A session recommendation method based on a denoising edge perception network

Abstract

The application provides a conversation recommendation method based on a denoising edge perception network, and belongs to the technical field of conversation recommendation. The method comprises the following steps: mining items in all historical conversation data and jump order neighbors of the items, and constructing a frequency perception graph by using co-occurrence frequencies of the items and the jump order neighbors of the items; obtaining an item embedding matrix according to information carried by edges, assigning learning parameters to the edges by using a denoising edge perception network, obtaining item representation learning, and obtaining item representation by calculation; and obtaining conversation representation by using a soft attention mechanism; calculating a ground truth one-hot encoding vector according to the conversation representation and the item embedding matrix, judging a next target item after a user inputs a conversation according to the ground truth one-hot encoding vector and individual contributions of the items, and completing conversation recommendation; and the application solves the problems of ignoring modeling of frequency relationships between items and being unable to effectively process noise in existing conversation recommendation algorithms.

Description

Technical Field

[0001] The present invention belongs to the technical field of conversation recommendation, and in particular relates to a conversation recommendation method based on a denoising edge perception network. Background Art

[0002] Recommender systems are essential for online platforms, extracting user preferences from their profiles and behaviors to combat information overload. However, users are often anonymous and transient on websites, such as news and streaming platforms, making session-based recommendations crucial. They focus on predicting the next item based on anonymous interaction sequences.

[0003] Traditional session-based recommendation algorithms model complex item transition relationships by assuming that the next item is only related to the last one or by leveraging deep learning's strengths in extracting latent information and complex relationships from raw data. However, traditional methods only aggregate information from adjacent items to facilitate learning item representations, and most ignore the potential impact of noisy neighbors caused by user misclicks or wandering interests. A few algorithms consider noise issues by removing noisy items from the input session. However, due to the short input sequences of session recommendations, denoising strategies that directly apply to the input result in sparse data. Furthermore, these strategies require accurate extraction of the user's overall intent, potentially leading to items being misclassified as noise. Summary of the Invention

[0004] In response to the above-mentioned deficiencies in the prior art, the present invention provides a conversation recommendation method based on a denoising edge perception network, which solves the technical problems of the existing conversation recommendation algorithm ignoring the modeling problem of the frequency relationship between items and the inability to effectively handle noise.

[0005] To achieve the above objectives, the present invention adopts a technical solution: a conversation recommendation method based on a denoising edge perception network, comprising the following steps:

[0006] S1. Mining all items and their hop-order neighbors in all historical conversation data, and constructing a frequency perception graph containing each item using the co-occurrence frequency of each item and its hop-order neighbors;

[0007] S2. Obtain the item embedding matrix based on the information carried by the edge in the frequency perception graph, use the denoising edge perception network to assign learning parameters to the edge, obtain the item representation learning and obtain the item representation through calculation;

[0008] S3. Based on the item representation, use the soft attention mechanism to obtain the session representation;

[0009] S4. Based on the session representation and the item embedding matrix, the ground truth one-hot encoding vector is calculated. Based on the ground truth one-hot encoding vector and the individual contribution of each item, combined with the current item after the user inputs the session, the next item is predicted to complete the session recommendation.

[0010] The beneficial effects of the present invention are as follows: the present invention takes sparse consideration of adjacent items, eliminates the influence of potential noise neighbors, improves the accuracy of conversation recommendations, and refines the connection relationship between items in the input conversation by using a frequency perception map, thereby improving the representativeness of the items and making the recommendation method more accurate and comprehensive. The present invention can be used on multiple platforms with a single network and is not limited to recommendations on a specific platform. It improves the practicality of conversation recommendations, enhances the generalizability of conversation recommendations, and realizes the ability to accurately recommend the next item based on the conversation even if the user accidentally touches it.

[0011] Furthermore, the S1 includes the following steps:

[0012] S101, mining items and their hop-order neighbors in all historical conversation data, combining items and their hop-order neighbors into adjacent item pairs, calculating the number of adjacent item pairs in all historical conversation data, and using the number of adjacent item pairs as edge weights in the frequency perception graph;

[0013] S102, normalizing the edge weight of each item according to the edge weight to obtain a frequency representation of the edge weight of each item;

[0014] S103. Classify the edge types according to the frequency representation of the edge weights of each item, and combine the frequency representation of the edge weights of the hopping neighbors of each item to obtain the co-occurrence frequency of each item and its hopping neighbors, and construct a frequency perception map containing each item.

[0015] The beneficial effect of the above further solution is that the present invention constructs a frequency perception graph in the denoising edge perception network to understand the relative importance of each neighbor node to the central item, thereby improving the noise item perception ability of the denoising edge perception network.

[0016] Furthermore, the S2 includes the following steps:

[0017] S201: Obtain a specific input session based on information carried by edges in the frequency perception graph, initialize the vector form of the input session to obtain a vectorized session, and combine the vectorized sessions to obtain an item embedding matrix;

[0018] S202, using the edge-aware graph attention subnetwork in the denoising edge-aware network to assign learnable parameters to the edges;

[0019] S203. Calculate the attention score in the semantic space from item to item neighbor using vectorized session and learnable parameters in combination with leaky rectified linear units. Use a self-attention mechanism based on the α-entmax activation function and calculate the sparse weight score using the attention score.

[0020] S204. Obtain the final attention weight according to the sparse weight score of each item;

[0021] S205. Perform a linear combination of the final attention weight and the vectorized session associated with the final attention weight to calculate the item representation.

[0022] The beneficial effects of the above further scheme are: the present invention uses the edge-aware graph attention sub-network in the denoising edge-aware network, and utilizes the edge-aware graph attention sub-network to more accurately model the effect of the graph structure according to the information carried by the edge, assigns different learning parameters to the edges, refines the learning process, improves the accuracy of item representation learning, and uses the self-attention mechanism to sparsely consider the learning parameters, eliminating the influence of noise items on the central items, so that the conversation recommendation method of the present invention can effectively handle noise.

[0023] Furthermore, the S3 includes the following steps:

[0024] S301, calculating the average value of the item representations, and representing the user's interaction information as the average value of the item representations;

[0025] S302: Use the attention mechanism to obtain the weight of each item, and linearly combine the weight of each item with the item representation to calculate the session representation.

[0026] Furthermore, the expression of the session representation is as follows:

[0027]

[0028] Among them, R u represents the session representation of session u, n represents the length of the session, i represents the i-th item in the session, δ i represents the weight of the i-th item, represents the i-th item representation of session u, represents the i-th interaction information of user u in session, b1, W1, and W2 all represent learnable parameters; b1∈R d ,W1,W2∈R d×d , R d represents a set of points in d-dimensional real space, R d×d Represents a set of d×d real number matrices, where d represents the dimension.

[0029] Further, the expression of the one-hot encoding vector of the ground truth is as follows:

[0030] y i = softmax((R u ) T E i )

[0031] wherein y i represents the one-hot encoding vector of the ground truth, softmax(·) represents a normalized exponential function, R u represents the session representation of the session u, E i represents the item embedding matrix.

[0032] The above further scheme has the beneficial effect that the present application improves the ability to distinguish and obtain the individual contribution of each item by adopting a soft attention mechanism to obtain the session representation, and improves the accuracy of the next item prediction by combining the individual contribution of each item in the current session into the prediction of the next item through the session representation. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is a flow chart of the method of the present application.

[0034] Figure 2 is a whole structure diagram of the denoising edge perception network in the present embodiment. DETAILED DESCRIPTION

[0035] The specific embodiments of the present application are described below to facilitate the understanding of the present application by those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0036] Before explaining the present embodiment, the following terms are explained:

[0037] EA-GAT: edge-aware graph attention network;

[0038] D-EAN: denoising edge-aware network;

[0039] alpha-entmax: alpha-entropy maximization normalization;

[0040] Sigmoid: sigmoid function;

[0041] Softmax: normalized exponential function.

[0042] EMBODIMENT

[0043] In this embodiment, the application provides a session recommendation method based on a denoising edge-aware network, which establishes a frequency-aware graph as a basic idea, captures more detailed connection relationships according to the global frequency of the connection between items and adjacent items in a user input session, identifies edge connection relationships and processes neighbor noise by using an edge-aware graph attention subnetwork (EA-GAT) in a denoising edge-aware network (D-EAN), discriminates and combines by using a soft attention mechanism, combines the individual contribution of each item in the current session to the next item prediction, and solves the problem of recommendation bottleneck caused by the fact that noise cannot be effectively processed.

[0044] In view of the problem that the frequency relationship between items and item connections in a session is ignored, resulting in that the relationship with adjacent items cannot be well obtained, a frequency-aware graph is established. When learning item representation by using an edge-aware graph attention network, parameters are assigned to each edge according to the types of connected edges in the frequency-aware graph. A soft attention mechanism is used to obtain a session representation for prediction, and finally a one-hot encoding of each type is obtained after softmax, so that the next target item that is most likely to appear is determined according to the one-hot encoding, and the session recommendation based on the denoising edge-aware network is completed.

[0045] As shown in Figure 1 The application provides a session recommendation method based on a denoising edge-aware network, and the implementation method is as follows:

[0046] S1, mine all the items and the jump order neighbors of the items in the historical session data, and use the co-occurrence frequency of each item and each jump order neighbor of the item to construct a frequency-aware graph containing each item, and the specific steps are as follows:

[0047] S101, mine all the items and the jump order neighbors of the items in the historical session data, and combine the items and the jump order neighbors of the items into adjacent item pairs, calculate the number of adjacent item pairs in all the historical session data, and use the number of adjacent item pairs as the weight of the edge in the frequency-aware graph;

[0048] S102, normalize the edge weight of each item according to the weight of the edge, and obtain the frequency representation of the edge weight of each item;

[0049] S103, classify the types of edges according to the frequency representation of the edge weight of each item, and combine the frequency representation of the jump order neighbor edge weight of each item to obtain the co-occurrence frequency of each item and each jump order neighbor, and construct a frequency-aware graph containing each item.

[0050] In this embodiment, the frequency-aware graph of each item is constructed by using the co-occurrence frequency of the item and its skip-order neighbor mined from historical all-session data. First, the number of adjacent item pairs in all sessions is calculated, and then the number of adjacent item pairs is used as the weight of the edge in the frequency-aware graph. On this basis, the edge weight of each item is normalized to obtain its frequency representation, and the type of edge is classified according to the frequency size. By constructing the frequency-aware graph, the relative importance of each neighbor node to the center item can be understood to perceive the noise item. The type of edge is divided into four categories: super-high-frequency edge, high-frequency edge, medium-frequency edge and low-frequency edge.

[0051] S2, according to the information carried by the edge in the frequency-aware graph, obtaining the item embedding matrix, using the denoising edge-aware network to assign learning parameters to the edge, obtaining the item representation learning and obtaining the item representation by calculation, the specific steps are as follows:

[0052] S201, according to the information carried by the edge in the frequency-aware graph, obtaining the specific input session, initializing the vector form of the input session to obtain the vectorized session, and combining each vectorized session to obtain the item embedding matrix;

[0053] S202, using the edge-aware graph attention subnetwork in the denoising edge-aware network to assign learnable parameters to the edge;

[0054] S203, using the vectorized session and the learnable parameter and combining the modified linear unit with leakage to calculate the attention score of the item to the item neighbor in the semantic space. An attention mechanism based on an alpha-entmax activation function is used, and sparse weight scores are calculated using the attention score;

[0055] S204, according to the sparse weight score of each item, obtaining the final attention weight;

[0056] S205, linearly combining the final attention weight and the vectorized session associated with the final attention weight to calculate the item representation.

[0057] In this embodiment, as shown in Figure 2 , the overall architecture of the denoising edge-aware network is shown in Figure 2 . The EA-GAT subnetwork in the denoising edge-aware network can more accurately model the graph structure according to the information carried by the edge. Different learning parameters are assigned to the edge to refine the learning process, so as to obtain item representation learning and item representation;

[0058] In the process of item representation learning of the EA-GAT subnetwork, the influence of adjacent noise items in the EA-GAT learning process is considered, the learning parameters are sparsely considered using alpha-entmax, and the weight score of the noise item to the center item is minimized.

[0059] In this embodiment, the session information carried by the edge, the specific input session Initialize the vector form of the input session as Vectorize the session into a matrix form as the item embedding matrix E i , considering the semantic and structural noise, calculate the attention score of item i to its neighbors in the semantic space, the calculation formula is as follows:

[0060]

[0061] Wherein, represents the attention score of item i to item j in the semantic space, LeakyReLU(·) represents the leaky rectified linear unit, r i,j represents the type of edge e ij between item i and item j in the frequency-aware graph, represents the learnable parameter assigned to the edge e ij between item i and item j, and R d×d represents the vector space, represents the vectorized session of item i, represents the vectorized session of item j, and represents matrix dot product.

[0062] Adopting the self-attention mechanism based on alpha-entmax, the influence of noise is completely eliminated, and the sparse consideration weight score is calculated, the sparse weight score, the calculation formula is as follows:

[0063]

[0064] Wherein, represents the sparse weight score, represents the hyperparameter;

[0065] The final attention weight is obtained from the sparse weight score, and then the features associated with the final attention weight are linearly combined to calculate the item representation The calculation formula of the item representation is as follows:

[0066]

[0067] Wherein, represents the item representation of item i, represents the final attention weight of item i to item j, represents the vectorized session of item j, represents the neighbor node set of node i in session u, which contains all the nodes connected to node i.

[0068] S3, according to the item representation, the soft attention mechanism is used to obtain the session representation, and the specific steps are as follows:

[0069] S301, the average value of the item representation is calculated, and the user's interaction information is represented as the average value of the item representation.

[0070] S302, the weight of each item is obtained by using the attention mechanism, and each item weight is linearly combined with the item representation, and the session representation is calculated.

[0071] In this embodiment, for the generated item representation The soft attention mechanism is used to obtain the session representation for distinguishing and combining the individual contribution of each item in the current session into the prediction of the next item, and the specific steps are as follows:

[0072] The user's interaction information is represented as the average value of the item representation, and the expression is as follows:

[0073]

[0074] Wherein, represents the i-th interaction information of the user, and m represents the total number of items, represents the i-th item representation;

[0075] The weight of each item is obtained by using the attention mechanism, and the calculation expression is as follows:

[0076]

[0077] Wherein, i represents the i-th item in the session, and δ i represents the weight of the i-th item, represents the i-th interaction information of the user u in the session, represents the i-th item representation of the session u, b1, W1, W2 all represent learnable parameters; wherein b1∈R d , W1,W2∈R d×d , R d represents a set of points in a d-dimensional real space, and R d×d represents a set of dxd real number matrices, and d represents the dimension;

[0078] And the session representation of the session u is obtained by using linear combination, and the calculation expression is as follows:

[0079]

[0080] Wherein, R u represents the session representation of the session u, n represents the length of the session, i represents the i-th item in the session, and δ irepresents the weight of the i-th item, represents the i-th item representation of session u.

[0081] In this embodiment, U represents a session set, V = {v1, v1, ..., v m} represents a unique set derived from the session set, m represents the total number of items, represents a specific session, u∈U represents a session, n represents the session length, represents the interaction term at time step p, Represents a vectorized session, which is combined into a matrix to form the project embedding matrix E i , by a specific session Embedded into vector space.

[0082] S4. Based on the session representation and the item embedding matrix, the ground truth one-hot encoding vector is calculated. Based on the ground truth one-hot encoding vector and the individual contribution of each item, combined with the current item after the user inputs the session, the next item is predicted to complete the session recommendation.

[0083] In this example, the session representation and item embedding matrices are multiplied, and softmax is used to scale the result. The calculation formula is as follows:

[0084] y i =softmax((R u ) T E i )

[0085] Among them, y i represents the one-hot encoded vector of the ground truth, softmax(·) represents the normalized exponential function, and R u represents the conversation representation of conversation u, E i represents the item embedding matrix;

[0086] The calculated one-hot encoding is used to determine the user's most likely next target item. Based on the ground truth one-hot encoding vector and the individual contribution of each item, and combined with the current item after the user enters the session, the next item is predicted to complete the session recommendation based on the denoising edge perception network.

Claims

1. A conversation recommendation method based on a denoising edge perception network, characterized in that: The following steps are involved: S1. Mining all items and their hop-order neighbors in all historical conversation data, and constructing a frequency perception graph containing each item using the co-occurrence frequency of each item and its hop-order neighbors. Specifically: S101, mining items and their hop-order neighbors in all historical conversation data, combining items and their hop-order neighbors into adjacent item pairs, calculating the number of adjacent item pairs in all historical conversation data, and using the number of adjacent item pairs as edge weights in the frequency perception graph; S102, normalizing the edge weight of each item according to the edge weight to obtain a frequency representation of the edge weight of each item; S103, classifying edge types based on the frequency representation of the edge weights of each item, and combining the frequency representation of the edge weights of each item's hopping neighbors to obtain the co-occurrence frequency of each item and its hopping neighbors, and constructing a frequency perception map containing each item; S2. According to the information carried by the edge in the frequency perception graph, the item embedding matrix is ​​obtained, and the denoising edge perception network is used to assign learning parameters to the edge to obtain the item representation learning and obtain the item representation through calculation. Specifically: S201: Obtain an input session based on information carried by edges in the frequency perception graph, initialize the vector form of the input session to obtain a vectorized session, and combine the vectorized sessions to obtain an item embedding matrix; S202, using the edge-aware graph attention subnetwork in the denoising edge-aware network to assign learnable parameters to the edges; S203, using vectorized sessions and learnable parameters combined with leaky rectified linear units, calculate the attention score in the semantic space from item to item neighbor, using Activation function-based self-attention mechanism, and sparse weight scores are calculated using attention scores; S204. Obtain the final attention weight according to the sparse weight score of each item; S205, performing a linear combination of the final attention weight and the vectorized session associated with the final attention weight to calculate an item representation; S3. Based on the item representation, use the soft attention mechanism to obtain the session representation; S4. Based on the session representation and the item embedding matrix, the ground truth one-hot encoding vector is calculated. Based on the ground truth one-hot encoding vector and the individual contribution of each item, combined with the current item after the user inputs the session, the next item is predicted to complete the session recommendation.

2. The conversation recommendation method based on denoising edge perception network according to claim 1, characterized in that The S3 includes the following steps: S301, calculating the average value of the item representations, and representing the user's interaction information as the average value of the item representations; S302: Use the attention mechanism to obtain the weight of each item, and linearly combine the weight of each item with the item representation to calculate the session representation.

3. The conversation recommendation method based on denoising edge perception network according to claim 2, characterized in that: The expression of the session representation is as follows: in, Represents a session The conversational representation of Indicates the length of the session, Indicates the first projects, Indicates the i The weight of the project, Represents a session No. i Project representation, Represents a session User's i Interaction information, 、 、 、 denote learnable parameters; , , express d A set of points in a dimensional real space, express The set of matrices of real numbers, d Indicates the dimension.

4. The conversation recommendation method based on denoising edge perception network according to claim 1, characterized in that The expression of the one-hot encoded vector of the ground truth is as follows: in, The one-hot encoded vector representing the ground truth, represents the normalized exponential function, Represents a session The conversational representation of represents the item embedding matrix.

Images