A Dynamic Graph Neural Network Recommendation Method Based on Cellular Automata

Abstract

The present invention discloses a dynamic graph neural network recommendation method based on cellular automata. A user-item interaction dynamic graph is constructed according to user-item interaction information, and the interaction dynamic graph is partitioned into a user-perspective dynamic graph and an item-perspective dynamic graph. The user-perspective dynamic graph and the item-perspective dynamic graph are respectively learned. The process for the user-perspective dynamic graph is user interest deviation, user interest enhancement, calculation of the offset of the change in user cell state, and user interest aggregation. The learning process for the item-perspective dynamic graph is item neighbor information aggregation, judgment of item cell state information, and update of item embedding information. The user-item embedding representation is used to predict the representation at a future time and calculate the interaction probability between the user and the item, and recommendations are made according to the ranking of the interaction probability magnitudes. The present invention well captures the law of continuous evolution of users and items over time in the recommendation scenario, timely captures the changes in user preferences and the update of the information carried by items, finds the items that users are interested in, and makes personalized recommendations.

Description

Technical Field

[0001] The present invention relates to the technical field of computer information recommendation, and particularly to a dynamic graph neural network recommendation method based on cellular automata. Background Art

[0002] A recommendation system is an important tool to help users discover content they are interested in and relieve information overload. In recent years, with the continuous development of online recommendation services, dynamic recommendation has become an important research direction. On the one hand, users' preferences for items evolve over time; on the other hand, with the emergence of new items, the representation and popularity of items also change continuously. Therefore, dynamic modeling based on time is crucial. Previous work mainly focused on the static graph perspective, paid attention to the interaction sequences between users and items, and learned the embedding representations of users and items. However, in actual recommendation scenarios, users and items interact dynamically over time, and this collaborative information of dynamic interaction has an important impact on the performance of the recommendation system. At the same time, due to the different positions of users and items in the real recommendation scenario, users are in an active state when purchasing items; items are often in a passive state of being selected. There are the following three problems to be solved in existing dynamic graph neural network recommendation methods:

[0003] (1), Mostly use discrete dynamic graphs to learn node representations in the change of time;

[0004] (2), Users and items are in different positions in the recommendation and need to be treated separately;

[0005] (3), The representation learning of dynamic graph neural networks should be fine-grained learning according to the interaction moment. Summary of the Invention

[0006] The purpose of the present invention is to make up for the deficiencies of the existing technology and provide a dynamic graph neural network recommendation method based on cellular automata.

[0007] The present invention is realized through the following technical solutions:

[0008] A dynamic graph neural network recommendation method based on cellular automata specifically includes the following steps:

[0009] Step 1, construct a user-item interaction dynamic graph according to the interaction information of users and items over time;

[0010] Step 2, divide the interaction dynamic graph obtained in Step 1 into a user perspective dynamic graph and an item perspective dynamic graph;

[0011] Step 3, learn the user perspective dynamic graph obtained in Step 2, and the process is user interest deviation, user interest enhancement, calculation of the offset amount of user cell state change, and user interest aggregation;

[0012] Step 4: Learn the dynamic graph of the item perspective obtained in Step 2. The process includes item neighbor information aggregation, item cell state information judgment, and item embedding information update;

[0013] Step 5: Determine the objective function based on the user and item embedding representations obtained in Step 3 and Step 4, and make recommendations according to the prediction results.

[0014] In Step 1, construct a user-item interaction dynamic graph based on the interaction information of users and items over time. The process is as follows:

[0015] Construct the interaction at time t of the user and the item into a dynamic graph G t =(V t , E t ), where V t and E t are the node set and edge set in the dynamic graph G t respectively.

[0016] In Step 2, divide the interaction dynamic graph obtained in Step 1 into a user perspective dynamic graph and an item perspective dynamic graph. The specific process is as follows:

[0017] Centering on the user and item interacting at the current moment, with neighbor nodes being the nodes interacted at the current moment and before, divide the dynamic graph G t into a user perspective dynamic graph and an item perspective dynamic graph respectively.

[0018] In Step 3, learn the user perspective dynamic graph obtained in Step 2. The process includes user interest deviation, user interest enhancement, calculation of the user cell state change deviation, and user interest aggregation. The specific process is as follows:

[0019] Step 3.1: Take the user node interaction node information at the current moment as input, calculate the user interest deviation representation, and compare it with the representation at the previous moment to obtain the user interest deviation state difference. The calculation formula is as follows:

[0020]

[0021] is the parameter matrix, is the state context. φ u is the activation function. are the representations of the item at time t - and the item's own characteristics respectively. Update the embeddings of the user and the item by propagating the current interaction information. γ 1 represents the state difference of the user during interest deviation.

[0022] Step 3.2. Aggregate the second-order neighbor information of the user node at the current moment for user interest enhancement learning. The basic principle of second-order aggregation is to learn the collaborative relationship between users and items. The calculation formula is as follows:

[0023]

[0024] ζ is the aggregation method. Here, attention aggregation is used, and γ 2 represents the state difference of the user during interest enhancement.

[0025] Step 3.3. Calculate the historical preferences that the user node needs to inherit at the current moment and its own new features, and perform attention aggregation on the representations obtained in Steps 3.1 and 3.2. The calculation formula is as follows:

[0026]

[0027] Δt is the time interval between the current moment t and the previous interaction time t - θ u is the activation function, and w 0 is the parameter vector of the time interval Δt.

[0028]

[0029] h (u,t) is the user node embedding after the interaction update between user u and item v at time t, and F u is the aggregation function. Here, the sigmod activation function is used.

[0030] In Step 4, learn the item perspective dynamic graph obtained in Step 2. The process is item neighbor information aggregation, item cell state information judgment, and item embedding information update. The specific process is as follows:

[0031] Step 4.1. Aggregate the information of the user set that interacts with item v at the same time at the current moment t. The calculation formula is as follows:

[0032]

[0033] is the parameter matrix, and φ v is the activation function.

[0034] Step 4.2. Combine the item neighbor information before time t and the neighbor information of item node v at the current moment to determine the current state of the item node. The determination process is shown in Equations 8 and 9:

[0035]

[0036] W sis the status context parameter, which converts the change of item information into a status parameter.F normalize is the normalization function. When the judgment result ρ t ≥ k, the item status changes, updates the interest offset information carried by the item, and proceeds to the next update of the item embedding information; when the judgment result ρ t < k, the item status information does not change, and the item embedding representation only updates its own historical information.

[0037] Step 4.3: Aggregate the second-order neighbor information of the item node at the current moment to enhance the interest preference carried by the item. The calculation formula is as follows:

[0038]

[0039] ζ v is the aggregation method, using attention aggregation.

[0040] Step 4.4: Calculate the historical preference that the item node at the current moment needs to inherit and its own new features, and perform aggregation calculation based on the information obtained in Steps 4.1 - 4.3 to obtain the final representation of the item node. The calculation formula is as follows:

[0041]

[0042] θ v is the aggregation function.

[0043]

[0044] is the parameter that controls the influence degree of the update mechanism.F v is the aggregation function.

[0045] In Step 5, determine the objective function based on the user-item embedding representations obtained in Steps 3 and 4, and make recommendations according to the prediction results. The specific process is as follows:

[0046] Step 5.1: Predict the future embedding representations of the user and the item, and use to represent the predicted future embedding representation of the user. The calculation formula is as follows:

[0047]

[0048] is the time context parameter, which is used to convert the time interval into a vector, is a vector with all elements being 1.t + is the future moment when the user interacts with the next item;

[0049] Step 5.2: Calculate the future status information of the item according to the predicted future embedding representation of the user. The calculation formula is as follows:

[0050]

[0051] Step 5.3: Determine the objective function based on the future embedding representation of the user and the future state information of the item, and perform loss calculation. The calculation formula is as follows:

[0052]

[0053] represents the interaction time of user items arranged in chronological order. λ u , λ v is the smoothing coefficient, which restricts the update of user embedding and item state.

[0054] The advantages of the present invention are as follows: By using the continuous-time dynamic graph learning method, the present invention grasps the time fine-grained information and updates the user and item node information in a timely manner. At the same time, in order to prevent the overfitting effect of fine-grained time learning on the update of user and item representations, the idea of cellular automata is introduced, and different transition rules are formulated according to the different states of users and items. Finally, the user and item embeddings at future moments are predicted, so as to perform personalized recommendation and improve the accuracy of the recommendation results. Brief Description of the Drawings

[0055] Figure 1 is the overall architecture diagram of the model of the present invention.

[0056] Figure 2 is the user representation learning graph of the present invention.

[0057] Figure 3 The item representation learning graph of the present invention. Detailed Embodiment

[0058] The present invention will be further described below with reference to the drawings and embodiments.

[0059] A dynamic graph neural network recommendation method based on cellular automata according to the present invention, as shown in Figure 1 , 2 , 3, divides the constructed dynamic graph, uses the continuous dynamic graph learning method to learn the user interest deviation information, user interest enhancement information and item neighbor information, and then performs state transition calculations on the user and the item respectively according to the specified state transition rules to obtain the final embedding representations of the user and the item. The transition rules of the user and the item are different according to the different positions of the user and the item. The user state change is continuous, and the item state change is discrete. Finally, the embedding representations and state information of the user and the item at future moments are predicted, and the user-item interaction probability is calculated based on the prediction results and sorted, and the items with higher probabilities are selected to generate a recommendation list.

[0060] The present invention specifically includes the following steps:

[0061] Step 1: Construct a user-item interaction dynamic graph based on the interaction information of users and items over time. The specific process is as follows:

[0062] Construct an interaction at time t between a user and an item into a dynamic graph G t =(V t , E t ), where V t and E t are the node set and edge set in the dynamic graph G t respectively.

[0063] Step 2: Partition the interaction dynamic graph obtained in Step 1 into a user-perspective dynamic graph and an item-perspective dynamic graph. The specific process is as follows:

[0064] Centering on the users and items interacting at the current moment, with the neighbor nodes being the nodes interacted at the current moment and before, partition the dynamic graph G t into a user-perspective dynamic graph and an item-perspective dynamic graph respectively.

[0065] Step 3: Learn from the user-perspective dynamic graph obtained in Step 2. The process includes user interest deviation, user interest enhancement, calculation of the user cell state change deviation, and user interest aggregation. The specific process is as follows:

[0066] Step 3.1: Take the interaction node information of the user node at the current moment as input, calculate the user interest deviation representation, and compare it with the representation at the previous moment to obtain the user interest deviation state difference. The calculation formula is as follows:

[0067]

[0068] is the parameter matrix, is the state context. φ u is the activation function. f v are the representations of the item at time t - and the item's own features respectively. Update the embeddings of the user and the item by propagating the current interaction information. γ 1 represents the state difference of the user during interest deviation;

[0069] Step 3.2: Aggregate the second-order neighbor information of the user node at the current moment for learning user interest enhancement. The basic principle of second-order aggregation is to learn the collaborative relationship between users and items. The calculation formula is as follows:

[0070]

[0071] ζ is the aggregation method. Here, attention aggregation is used. γ 2Indicates that the user's state is poor when their interest increases.

[0072] Step 3.3: Calculate the historical preferences that the user node needs to inherit at the current moment and its own new features, and perform attention aggregation on the representations obtained in Steps 3.1 and 3.2. The calculation formula is as follows:

[0073]

[0074] Δt is the time interval between the current moment t and the previous interaction time t - and θ u is the activation function, and w 0 is the parameter vector of the time interval Δt.

[0075]

[0076] h (u,t) is the user node embedding after the interaction update between user u and item v at time t, and F u is the aggregation function, and here the sigmod activation function is used.

[0077] Step 4: Learn the item perspective dynamic graph obtained in Step 2. The process is item neighbor information aggregation, item cell state information judgment, and item embedding information update. The specific process is as follows:

[0078] Step 4.1: Aggregate the information of the set of users who interact with item v at the same time at the current moment t. The calculation formula is as follows:

[0079]

[0080] is the parameter matrix, and φ v is the activation function.

[0081] Step 4.2: Combine the item neighbor information before time t and the neighbor information of item node v at the current moment to determine the current state of the item node. The determination process is shown in Equations 8 and 9:

[0082]

[0083] W s is the state context parameter, which converts the change in item information into a state parameter, and F normalize is the normalization function. When the determination result ρ t ≥k, the item state changes, and the interest offset information carried by the item is updated for the next item embedding information update; when the determination result ρ t <k, the item state information does not change, and the item embedding representation only updates its own historical information.

[0084] Step 4.3. Aggregate the second-order neighbor information of the item node at the current moment to enhance the item's carrying interest preference. The calculation formula is as follows:

[0085]

[0086] ζ v is the aggregation method, and attention aggregation is adopted.

[0087] Step 4.4. Calculate the historical preference that the item node needs to inherit and its own new features at the current moment, and perform aggregation calculation based on the information obtained in Steps 4.1 - 4.3 to obtain the final representation of the item node. The calculation formula is as follows:

[0088]

[0089] θ v is the aggregation function;

[0090]

[0091] is the parameter that controls the influence degree of the update mechanism, and F v is the aggregation function.

[0092] Step 5. Determine the objective function based on the user-item embedding representations obtained in Steps 3 and 4, and make recommendations according to the prediction results. The specific process is as follows:

[0093] Step 5.1. Predict the future embedding representations of the user and the item. Use to represent the predicted future embedding representation of the user. The calculation formula is as follows:

[0094]

[0095] is the time context parameter, which is used to convert the time interval into a vector, is a vector with all elements being 1, and t + is the future moment when the user interacts with the next item.

[0096] Step 5.2. Calculate the future state information of the item based on the predicted future embedding representation of the user. The calculation formula is as follows:

[0097]

[0098] Step 5.3. Determine the objective function based on the user's future embedding representation and the item's future state information, and perform loss calculation. The calculation formula is as follows:

[0099]

[0100] Denote the user-item interaction time arranged in chronological order, λ u , λ v is the smoothing coefficient, which constrains the user embedding and item state update.

[0101] The embodiments described in the present invention are only descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. Without departing from the design idea of the present invention, various modifications and improvements made by those skilled in the art to the technical solutions of the present invention should fall within the protection scope of the present invention. The technical content claimed by the present invention has been fully recorded in the claims.

Claims

1. A dynamic graph neural network recommendation method based on cellular automata, characterized in that: Specifically, it includes the following steps: Step 1: Construct a user-item interaction dynamic graph based on the interaction information of users and items over time; Step 2: Divide the interaction dynamic graph obtained in Step 1 into a user-perspective dynamic graph and an item-perspective dynamic graph; Step 3: Learn the user-perspective dynamic graph obtained in Step 2. The process includes user interest deviation, user interest enhancement, calculation of the deviation amount of user cell state change, and user interest aggregation; Step 4: Learn the item-perspective dynamic graph obtained in Step 2. The process includes item neighbor information aggregation, item cell state information judgment, and item embedding information update; Step 5: Determine the objective function based on the user and item embedding representations obtained in Step 3 and Step 4, and make recommendations according to the prediction results; The learning of the user-perspective dynamic graph obtained in Step 2 in Step 3, the process includes user interest deviation, user interest enhancement, calculation of the deviation amount of user cell state change, and user interest aggregation. The specific process is as follows: Step 3.1: Take the interaction node information of the user node at the current moment as input, calculate the user interest deviation representation, and compare it with the representation at the previous moment to obtain the user interest deviation state difference. The calculation formula is as follows: is the parameter matrix, is the state context, φ u is the activation function, h (v,t-) , f v are respectively the representation of the item at time t - and the characteristics of the item itself; Updating the embeddings of users and items by propagating the current interaction information, γ 1 indicating the state difference of the user when there is an interest deviation; Step 3.2: Aggregate the second-order neighbor information of the user node at the current moment for the learning of user interest enhancement. The second-order aggregation is to learn the collaborative relationship between users and items. The calculation formula is as follows: ζ is an aggregation method that uses attention aggregation, γ 2 represents the state difference of the user when the interest is enhanced; Step 3.3: Calculate the historical preferences that the user node at the current moment needs to inherit and its own new features, and perform attention aggregation on them with the representations obtained in Step 3.1 and Step 3.

2. The calculation formula is as follows: Δt is the time interval between the current moment t and the previous interaction time t - , θ u is the activation function, w 0 is the parameter vector of the time interval Δt, is the parameter matrix, f u is the user's own characteristics, h (u,t) is the user node embedding after the interaction update between user u and item v at time t, and F u is the aggregation function, using the sigmod activation function.

2. A dynamic graph neural network recommendation method based on cellular automata according to claim 1, characterized in that: The construction of the user-item interaction dynamic graph according to the interaction information of users and items over time in Step 1. The specific process is as follows: Construct the interaction between the user and the item at time t into a dynamic graph G t =(V t , E t ), where V t , E t are the node set and edge set in the dynamic graph G t respectively.

3. A dynamic graph neural network recommendation method based on cellular automata according to claim 2, characterized in that: The division of the interaction dynamic graph obtained in Step 1 into a user-perspective dynamic graph and an item-perspective dynamic graph in Step 2. The specific process is as follows: Centered around the users and items interacting at the current moment, with neighbor nodes being the nodes interacted at the current moment and before, partition the dynamic graph G t into a user-perspective dynamic graph and an item-perspective dynamic graph respectively.

4. A dynamic graph neural network recommendation method based on cellular automata according to claim 1, characterized in that: The learning of the item-perspective dynamic graph obtained in Step 2 in Step 4. The process includes item neighbor information aggregation, item cell state information judgment, and item embedding information update. The specific process is as follows: Step 4.1: Aggregate the information of the set of users that interact with item v at the same time at the current moment t. The calculation formula is as follows: is the parameter matrix, φ v is the activation function, is the set of users who purchase item v at time t, and n represents the number of users; Step 4.2: Combine the item neighbor information before time t and the neighbor information of the item node v at the current moment to determine the current state of the item node. The determination process is shown in equations (8) and (9): W s is a status context parameter that converts changes in item information into status parameters, ρ t represents the value of the status mapping of item v at time t, k is the threshold for the status change of item v, F normalize is a normalization function. When the judgment result ρ t ≥ k, the item status changes, the interest offset information carried by the item is updated, and the item embedding information is updated in the next step; when the judgment result ρ t < k, the item status information does not change, and the item embedding representation only updates its own historical information; Step 4.3: Aggregate the second-order neighbor information of the item node at the current moment to enhance the interest preference carried by the item. The calculation formula is as follows: ζ v is an aggregation method that uses attention aggregation; Step 4.4: Calculate the historical preferences that the item node needs to inherit at the current moment and its own new features, and perform an aggregation calculation based on the information obtained in Steps 4.1 - 4.3 to obtain the final representation of the item node. The calculation formula is as follows: θ v is an aggregation function. A parameter for controlling the influence degree of the update mechanism, F v is an aggregation function.

5. A dynamic graph neural network recommendation method based on cellular automata according to claim 4, wherein: The target function is determined based on the user and item embedding representations obtained in Steps 3 and 4 in Step 5, and recommendations are made according to the prediction results. The specific process is as follows: Step 5.

1. Predict the future embedded representations of users and items. Use to represent the predicted future embedded representation of the user. The calculation formula is as follows: is a time context parameter for converting a time interval into a vector, is a vector with all elements being 1, t + is the future moment when the user interacts with the next item, MLP u represents a fully connected network that calculates the user's future moment representation; Step 5.2: Calculate the future state information of the item based on the predicted future embedding representation of the user. The calculation formula is as follows: Indicates the state of the predicted item v at time t + ; Step 5.3: Determine the target function based on the future embedding representation of the user and the future state information of the item, and perform loss calculation. The calculation formula is as follows: represents the user-item interaction time arranged in chronological order, λ u , λ v is the smoothing coefficient that constrains the user embedding and item state update.

Images