Interpretable sequence recommendation method integrating time perception and path reasoning

Abstract

The invention discloses an interpretable sequence recommendation method fusing time perception and path reasoning. The method comprises the following steps: 1) constructing a time sequence collaborative knowledge graph; 2) node representation learning is carried out on the time sequence collaborative knowledge graph through TransE; 3) designing a time sensing system, updating an article vector at the current moment through a path and an article attention mechanism, and generating an article prediction vector in combination with a user vector; 4) designing a path reasoning system, performing correlation path reasoning by using a reinforcement learning framework, and obtaining a multi-path vector through path representation learning; and 5) model training: training the time sensing system and the path reasoning system to obtain an article prediction result. According to the method, double systems are introduced to simulate the human reasoning process, and the accuracy and interpretability of the recommendation result are improved by learning the time dependence characteristics of the user behaviors in the time sequence collaborative knowledge graph.

Description

Technical Field

[0001] The present invention belongs to the fields of knowledge graphs, graph neural networks, recommendation systems, etc., and particularly provides an explainable sequence recommendation method that integrates time perception and path reasoning. Background Art

[0002] With the rapid development of Internet technology, the amount of information available to users has increased dramatically, leading to an increasingly serious problem of "information overload". How to find content that users are interested in among the massive amount of information and improve the efficiency of information acquisition has become an urgent problem to be solved. The recommendation system analyzes the user's behavioral preferences and intelligently selects the content that users are most interested in, thereby improving the user experience.

[0003] A knowledge graph is a semantic network whose nodes represent entities or concepts and whose edges represent various semantic relationships between entities / concepts. A knowledge graph consists of several triples (h, r, t), where h and t represent the head node and tail node, and r represents the relationship. The recommendation model based on the knowledge graph introduces the knowledge graph (KG) as auxiliary information into the recommendation system, and mines the potential connections between users and objects through the rich semantic associations and structural information in the knowledge graph to improve the performance of recommendations. It mainly includes embedding-based recommendation systems, path-based recommendation systems, and propagation-based recommendation systems.

[0004] Embedding-based recommender systems use knowledge graph embedding algorithms to preprocess KG and incorporate the learned entity embeddings into the recommendation framework. CKE (Proceedings of the 37th international ACM SIGIRconference on Research&development in information retrieval, 2014) combines the CF module with item knowledge embedding, text embedding, and image embedding in a unified Bayesian framework. DKN (IEEE Transactions on Knowledge and Data Engineering, 2017) treats entity embeddings and word embeddings as different channels, and then designs a CNN framework to combine them for news recommendation. KTUP (Proceedings ofthe World Wide Web Conference, 2019) transfers the learned relation embeddings and entity embeddings from KG to TUP, while training the completeness and recommendation tasks of KG. Jointly learn representations of users, items, entities, and relations. Jointly learn TUP and TransH.

[0005] Path-based recommender systems provide additional guidance for recommendations by exploring various connection patterns between items in KG. PER (ACM International Conference on Web Search and Data Mining, 2014) treats KG as a heterogeneous network and extracts meta-path-based latent features to represent the connectivity between users and items along different types of relationship paths. Hete-CF (IEEE International Conference on Data Mining, 2014) calculates the user's preference for unrated items by obtaining a regularization term for the meta-path similarity between various entities. KPRN (AAAIConference on Artificial Intelligence, 2019) enumerates all eligible paths between each user-item pair and then trains a sequence RNN model from the extracted paths to predict the ranking scores of these pairs. PGPR (ACMSIGIR Conference on Research and Development in Information Retrieval, 2019) replaces the enumeration method with a reinforced reasoning method to determine meta-paths suitable for scalability. TMER (ACM International Conference on Web Search and Data Mining, 2021) captures the contextual information of historical item features and path definitions through an attention network, and further models the temporal meta-path.

[0006] The recommendation system based on the propagation method enhances the entity representation by aggregating the features of multi-hop neighbors in the knowledge graph. RippleNet (ACM International Conference on Information and Knowledge Management, 2018) borrowed the "ripple effect" in physics and believed that the user's interest in the product will spread like a ripple. By constructing a user-product interaction graph, the graph neural network is used to learn the potential features of users and products, and simulate the user's interaction process with the product to generate a personalized recommendation list. KGCN (World Wide Web Conference, 2019) uses a graph convolutional network to aggregate the neighbor information of the node, and obtains the final entity vector through continuous iteration to capture high-order information. KGAT (ACM Knowledge Discovery and Data Mining, 2019) uses a graph attention network to assign different weights to relationships to capture user preferences and interests. CKAN (ACM SIGIR Conference on Research and Development in Information Retrieval, 2020) proposes a heterogeneous propagation strategy to encode two types of information and then apply a knowledge-aware attention mechanism to distinguish the contributions of different knowledge-based neighbors.

[0007] However, existing recommendation systems only utilize static knowledge graph feature information, do not pay attention to the dynamic changes of user interests and behaviors, and cannot model the dynamics of user historical behaviors, resulting in inaccurate recommendation results and unconvincing explanations. Although some methods consider temporal information, they are mainly used to enhance data representation rather than reasoning on knowledge graphs, which makes them unable to infer more information from temporal information. Summary of the invention

[0008] In order to overcome the shortcomings of the existing technologies and to improve the accuracy and interpretability of the recommendation system, the present invention proposes an interpretable sequence recommendation system that integrates time perception and path reasoning. Drawing on the dual-process theory of cognitive science, the present invention designs a time perception system and a path reasoning system. The time perception system models temporal dependencies through temporal perception, and the path reasoning system selects appropriate associated paths as recommendation reasons through path reasoning. The two systems work iteratively and complement each other, further improving the accuracy and interpretability of the recommendation results.

[0009] The technical solution adopted by the present invention to solve its technical problem is:

[0010] An explainable sequence recommendation method integrating time perception and path reasoning, the method comprising the following steps:

[0011] (1) Building a time-series collaborative knowledge graph

[0012] According to the historical interaction information of users, the collaborative knowledge graph is sliced ​​to obtain a series of ordered snapshots and a temporal collaborative knowledge graph is established. The temporal collaborative knowledge graph TCKG is expressed as in represents a snapshot of CKG at time τ, where h represents the head entity, t represents the tail entity, and r τ represents the relationship at time τ, ε represents the entity set, Represents a set of relations;

[0013] (2) Learning the representation of knowledge graphs for temporal collaboration

[0014] The TransE translation model is used to embed entities and relations in TCKG. The triples (h,r τ ,t), where h,r τ ,t have d-dimensional embedding, that is, The scoring function is defined as e h +r τ To e t The distance between them is given by the following formula:

[0015]

[0016] Smaller score indicating that the triple is more likely to be correct than vice versa;

[0017] The loss function for the relative order between positive and negative triplets is defined as:

[0018]

[0019] in And (h, r, t′) is the negative triplet after the tail entity is replaced by the random entity t′, σ(·) is the sigmoid function;

[0020] Through training, we get the user vector set {h u}、Item vector set {h v} and the set of relation vectors {h r};

[0021] (3) Designing a time-aware system

[0022] The time-aware system takes the current item vector, user vector, and multiple path vectors as input to generate an item prediction vector;

[0023] (4) Designing a path reasoning system

[0024] The path reasoning system performs associated path reasoning under the guidance of the item prediction vector generated by the time perception system, and then obtains the above-mentioned multiple path vectors through path representation learning. The system uses a reinforcement learning framework to establish a reasoning environment, which provides the policy decision module with information about the current environment state and executable operations in TCKG, and rewards the decision module according to the degree of match between the current policy and user interaction;

[0025] (5) Model training

[0026] The two systems work iteratively and complement each other. The item prediction vector output by the time perception system is passed to the path reasoning system for associated path reasoning. Then, multiple path vectors output by the path reasoning system and the item vector and user vector of the current time step are input into the time perception system to obtain the item prediction vector.

[0027] Get a trained model to provide users with accurate recommendation results.

[0028] Further, the process of (3) is as follows:

[0029] First, multiple path vectors are passed to the time perception system. These paths reflect different purchasing preferences of users. The path attention mechanism is used to aggregate multiple path vectors. The resulting multi-path aggregation vector can better represent the complex relationship between entities. The calculation formula of the path attention mechanism is:

[0030] MultiHead(Q φ ,K φ ,V φ )=Concat(head1,...,head1)W O (3)

[0031]

[0032] Among them head i yes f(·) is the softmax function, Q, K, V are the self-attention variables associated with path φ, W is the weight matrix, d k is the dimension, and Concat(·) is the concatenation operation.

[0033] After obtaining the multi-path aggregation vector, the item attention mechanism is used to combine the current item's own information with the path aggregation vector to integrate the historical time series information and update the item vector at the current moment. The calculation formula of the item attention mechanism is:

[0034]

[0035] where pi represents the initial representation of the current item, φ i-1→i is the path instance from the i-1th item to the i-th item, W is the weight matrix, b is the bias, and g(·) is the ReLU function;

[0036] The user vector and the item vector are concatenated and input into the multi-layer perceptron to generate the item prediction vector at that moment. The calculation formula is:

[0037]

[0038] Since there is no historical time series information before the first item, the present invention concatenates the user vector and the first item vector, inputs the concatenated vector into a multi-layer perceptron, and generates the item prediction vector at that moment. The calculation formula is:

[0039] h1=MLP(Concat(h u ,h1)) (7)

[0040] The generated item prediction vector is input into the path reasoning system for associated path reasoning.

[0041] Furthermore, the process of (4) is as follows:

[0042] (4.1) State: The initial state is s0 = [u,u,φ,φ], and the state s at time step t is t =[u,h t ,e t ,e' t+1 ], where u represents the user entity, h t represents a history of t steps (e0,e1,...,e t ), e t Represents the entity of the current step, e' t+1 represents the entity estimation vector generated by the time-aware system at time step t+1;

[0043] (4.2) Action: At time step t, the action is expressed as where e t+1 Indicates the next entity to be visited, r t+1 Indicates the current entity and e t+1 the relationship between;

[0044] (4.3) Pruning strategy: In order to improve model efficiency and control space complexity, it is necessary to perform effective reasoning while maintaining the maximum out-degree benchmark of the action space. The score function is used to calculate the score for each action, which is defined as follows:

[0045]

[0046] where <·,·> represents the dot product operation, r t′ ∈{r kg ∪r interact} is the relation embedding of t′ jump time, sort all actions in the action space according to the score, and keep the top ranked actions in the pruned action space In, it is defined as:

[0047]

[0048] where ∈ is the predefined action space size, which is used to limit the number of actions after pruning.

[0049] (4.4) State transition: given state s t and action a t As well as the item representation delivered by the time-aware system, the state transition is represented as:

[0050] s k+1 =δ(s t ,a t )=[u,h t+1 ,e t+1 ,e' t+2 ] (10)

[0051] (4.5) Reward function: The goal of reinforcement learning is to find user-item path instances and item-item path instances. In order to evaluate the quality of different paths and the flexibility of selection, soft rewards are used to achieve the above goals. The reward function is defined as:

[0052]

[0053] Where v represents the item entity, V represents the item entity set, and T represents the maximum number of steps for path search; if the target entity is found before T steps, the reward is calculated, otherwise the reward is 0;

[0054] (4.6) Optimization strategy: Parameterize the path reasoning strategy using state information and action space. The state s at time step t t =[u,h t ,e t ,e' t+1 ], including search history h t , using bidirectional LSTM to encode the state vector s t , the formula is as follows:

[0055]

[0056] Where W1 is a linear parameter, and the policy network π outputs in the pruning space The probability of each action in:

[0057]

[0058] Where W a is a linear parameter that needs to be learned. In the policy gradient method, we hope to optimize the strategy by maximizing the cumulative reward. In order to improve the training efficiency, a baseline is introduced to reduce the variance. The value network M(s t ) The state vector x t Mapping to real values ​​in reinforcement learning as a baseline:

[0059] M(s t )=x t W c (14)

[0060] Where W c are the linear parameters that need to be learned. The two networks are trained by maximizing the expected reward of any user u, optimized by the REINFORCE algorithm, which updates the model parameters Θ using the following policy gradient:

[0061]

[0062] Where R T Indicates that from state s t Discounted cumulative rewards from start to end state;

[0063] (4.7) Path representation learning: Obtain candidate user-item and item-item path instances and represent the path data as where p i1 represents the feature vector of path i at time step t, T i is the length of path i. The path data is input into the LSTM layer. LSTM maps the input sequence to an embedding vector of a fixed size. The calculation is shown as follows:

[0064]

[0065] Where W LSTM is the weight matrix, h t-1 is the hidden state of the previous time step, the last hidden state output by the LSTM layer It is represented as a path vector, which is transmitted to the time-aware system as guidance information.

[0066] In (5), the item prediction vector and user vector finally obtained after multiple iterations are input into a multi-layer perceptron to calculate the user-item prediction score:

[0067] r u,v =MLP(Concat(h u ,h v)) (17)The MLP contains two hidden layers, uses ReLU as the activation function, and the Sigmoid function as the output layer;

[0068] Use the implicit feedback loss of negative sampling as the entire loss function:

[0069]

[0070] in Represents the calculation of the expected value on the negative sample set, where P neg represents the distribution of negative samples, and j represents the samples drawn from this distribution.

[0071] The beneficial effects of the present invention are mainly manifested in: In view of the problem that the existing recommendation system ignores historical dynamic time information and lacks interpretability, the present invention proposes an interpretable sequence recommendation system that integrates time perception and path reasoning. By introducing the dual-process theory, it simulates the reasoning process of human thinking, captures the time-dependent characteristics in user purchasing behavior, and further enhances the accuracy and interpretability of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a temporal collaborative knowledge graph.

[0073] Figure 2 It is a schematic diagram of the recommendation system framework. DETAILED DESCRIPTION

[0074] The present invention will be further described below in conjunction with the accompanying drawings.

[0075] Reference Figure 1 and Figure 2 , an explainable sequence recommendation method integrating time perception and path reasoning, comprising the following steps:

[0076] (1) Building a time-series collaborative knowledge graph

[0077] The knowledge graph is sorted and sliced ​​according to the timestamps of the interaction records to form a series of ordered time snapshots, which together constitute the temporal collaborative knowledge graph. For example, Figure 1 The paper presents a temporal collaborative knowledge graph built based on the Amazon Beauty dataset, in which each interaction record is accompanied by a timestamp, such as (B001MW1MK2, A2GJX2KCUSR0EI, 5, 1422316800). The interaction records are sorted by these timestamps and sliced ​​according to the user's temporal interaction records to generate a time snapshot.

[0078] (2) Learning the representation of temporal collaborative knowledge graphs

[0079] TransE is used to learn the representation of the temporal collaborative knowledge graph. First, the parameters, relationship vectors, and entity vectors are initialized using normal distribution. Then, during the iterative training process, the closeness between the head entity vector plus the relationship vector and the tail entity vector is evaluated according to formula (1). Then, based on the comparison result between the energy value and the preset threshold, the entity and relationship vectors are updated, and these vectors are optimized using formula (2). In each iteration, the vector is updated using the gradient descent method, and the training is stopped when the preset number of iterations is reached.

[0080] (3) Designing a time-aware system

[0081] Input multiple path vectors generated by the path reasoning system, the item vector at the current moment, and the user vector. Input multiple path vectors into the path attention mechanism network, obtain the multi-path aggregation vector according to formulas (3) and (4), input the learned multi-path aggregation vector and item vector into the item attention mechanism network, learn the temporal characteristics of the item according to formula (5), input the updated n-dimensional feature vector and user vector into the MLP layer according to formula (6), generate the item prediction vector and transmit it to the path reasoning system.

[0082] In particular, since there is no historical time series information before the first item, according to formula (7), the user vector and the item vector are concatenated, and the concatenated vector is input into the MLP layer to generate the item prediction vector at that moment.

[0083] (4) Designing a path reasoning system

[0084] The item prediction vector generated by the time-aware system is concatenated with the search history vector, and the current state is input into the bidirectional LSTM to encode the state vector s according to formula (12): t , according to the current state s t And the strategy formula (13) selects an action a t , use the scoring function to calculate the score for each action, sort all actions in the action space according to the score, and keep the top ranked actions in the pruned action space In. Execute actions, the agent performs actions a in the environment t , the environment returns the next state s according to formula (10) t+1 , and calculate the reward for executing the current action according to formula (11). Input the obtained path into the LSTM network and obtain the path vector according to formula (16). Use the REINFORCE algorithm to calculate the policy gradient according to formula (15) and update the policy network parameters.

[0085] (5) Model training

[0086] First, the item prediction vector output by the time perception system is passed to the path reasoning system for associated path reasoning, and then the multiple path vectors output by the path reasoning system, the item vector of the current time step and the user vector are input into the time perception system to obtain the item prediction vector. The two systems work iteratively and complement each other to obtain the final item prediction vector, and the obtained vector is input into formula (17) to obtain the recommendation probability. Secondly, according to the loss function of formula (18), the loss between the probability calculated in formula (17) and the correct label is reduced, so that the calculated probability is constantly close to the correct label, so that the user and item prediction vectors are more accurate, and then the training process is continuously cycled to continuously reduce the loss so that the model converges, and finally recommendations are made based on the converged user vector and item prediction vector.

[0087] The contents described in the embodiments of this specification are merely enumerations of implementation forms of the inventive concept and are for illustrative purposes only. The protection scope of the present invention should not be considered to be limited to the specific forms described in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that can be thought of by ordinary technicians in this field based on the inventive concept.

Claims

1. An explainable sequence recommendation method integrating time perception and path reasoning, characterized in that: The method comprises the following steps: (1) Building a time-series collaborative knowledge graph According to the historical interaction information of users, the collaborative knowledge graph is sliced ​​to obtain a series of ordered snapshots and a temporal collaborative knowledge graph is established. The temporal collaborative knowledge graph TCKG is expressed as in represents a snapshot of CKG at time τ, where h represents the head entity, t represents the tail entity, and r τ represents the relationship at time τ, ε represents the entity set, Represents a set of relations; (2) Learning the representation of temporal collaborative knowledge graphs The TransE translation model is used to embed entities and relations in TCKG. The triples (h,r τ ,t), where h,r τ ,t have d-dimensional embedding, that is, The scoring function is defined as e h +r τ To e t The distance between them is given by the following formula: Smaller score indicating that the triple is more likely to be correct than vice versa; The loss function for the relative order between positive and negative triplets is defined as: in And (h, r, t′) is the negative triplet after the tail entity is replaced by the random entity t′, σ(·) is the sigmoid function; Through training, we get the user vector set {h u }、Item vector set {h v } and the set of relation vectors {h r }; (3) Designing a time-aware system The time-aware system takes the current item vector, user vector, and multiple path vectors as input to generate an item prediction vector; (4) Designing a path reasoning system The path reasoning system performs associated path reasoning under the guidance of the item prediction vector generated by the time perception system, and then obtains the above-mentioned multiple path vectors through path representation learning. The system uses a reinforcement learning framework to establish a reasoning environment, which provides the policy decision module with information about the current environment state and executable operations in TCKG, and rewards the decision module according to the degree of match between the current policy and user interaction; (5) Model training The two systems work iteratively and complement each other. The item prediction vector output by the time perception system is passed to the path reasoning system for associated path reasoning. Then, multiple path vectors output by the path reasoning system and the item vector and user vector of the current time step are input into the time perception system to obtain the item prediction vector. Get a trained model to provide users with accurate recommendation results.

2. The explainable sequence recommendation method integrating time perception and path reasoning as claimed in claim 1, characterized in that: The process of (3) is as follows: First, multiple path vectors are passed to the time perception system. These paths reflect different purchasing preferences of users. The path attention mechanism is used to aggregate multiple path vectors. The resulting multi-path aggregation vector can better represent the complex relationship between entities. The calculation formula of the path attention mechanism is: MultiHead(Q φ ,K φ ,V φ )=Concat(head1,...,head1)W O (3) Among them head i yes f(·) is the softmax function, Q, K, V are the self-attention variables associated with path φ, W is the weight matrix, d k is the dimension, Concat(·) is the concatenation operation; After obtaining the multi-path aggregation vector, the item attention mechanism is used to combine the current item's own information with the path aggregation vector to integrate the historical time series information and update the item vector at the current moment. The calculation formula of the item attention mechanism is: where p i represents the initial representation of the current item, φ i-1→i is the path instance from the i-1th item to the i-th item, W is the weight matrix, b is the bias, and g(·) is the ReLU function; The user vector and the item vector are concatenated and input into the multi-layer perceptron to generate the item prediction vector at that moment. The calculation formula is: Since there is no historical time series information before the first item, the present invention concatenates the user vector and the first item vector, inputs the concatenated vector into a multi-layer perceptron, and generates the item prediction vector at that moment. The calculation formula is: h1=MLP(Concat(h u ,h1)) (7) The generated item prediction vector is input into the path reasoning system for associated path reasoning.

3. The explainable sequence recommendation method integrating time perception and path reasoning as claimed in claim 1 or 2, characterized in that: The process of (4) is as follows: (4.1) State: The initial state is s0 = [u,u,φ,φ], and the state s at time step t is t =[u,h t ,e t ,e' t+1 ], where u represents the user entity, h t represents a history of t steps (e0,e1,...,e t ), e t Represents the entity of the current step, e' t+1 represents the entity estimation vector generated by the time-aware system at time step t+1; (4.2) Action: At time step t, the action is expressed as where e t+1 Indicates the next entity to be visited, r t+1 Indicates the current entity and e t+1 the relationship between; (4.3) Pruning strategy: In order to improve model efficiency and control space complexity, it is necessary to perform effective reasoning while maintaining the maximum out-degree benchmark of the action space. The score function is used to calculate the score for each action, which is defined as follows: Where <·,·> represents the dot product operation, r t′ ∈{r kg ∪r interact } is the relation embedding of t′ jump time, sort all actions in the action space according to the score, and keep the top ranked actions in the pruned action space In, it is defined as: Where ∈ is the predefined action space size, which is used to limit the number of actions after pruning; (4.4) State transition: given state s t and action a t As well as the item representation delivered by the time-aware system, the state transition is represented as: s k+1 =δ(s t ,a t )=[u,h t+1 ,e t+1 ,e’ t+2 ] (10) (4.5) Reward function: The goal of reinforcement learning is to find user-item path instances and item-item path instances. In order to evaluate the quality of different paths and the flexibility of selection, soft rewards are used to achieve the above goals. The reward function is defined as: Where v represents the item entity, V represents the item entity set, and T represents the maximum number of steps for path search; if the target entity is found before T steps, the reward is calculated, otherwise the reward is 0; (4.6) Optimization strategy: Parameterize the path reasoning strategy using state information and action space. The state s at time step t t =[u,h t ,e t ,e' t+1 ], including search history h t , using bidirectional LSTM to encode the state vector s t , the formula is as follows: Where W1 is a linear parameter, and the policy network π outputs in the pruning space The probability of each action in: Where W a is a linear parameter that needs to be learned. In the policy gradient method, we hope to optimize the strategy by maximizing the cumulative reward. In order to improve the training efficiency, a baseline is introduced to reduce the variance. The value network M(s t ) The state vector x t Mapping to real values ​​in reinforcement learning as a baseline: M(s t )=x t W c (14) Where W c are the linear parameters that need to be learned. The two networks are trained by maximizing the expected reward of any user u, optimized by the REINFORCE algorithm, which updates the model parameters Θ using the following policy gradient: Where R T Indicates that from state s t Discounted cumulative rewards from start to end state; (4.7) Path representation learning: Obtain candidate user-item and item-item path instances and represent the path data as where p i1 represents the feature vector of path i at time step t, T i is the length of path i. The path data is input into the LSTM layer. LSTM maps the input sequence to an embedding vector of a fixed size. The calculation is shown as follows: Where W LSTM is the weight matrix, h t-1 is the hidden state of the previous time step, the last hidden state output by the LSTM layer It is represented as a path vector, which is transmitted to the time-aware system as guidance information.

4. The explainable sequence recommendation method integrating time perception and path reasoning as claimed in claim 1 or 2, characterized in that: In (5), the item prediction vector and user vector finally obtained after multiple iterations are input into a multi-layer perceptron to calculate the user-item prediction score: r u,v =MLP(Concat(h u ,h v )) (17) The MLP contains two hidden layers, uses ReLU as the activation function, and the Sigmoid function as the output layer; Use the implicit feedback loss of negative sampling as the entire loss function: in Represents the calculation of the expected value on the negative sample set, where P neg represents the distribution of negative samples, and j represents the samples drawn from this distribution.

Images