and outbound flow networks based on a hybrid prototype according to claim 4, characterized in that, Step 3 is as follows: After SEG, based on the N-layer global-local enhancement module, the local update of the nth-layer global-local enhancement module is formulated as follows: ； ； in To locally update the weight coefficients, These are learnable parameters. This represents a global hybrid representation of the (n-1)-layer global-local enhancement modules. For the local item representation of the (n-1)th layer global-local enhancement module; the global update representation is: ； ； in To update the weight coefficients globally, These are learnable parameters; Then, a soft attention mechanism is used to learn the conversation representation; first, the item embeddings are connected with the learnable location embeddings: ； in For items The embedding vector output by the Nth layer global-local enhancement module, where || represents concatenation. and These are learnable parameters. It is the first The location embedding of each item; then, the weights learned from the attention mechanism are used to aggregate the item representation: ； ； in These are the weighting coefficients. and These are learnable parameters. It is a conversation The average polling result represented by the items in the middle, It is a conversation The conversation is represented by [the subject / method].

6. The session recommendation denoising method for inbound and outbound flow networks based on a hybrid prototype according to claim 5, characterized in that, Step 4 is as follows: For the conversation Its neighborhood is defined as: ； in It's a small batch of data. and Is assigned to a session and conversation The representation of the hybrid prototype of the next interactive item; Using a contrastive learning task To maximize mutual information between collaborative sessions: ； in, yes The average pooling result, For the session The conversation indicates that, It is a temperature parameter that controls the discrimination of negative samples. This is a penalty item; The next item prediction task uses cross-entropy loss as the learning objective: ； ； in, This is the one-hot encoded vector of the real label. The probability of each item being the next item. For the representation of an item, i.e., an item Embedded representation, Session representation, item representation, and session representation are achieved through... Normalization smooths the distribution of items; The total loss is expressed as: ； Where λ is a hyperparameter.

7. The session recommendation denoising method for inbound and outbound flow networks based on a hybrid prototype as described in claim 6, characterized in that, The The function is as follows: ； in To determine the set of variables, To determine the variable, It is a single threshold. It is a vector of all 1s, and for the score it satisfies The probability of the entry is set to zero; For the session The weights are calculated as follows: ； in , Both are parameters that the model can learn. It is an activation function. This represents the session length.

Images