A user intention enhanced hypergraph conversation recommendation method

Abstract

The present invention relates to the field of personalized recommendation technology, and specifically to a user intent enhanced hypergraph session recommendation method. A hypergraph is established by connecting all items in a session through hyperedges, node features are obtained through a convolutional neural network, the node features are normalized to obtain normalized node features, session features are calculated through normalized node features, and the weight coefficient of each item is further calculated. The general intent feature of the entire session is calculated through the weight coefficient, and the predicted probability of each item becoming the next item in the current session is predicted in combination with the general intent. The items are displayed in the current session according to the predicted probability from large to small.

Description

Technical Field

[0001] The present invention relates to the technical field of personalized recommendation, and in particular to a user intention enhanced hypergraph conversation recommendation method. Background Art

[0002] Early studies on conversational recommendation algorithms mainly relied on neighbor nodes. Some of these studies used cosine similarity to calculate similarity scores, but these methods ignored the transition patterns between items. Subsequently, many sequential methods that use time sequence to simulate user intent have been proposed. For example, Markov chains and personalized matrix factorization are used to capture users' sequential patterns and long-term preferences, but Markov chain-based methods usually focus on the transitions of adjacent items and are difficult to capture more complex and higher-order sequential relationships.

[0003] In addition, existing technologies overly rely on the strict order between items, which may lead to overfitting problems and fail to capture the relationship between non-adjacent items well. Although graph neural networks (GNNs) have advantages in modeling complex transformation relationships between nodes, their ability to capture higher-order relationships beyond adjacency is limited, resulting in insufficient performance in complex conversational recommendation tasks. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides the following technical solutions: a user intention enhanced hypergraph conversation recommendation method, comprising the steps.

[0005] S1. Connect all items in the session through hyperedges to build a hypergraph.

[0006] S2. Obtain the high-order relationship between all items in the conversation in the hypergraph through a convolutional neural network to obtain node features, normalize the node features to obtain normalized node features, connect the normalized node features with the position information of the corresponding items, perform linear changes through the learning weight matrix, apply a nonlinear activation function plus a learnable bias vector to obtain normalized node features containing position information, fuse all normalized node features in the conversation to obtain the conversation features of the entire conversation, obtain the weight coefficient of each item through the soft attention mechanism algorithm combined with the normalized node features containing position information and the conversation features, calculate the general intent features of the entire conversation through the weight coefficient combined with all normalized node features in the entire conversation using the general intent calculation formula, obtain the candidate score of each item according to the general intent through the inner product calculation formula, and use the softmax function combined with the candidate score to calculate the predicted probability of each item becoming the next item in the current conversation.

[0007] S3. The probability of each item becoming the next item in the current session is optimized through the cross entropy loss function.

[0008] S4. Through the self-supervised contrastive learning method, obtain the positive sample pairs consisting of the general intent features of the specific item in the same session and the node features of the next item, and obtain the negative sample pairs consisting of the general intent features of the specific item and the node features in different sessions, combine with the noise contrast estimation method, calculate the binary cross entropy loss between the positive sample pairs and the negative sample pairs, combine the binary cross entropy loss with the cross entropy loss to establish a joint framework to calculate the joint loss, until the joint loss meets the convergence criterion;

[0009] S5. Display each item that meets the convergence criteria in a set position in the current session in a set manner in order of predicted probability from large to small.

[0010] Preferably, in step S2, the calculation formula for obtaining node features through the convolutional neural network is: , where Indicates that the tth node is in ( ) layer node features, For the The node and The correlation strength of a hyperedge, For the The node and The correlation strength of a hyperedge, For the The weight of a hyperedge, is the i-th node in Node features of the layer.

[0011] The calculation formula for normalizing the node features to obtain the normalized node features is: , where D is the degree matrix of the nodes in the hypergraph, is the degree matrix of the hyperedges in the hypergraph. is the incidence matrix of the hypergraph, is the weight matrix of the hyperedge, For the The collection of all project features of a layer, Represents the aggregation process of features from nodes to hyperedges. is the normalized node feature.

[0012] The calculation formula for obtaining the normalized node features containing position information is: , where is the normalized node feature containing position information; For the Normalized node features of each project, is the total number of items in a session, For the Items in total Reverse position information of items in the session, and are learnable parameters.

[0013] The calculation formula for obtaining session features is: , where is the session feature of a session containing m items.

[0014] The calculation formula for calculating the weight coefficient of each item through the soft attention mechanism is: , where For the The weight coefficient of each item, , , , These are all learnable parameters.

[0015] The calculation formula for calculating the general intent feature is , where It is the general intent feature of the entire current session.

[0016] The formula for calculating the candidate score for each item is , where Score the candidate. is the transpose of the general intent feature.

[0017] The formula for calculating the predicted probability of each item being the next item in the current session is , where For each item, the predicted probability of being the next item in the current session.

[0018] Preferably, in step S3, the calculation formula of the cross entropy loss function is: , where is the true label, is the cross entropy loss.

[0019] Preferably, in step S4, the calculation formula for calculating the binary cross entropy loss between the positive sample pair and the negative sample pair is combined with the noise contrast estimation method: , where is the binary cross entropy loss between positive and negative sample pairs, To obtain a positive sample pair consisting of the general intent feature of a specific item in the same session and the node feature of the next item, To obtain the negative sample pairs composed of the user's general intention features for a specific item and the node features in different sessions , sim() is the similarity calculation function, is the sigmoid activation function.

[0020] Preferably, the calculation formula of the joint framework in step S4 is: , where is a hyperparameter used to control the size of the self-supervised task, For joint losses.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] (1) All items in the session are connected through hyperedges to build a hypergraph, node features are obtained through convolutional neural networks, node features are normalized to obtain normalized node features, session features are calculated through normalized node features, and the weight coefficient of each item is further calculated. The general intention feature of the user is calculated through the weight coefficient, and the predicted probability of each item becoming the next item in the current session is predicted based on the general intention.

[0023] (2) By integrating the cross-entropy loss function with the self-supervised contrastive learning method, a joint framework of binary cross-entropy loss function and cross-entropy loss function is established to further improve the prediction accuracy of the predicted probability.

[0024] Other advantages, objectives and features of the present invention will be set forth in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the process of the present invention;

[0026] Figure 2 It is a schematic diagram of the model of the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] See also Figure 1-Figure 2 , the present invention provides a user intention enhanced hypergraph session recommendation method, comprising the steps.

[0029] S1. Connect all items in a session through hyperedges to build a hypergraph, so that the relationship between all items in the same session does not depend on the order of the items.

[0030] S2. Obtain the high-order relationship between the items of all sessions in the hypergraph through the convolutional neural network, obtain the node features, and obtain the normalized node features after normalizing the node features. The calculation formula for obtaining the node features through the convolutional neural network is: , where Indicates that the tth node is in ( ) layer node features, For the The node and The correlation strength of a hyperedge, For the The node and The correlation strength of a hyperedge, For the The weight of a hyperedge, is the i-th node in Node features of the layer.

[0031] The calculation formula for normalizing the node features to obtain the normalized node features is: , where D is the degree matrix of the nodes in the hypergraph, is the degree matrix of the hyperedges in the hypergraph. is the incidence matrix of the hypergraph, is the weight matrix of the hyperedge, For the The collection of all project features of a layer, Represents the aggregation process of features from nodes to hyperedges. is the normalized node feature.

[0032] The normalized node features are connected to the location information of the corresponding items, and the linear changes are made through the learning weight matrix. The nonlinear activation function is applied plus the learnable bias vector to obtain the normalized node features containing the location information. The calculation formula for obtaining the normalized node features containing the location information is: , where is the normalized node feature containing position information, For the Normalized node features of items, is the total number of items in a session, For the Items in total The reverse position information of the item in the session is the The position information of each item in the session, and are learnable parameters.

[0033] All normalized node features in the session are fused to obtain the session features of the entire session. The calculation formula for obtaining the session features is: , where is the session feature of a session containing m items.

[0034] The weight coefficient of each item is obtained by combining the normalized node features with position information and the session features through the soft attention mechanism algorithm. The calculation formula for calculating the weight coefficient of each item through the soft attention mechanism is: , where For the The weight coefficient of each item, , , , These are all learnable parameters.

[0035] The general intent feature of the entire session is calculated using the general intent calculation formula by combining the weight coefficients of all items with all the corresponding normalized node features in the entire session. The calculation formula for calculating the general intent feature is: , where It is the general intent feature of the entire current session.

[0036] The candidate score of each item is calculated based on the general intent features of the current session through the inner product calculation formula. The calculation formula for calculating the candidate score of each item is: , where Score the candidate. is the transpose of the general intent feature.

[0037] The softmax function is used in combination with the candidate scores to calculate the predicted probability of each item becoming the next item in the current session. The formula for calculating the predicted probability of each item becoming the next item in the current session is: , where For each item, the predicted probability of being the next item in the current session.

[0038] S3. Obtain the predicted probability of each item becoming the next item After that, the predicted probability of each item becoming the next item in the current session is optimized by the cross entropy loss function. The calculation formula of the cross entropy loss function is: , where is the true label, is the cross entropy loss, The smaller the value, the true label. With predicted probability The smaller the difference between them, the better the prediction probability. to make it more accurate.

[0039] S4. Through the self-supervised contrastive learning method, we obtain the positive sample pairs consisting of the general intent features of a specific item in the same session and the node features of the next item. , and obtain negative sample pairs consisting of general intent features of specific items and node features in different sessions , combined with the noise contrast estimation method, the binary cross entropy loss between the positive sample pairs and the negative sample pairs is calculated. The calculation formula for the binary cross entropy loss is: , where is the binary cross entropy loss between positive sample pairs and negative sample pairs, sim() is the similarity calculation function, is the sigmoid activation function.

[0040] By combining the binary cross entropy loss with the above cross entropy loss to establish a joint framework, the prediction probability is further improved. The accuracy of the joint framework is calculated as follows: , where is a hyperparameter used to control the size of the self-supervision task, with a value range of 0-1, including this number. For joint losses, The smaller the prediction probability, The more accurate, the next step is executed when the joint loss value reaches the convergence criterion.

[0041] In the joint loss calculation of the cross entropy loss function, binary cross entropy loss function and joint framework, training sets and test sets are generated at the same time. When the loss value of the training set drops to a certain level and enters a fluctuating state, the performance of the test set no longer improves after multiple rounds of training. That is, the training set is set to stop training when the convergence criterion is reached. The loss value at this time is the final loss value.

[0042] S5. Display each item that meets the convergence criteria in a set position in the current session in a set manner in order of predicted probability from large to small. The greater the predicted probability, the greater the probability that the item will become the next item in the current session. The specific display position in the current session, such as below or to the side of the current session, is manually adjusted according to actual conditions.

[0043] On the basis of the above implementation scheme, the processors, modules, corresponding control programs, algorithm programs and other supporting technologies mentioned in the present invention can be implemented in combination with existing electrical technology, information technology, software technology and general protocols. They are not within the scope of protection required by the present invention and will not be described in detail in this application.

[0044] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be considered as exemplary and non-restrictive in all respects.

Claims

1. A user intention enhanced hypergraph conversation recommendation method, characterized in that: Includes steps: S1. Connect all items in the session through hyperedges to build a hypergraph; S2. Obtain the high-order relationship between all items in the conversation in the hypergraph through a convolutional neural network to obtain node features, normalize the node features to obtain normalized node features, connect the normalized node features with the location information of the corresponding items, perform linear changes through a learning weight matrix, apply a nonlinear activation function plus a learnable bias vector to obtain normalized node features containing location information, fuse all normalized node features in the conversation to obtain the conversation features of the entire conversation, obtain the weight coefficient of each item through a soft attention mechanism algorithm in combination with the normalized node features containing location information and the conversation features, calculate the general intent features of the entire conversation through the weight coefficient in combination with all normalized node features in the entire conversation using a general intent calculation formula, obtain the candidate score of each item according to the general intent through an inner product calculation formula, and calculate the predicted probability of each item becoming the next item in the current conversation using a softmax function in combination with the candidate score; S3. Optimizing the probability of each item becoming the next item in the current session by a cross entropy loss function; S4. Obtain a positive sample pair consisting of the general intent feature of a specific item in the same session and the node feature of the next item through a self-supervised contrastive learning method, and obtain a negative sample pair consisting of the user general intent feature of a specific item and the node feature in different sessions, and calculate the binary cross entropy loss between the positive sample pair and the negative sample pair in combination with a noise contrast estimation method, and combine the binary cross entropy loss with the cross entropy loss to establish a joint framework to calculate the joint loss until the joint loss meets the convergence criterion; S5. Display each item that meets the convergence criteria in a set position in the current session in a set manner in order of predicted probability from large to small.

2. The user intention enhanced hypergraph conversation recommendation method according to claim 1, characterized in that: The calculation formula for obtaining node features through the convolutional neural network in step S2 is: , where Indicates that the tth node is in ( ) layer node features, For the The node and The correlation strength of a hyperedge, For the The node and The correlation strength of a hyperedge, For the The weight of a hyperedge, is the i-th node in Node characteristics of the layer; The node features are normalized to obtain the calculation formula of the normalized node features: , where D is the degree matrix of the nodes in the hypergraph, is the degree matrix of the hyperedges in the hypergraph, is the incidence matrix of the hypergraph, is the weight matrix of the hyperedge, For the The collection of all project features of a layer, represents the aggregation process of features from nodes to hyperedges, is the normalized node feature; The calculation formula for obtaining the normalized node features containing position information is: , where is the normalized node feature containing position information; For the Normalized node features of items, is the total number of items in a session, For the Items in total Reverse position information of items in the session, and is a learnable parameter; The calculation formula for obtaining the session features is: , where is the session feature of a session containing m items; The calculation formula for calculating the weight coefficient of each item through the soft attention mechanism is: , where For the The weight coefficient of each item, , , , are all learnable parameters, is the sigmoid activation function; The calculation formula for calculating the general intention feature is: , where It is the general intention feature of the entire current session; The calculation formula for calculating the candidate score of each item is: , where Score the candidate. is the transposition of the general intention feature; The formula for calculating the predicted probability of each item becoming the next item in the current session is: , where For each item, the predicted probability of being the next item in the current session.

3. The user intention enhanced hypergraph session recommendation method according to claim 2, characterized in that: The calculation formula of the cross entropy loss function in step S3 is: , where is the true label, is the cross entropy loss.

4. The user intention enhanced hypergraph conversation recommendation method according to claim 3, characterized in that: In step S4, the calculation formula for calculating the binary cross entropy loss between the positive sample pair and the negative sample pair is combined with the noise contrast estimation method: , where is the binary cross entropy loss between positive and negative sample pairs, The positive sample pairs are formed by obtaining the general intent feature of a specific item in the same session and the node feature of the next item. To obtain the negative sample pairs consisting of the general intent features of a specific project and the node features in different sessions , sim() is the similarity calculation function, is the sigmoid activation function.

5. The user intention enhanced hypergraph conversation recommendation method according to claim 4, characterized in that: In step S4, the calculation formula of the joint framework is: , where is a hyperparameter used to control the size of the self-supervised task, For joint losses.

Images