Decoupled evolutionary modeling and service recommendation method for user interest state distribution

Abstract

This invention relates to a decoupled evolutionary modeling and service recommendation method for user interest state distribution. The method first collects user interaction data within a service system to construct a sequence of user-item-time triples. A graph convolution algorithm is used to perform initial static embedding learning of users and items, extracting collaborative signals from the user-item interaction graph. A decoupled state space model is constructed, representing user interest states as distributed variables composed of an orthogonal rotation matrix and a diagonal scaling matrix, respectively capturing the evolution of interest direction and the changing diversity of interest. A user interest precision matrix is ​​generated based on the rotation-scaling transformation, and the anisotropic Mahalanobis distance is used to measure the degree of match between users and items. A recommendation list is constructed based on the distance scores, with the top K best items selected as the recommended results. This method overcomes the static embedding limitations of user interest modeling, achieving invariant preservation and potential preference extrapolation in a dynamically evolving structure, thereby improving the expressiveness and generalization performance of the recommendation system.

Description

Technical Field

[0001] The present invention relates to a decoupled evolutionary modeling and service recommendation method for user interest state distribution, which belongs to a modeling optimization technology in personalized recommendation and intelligent service systems, and belongs to the field of data mining and recommendation technology. Background Art

[0002] Modeling the dynamic evolution of user interests is the key to improving the responsiveness and intelligence of personalized recommendation systems. Its core lies in capturing the fine-grained changes in user preferences over time to support more accurate and explainable behavior predictions. In practical applications, recommendation systems need to simultaneously perceive the intensity and directionality of interests and characterize their stable evolution path during the interaction process. However, existing methods have limitations in two key aspects: recommendation retrieval and interest evolution modeling. On the one hand, user interest states are usually modeled as static high-dimensional embedded points, which makes it difficult to express their time-varying and distributional characteristics, resulting in a lack of characterization of the dynamic evolution and uncertainty of interests in the recommendation process. On the other hand, the update of interest states mostly adopts additive or aggregate transformations of sequences, which is limited to the convex hull of historical interaction content. It is difficult to express potential preference extrapolation, retain structural invariants in the interest evolution process, or capture diversity changes. In view of the above problems, the present invention proposes a paradigm optimization strategy that integrates telescoping rotation evolution and interest distribution modeling. On the one hand, the user's interest state is expanded from a static embedding point to a high-dimensional random variable to reflect its dynamic evolution properties and potential distribution structure. On the other hand, the interest evolution process is decoupled into orthogonal rotation transformations and scaling transformations. The former, as an isometric transformation, is used to maintain the invariant structure during the migration of interest directions, while the latter characterizes the dynamic changes in user interest diversity. Although interest distribution modeling is introduced to reflect the uncertainty of evolution, recommendation retrieval itself is still carried out within a deterministic framework. Relying on the joint modeling of orthogonal rotation groups and diagonally positive matrix groups, combined with the anisotropic Mahalanobis distance, it achieves precise measurement based on structural alignment, thus breaking through the limitations of the historical convex hull and improving the recommendation system's modeling capabilities and generalization effect for complex interest dynamics. Summary of the Invention

[0003] In order to overcome the shortcomings of the existing technology, the present invention provides a decoupled evolutionary modeling and service recommendation method for user interest state distribution, aiming to break through the static embedding limitations of user interest modeling, realize the invariant preservation and potential preference extrapolation in the dynamic evolution structure, and improve the expressiveness and generalization performance of the recommendation system.

[0004] The decoupled evolutionary modeling and service recommendation method for user interest state distribution includes the following steps:

[0005] Step 1: Collect user interaction data in the service system, construct an interaction sequence consisting of user ID, item ID, and interaction time triples, and partition the dataset. This data can come from real-world scenarios with user behavior logs, such as music recommendations and geo-social networks.

[0006] Step 2: Pre-train the initial embedding vectors of items and users through the graph convolution algorithm, and use the lightweight graph convolution clustering method to extract static collaborative signals in the user-item interaction graph. The obtained embedding vectors are distributed in the Euclidean space and form a community distribution. A decoupled state space model is constructed, and the user interest state is represented as a distribution variable composed of an orthogonal rotation matrix and a diagonal scaling matrix, which respectively capture the evolution of interest direction and diversity change characteristics; the training goal of this stage is to minimize the divergence between the predicted distribution and the true distribution.

[0007] Step 3: Construct a state-space model of the rotational evolution of user interest states. For the historical interaction item embedding sequence, two independent content-aware state-space models are used to model the diagonal matrix describing the diversity of user interests in the sequence state and the rotation matrix describing the rotational evolution of user interests.

[0008] Step 4: Build an anisotropic recommendation retriever. Use the diagonal matrix obtained in step 3 and the rotation matrix to construct the user interest state precision matrix. Use the Mahalanobis distance as the recommendation indicator and take the top K items with the closest distance as the recommendation list.

[0009] Preferably, the graph convolution clustering algorithm is defined as follows:

[0010] For a given set of users , item collection and its historical interactions Constructing an equal-weighted undirected bipartite graph , assign initial embedding vectors to users and items by sampling from standard multivariate Gaussian distribution ;

[0011] For users and items they interacted with , the corresponding embedding vector graph convolution update formulas are defined as follows:

[0012] ,

[0013] ,in Respectively Layer user and item embedding vectors, are the neighbor node set of the node and the degree of the node respectively;

[0014] The final embedding of users and items is expressed by average pooling as follows: ,

[0015] The model predicts the interest distribution as , real interest distribution Obtained through historical interaction data statistics,

[0016] The goal of model training in this stage is to fit the true distribution, and the loss function is defined as the KL divergence between the true distribution and the predicted distribution: .

[0017] Preferably, the user interest state rotation evolution state space model is defined as follows:

[0018] To enhance the network modeling capability, the two linear state space models can be stacked separately using a point-by-point feedforward network to increase the network depth.

[0019] For users An interactive sequence , construct two state space modules and take the output of the last sequence position of each module as the prediction result and record it as 、 ,

[0020] According to the user interest state scale vector Construct a positive definite diagonal matrix ,

[0021] Generate meta-vectors based on user interest drift Construct an antisymmetric matrix element by element , and then obtain the corresponding orthogonal rotation matrix through exponential mapping ,

[0022] Here we assume that the model Lsu is an antisymmetric matrix of the implicit evolution time step. This is because it is difficult to directly estimate the time of the predicted target, and the antisymmetric matrix as the generator of the orthogonal rotation matrix requires a scalar to parameterize the rotation angle. The antisymmetric matrix itself only indicates a rotation axis and direction.

[0023] Among them, the antisymmetric matrix is ​​constructed element by element using Fill the diagonal elements with 0 The lower triangular part of the upper matrix is ​​filled with opposite numbers symmetrically in the upper triangular part.

[0024] The obtained user interest state precision matrix is , as a bilinear form for the subsequent measurement of the distance between the user's interest state and the static representation of the item.

[0025] The user interest state rotation evolution state space model includes a user interest diversity evolution module and a user interest rotation evolution module.

[0026] The user interest diversity evolution module is defined as:

[0027] , ,in For the Step scale vector hidden state, For the interaction sequence The embedding vector of each item can be added with a sinusoidal code corresponding to the timestamp. For the model Step user interest state scale vector, is the learnable parameter matrix, To expand the dimension of the hidden state, a Softplus activation function is added to the output layer of the network to ensure that the obtained diagonal matrix is ​​positive definite. .

[0028] The user interest rotation evolution module is defined as:

[0029] , ,in For the Step rotation hidden state, For the interaction sequence The embedding vector of each item can be added with a sinusoidal code corresponding to the timestamp. For the model Step user interest drift generates meta vector, is the learnable parameter matrix, Expand the dimension of the hidden state, It is used to construct antisymmetric matrices, that is, the Lie algebra of the orthogonal rotation group.

[0030] The anisotropic recommendation retriever is defined as follows:

[0031] For a certain user sequence Its recommendation list is defined as , where the Mahalanobis distance , that is, the user interest state precision matrix The corresponding anisotropic bilinear form,

[0032] The network optimization goal is to minimize the cross entropy loss ,in For users The set of all interaction sequence prefixes of , for The corresponding positive item label.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention proposes a decoupled evolutionary modeling and service recommendation method for user interest state distribution. By expanding user interest states from static high-dimensional embedding points to high-dimensional random variables, the modeling of the dynamics and uncertainty of user interests is enhanced. A decoupled evolutionary mechanism consisting of orthogonal rotation transformation and scaling transformation is introduced, which are used to maintain structural invariants in changes in interest directions and dynamic transformations that characterize diversity characteristics, respectively, to improve the structurality and interpretability of interest evolution modeling. At the same time, the present invention introduces anisotropic Mahalanobis distance as a deterministic measurement method in the retrieval process, effectively breaking through the expression bottleneck of traditional additive aggregation methods limited by the convex hull of historical trajectories, and enhancing the ability to identify unpopular interests and potential preferences. In addition, the constructed interest distribution supports visual interpretation through dimensionality reduction, and has good module compatibility. It can be flexibly integrated into existing recommendation architectures such as graph embedding and sequence modeling, taking into account model performance, explanatory power and engineering deployment efficiency, thereby achieving efficient, interpretable and generalizable modeling of the user interest evolution process. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0036] Figure 1 This is a framework diagram of the decoupled evolutionary modeling and service recommendation method for user interest state distribution in the present invention. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0038] Reference Figure 1 , a decoupled evolutionary modeling and service recommendation method for user interest state distribution, including the following steps:

[0039] Step 1: Collect user interaction data in the service system, construct an interaction sequence consisting of user ID, item ID, and interaction time triples, and partition the dataset. This data can come from real-world scenarios with user behavior logs, such as music recommendations and geo-social networks.

[0040] Specifically, in this embodiment, the user's interaction data in the service system is collected, including the user number set , item number collection and a check-in triple containing a timestamp First, according to the sign-in triple set Filter out the bipartite graph of user-item interactions ,in is a set of equal-weight edges.

[0041] Secondly, based on the check-in triple set, the interaction sequence prefix set is sampled for each user by time. , where all interaction item pairs are , the remaining pairs are prefix, For users The total number of items interacted with. The dataset is divided into sequences for each user proportionally and then sampled using the above method.

[0042] Step 2: Pre-train the initial embedding vectors of items and users through the graph convolution algorithm, and use the lightweight graph convolution clustering method to extract static collaborative signals in the user-item interaction graph. The obtained embedding vectors are distributed in the Euclidean space and form a community distribution. A decoupled state space model is constructed, and the user interest state is represented as a distribution variable composed of an orthogonal rotation matrix and a diagonal scaling matrix, which respectively capture the evolution of interest direction and diversity change characteristics; the training goal of this stage is to minimize the divergence between the predicted distribution and the true distribution.

[0043] Specifically, the following steps are included:

[0044] S2.1: Embedding layer

[0045] Assign initial embedding vectors to users and items by sampling standard multivariate Gaussian distribution , .

[0046] S2.2: Graph Convolution Update Layer

[0047] For users and items they interacted with , the corresponding embedding vector graph convolution update formulas are defined as follows:

[0048] , ,in Respectively Layer user and item embedding vectors, are respectively the set of neighbor nodes of the node and the degree of the node.

[0049] S2.3: Average Pooling Layer

[0050] The final embedding of users and items is expressed by average pooling as follows:

[0051] .

[0052] S2.4: Network Optimization Layer

[0053] The model predicts the interest distribution as ,Real interest distribution ,in For users and items The number of interactions.

[0054] The goal of model training in this stage is to fit the true distribution, and the loss function is defined as the KL divergence between the true distribution and the predicted distribution: .

[0055] Step 3: Construct a state-space model of the rotational evolution of user interest states. For the historical interaction item embedding sequence, two independent content-aware state-space models are used to model the diagonal matrix describing the diversity of user interests in the sequence state and the rotation matrix describing the rotational evolution of user interests.

[0056] In this embodiment, two content-aware state space models are introduced to respectively model the scale transformation parameters and rotation transformation parameters of the user interest state evolution sequence.

[0057] The concept of the state-space model is defined by the following set of ordinary differential equations:

[0058] , ,

[0059] This set of equations describes an observable dynamic system that is linearly driven by an external signal.

[0060] According to the solution of this ordinary differential equation and time step Discretize the above parametric equations and get the following discretized equations

[0061] , ,

[0062] in , the model trainable parameters are .

[0063] Modeling the required parameters using a content-aware state-space model involves the following five steps:

[0064] S3.1: Constructing content-aware variables

[0065] For input The superscript Represent the batch size and maximum sequence length used for training, respectively. First, the linear projection is used to expand the dimension to obtain:

[0066] , the two are equal, where It is used as input for subsequent residual connections. That is to say First, use one-dimensional convolution to capture the embedding vector relationship of neighboring items, and then use SiLU activation function to capture nonlinear relationships, and get Serves as the basis for subsequent input content perception mechanism.

[0067] S3.2: Construct the original parameter matrix

[0068] For the parameters of the state-space model It can be a structured matrix based on the High Order Polynomial Projection Operator (HiPPO), which is the input The feature dimensions construct independent state space parameters and only store diagonal elements to express The matrix on ; for the parameter , which come from two different linear projections, the information source is the above .

[0069] S3.2: Discretization Parameter Matrix

[0070] , is the basic size hyperparameter of the discretization step size.

[0071] According to the original parameter matrix obtained above and the content-aware discretization step size parameter Perform discretization parameter construction:

[0072] , ,

[0073] S3.4: Use the above parameters to get the output and construct the residual connection to get the network output

[0074] Will As input Apply the transfer equation:

[0075] ,

[0076] Get the hidden state tensor output by the selected state space model , using the above spare Perform residual connection to obtain:

[0077] , and finally The network output is obtained by linear projection

[0078] S3.1 to S3.4 are simple descriptions of the required state space model, which can be used as an application implementation method of this embodiment.

[0079] Specifically, this method takes the item embedding sequence as input and uses the two aforementioned networks to model scaling and rotation transformations respectively, which are described as follows:

[0080] The first state space model, namely the user interest diversity evolution module, is defined as:

[0081] , ,in For the Step scale vector hidden state, For the interaction sequence The embedding vector of each item can be added with a sinusoidal code corresponding to the timestamp. For the model Step user interest state scale vector, is the learnable parameter matrix, To expand the dimension of the hidden state, a Softplus activation function is added to the output layer of the network to ensure that the obtained diagonal matrix is ​​positive definite.

[0082] The second state space model, namely the user interest rotation evolution module, is defined as:

[0083] , ,in For the Step rotation hidden state, For the interaction sequence The embedding vector of each item can be added with a sinusoidal code corresponding to the timestamp. For the model Step user interest drift generates meta vector, is the learnable parameter matrix, Expand the dimension of the hidden state. It is used to construct antisymmetric matrices, that is, the Lie algebra of the orthogonal rotation group.

[0084] To enhance the network modeling capability, the two linear state space models can be stacked separately using a point-by-point feedforward network to increase the network depth.

[0085] For users An interactive sequence , take the output of the last sequence position of the two modules as the prediction result and record it as 、 .

[0086] According to the user interest state scale vector Construct a positive definite diagonal matrix .

[0087] Generate meta-vectors based on user interest drift Construct an antisymmetric matrix element by element , and then obtain the corresponding orthogonal rotation matrix through exponential mapping , here we assume that the model modeling Lsu is an antisymmetric matrix of the implicit evolution time step, because it is difficult to directly estimate the time of the predicted target, and the antisymmetric matrix as the generator of the orthogonal rotation matrix requires a scalar to parameterize the rotation angle, and the antisymmetric matrix itself only indicates a rotation axis and direction. Among them, the antisymmetric matrix is ​​constructed element by element using Fill the diagonal elements with 0 The lower triangular part of the upper matrix is ​​filled with opposite numbers symmetrically in the upper triangular part.

[0088] The obtained user interest state precision matrix is , as a bilinear form for the subsequent measurement of the distance between the user's interest state and the static representation of the item.

[0089] Step 4: Build an anisotropic recommendation retriever. Use the diagonal matrix obtained in step 3 and the rotation matrix to construct the user interest state precision matrix. Use the Mahalanobis distance as the recommendation indicator and take the top K items with the closest distance as the recommendation list.

[0090] For a certain user sequence Its recommendation list is defined as , where the Mahalanobis distance , that is, the user interest state precision matrix The corresponding anisotropic bilinear form.

[0091] The network optimization goal is to minimize the cross entropy loss ,in For users The set of all interaction sequence prefixes of , for The corresponding positive item label.

[0092] In summary, the present invention provides a paradigm optimization method for a recommendation system that integrates a rotational evolution mechanism and interest distribution modeling. By promoting the user interest state from a static embedding point to a high-dimensional random variable, the present invention effectively introduces the dynamic and uncertain expression of user interests. In terms of structural modeling, the present invention decouples the interest evolution process into an equidistant rotation transformation and a scale-adjustable scaling transformation, respectively characterizing the interest migration direction and diversity expansion, thereby enhancing the modeling ability of the model for complex interest changes. At the same time, combined with a joint modeling method based on orthogonal rotation groups and diagonal positive definite matrix groups, and an anisotropic Mahalanobis distance retrieval mechanism, the present invention breaks through the dependence of traditional sequence models on the historical interaction convex hull, and improves the recommendation system's ability to identify unpopular interests and potential preferences. This method has good interpretability, scalability, and engineering deployment adaptability, and can be widely used in various practical scenarios such as personalized recommendations and behavior modeling.

[0093] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It is apparent to those skilled in the art that various changes, modifications, substitutions, and variations to these embodiments may be made without departing from the principles and spirit of the present invention, and these changes and modifications still fall within the scope of protection of the present invention.

Claims

1. A decoupled evolutionary modeling and service recommendation method for user interest state distribution, characterized by: The following steps are involved: Step 1: Collect user interaction data in the service system, construct an interaction sequence containing user ID-item ID-interaction time triples, and divide the data set; Step 2: Pre-train the initial embedding vectors of items and users through the graph convolution algorithm. Then, use the lightweight graph convolution clustering method to extract the static collaborative signals in the user-item interaction graph. The obtained embedding vectors are distributed in the Euclidean space and form a community distribution. Step 3: Construct a state-space model of the rotational evolution of user interest states. For the historical interaction item embedding sequence, two independent content-aware state-space models are used to model the diagonal matrix describing the diversity of user interests in the sequence state and the rotation matrix describing the rotational evolution of user interests. Step 4: Build an anisotropic recommendation search engine. Use the diagonal matrix obtained in step 3 and the rotation matrix to construct the user interest state precision matrix. Use Mahalanobis distance as the recommendation metric and select the top K items with the closest distance as the recommendation list. The user interest state rotation evolution state space model is defined as follows: For users An interactive sequence , construct two state space modules and take the output of the last sequence position of each module as the prediction result and record it as 、 , According to the user interest state scale vector Construct a positive definite diagonal matrix , Generate meta-vectors based on user interest drift Construct an antisymmetric matrix element by element , and then obtain the corresponding orthogonal rotation matrix through exponential mapping , where the antisymmetric matrix is ​​constructed element by element using Fill the diagonal elements with 0 The lower triangular part of the upper matrix is ​​filled with opposite numbers symmetrically in the upper triangular part. The obtained user interest state precision matrix is , as a bilinear form for the subsequent measurement of the distance between the user's interest state and the static representation of the item.

2. The decoupled evolutionary modeling and service recommendation method for user interest state distribution according to claim 1 is characterized by: The graph convolution algorithm is defined as follows: For a given set of users , item collection and its historical interactions Constructing an equal-weighted undirected bipartite graph , assign initial embedding vectors to users and items by sampling from standard multivariate Gaussian distribution ; For users and items they interacted with , the corresponding embedding vector graph convolution update formulas are defined as follows: , ,in Respectively Layer user and item embedding vectors, are the node's neighbor node set and the node's degree respectively; The final embedding of users and items is expressed by average pooling as follows: , The model predicts the interest distribution as , real interest distribution Obtained through historical interaction data statistics, The goal of stage model training is to fit the true distribution, and the loss function is defined as the KL divergence between the true distribution and the predicted distribution: .

3. The decoupled evolutionary modeling and service recommendation method for user interest state distribution according to claim 1 is characterized by: The user interest state rotation evolution state space model includes a user interest diversity evolution module and a user interest rotation evolution module.

4. The decoupled evolutionary modeling and service recommendation method for user interest state distribution according to claim 3 is characterized by: The user interest diversity evolution module is defined as: , ,in For the Step scale vector hidden state, For the interaction sequence The embedding vector of each item can be added with a sinusoidal code corresponding to the timestamp. For the model Step user interest state scale vector, is the learnable parameter matrix, To expand the dimension of the hidden state and ensure that the obtained diagonal matrix is ​​positive, a Softplus activation function is added to the network output layer .

5. The decoupled evolutionary modeling and service recommendation method for user interest state distribution according to claim 3 is characterized by: The user interest rotation evolution module is defined as: , ,in For the Step rotation hidden state, For the interaction sequence The embedding vector of each item can be added with a sinusoidal code corresponding to the timestamp. For the model Step user interest drift generates meta vector, is the learnable parameter matrix, Expand the dimension of the hidden state, It is used to construct antisymmetric matrices, that is, the Lie algebra of the orthogonal rotation group.

6. The decoupled evolutionary modeling and service recommendation method for user interest state distribution according to claim 1 is characterized by: The anisotropic recommendation retriever specifically includes: For a certain user sequence Its recommendation list is defined as , where the Mahalanobis distance , that is, the user interest state precision matrix The corresponding anisotropic bilinear form, The network optimization goal is to minimize the cross entropy loss ,in For users The set of all interaction sequence prefixes of , for The corresponding positive item label.

7. The decoupled evolutionary modeling and service recommendation method for user interest state distribution according to claim 1 is characterized by: The step 1 specifically includes: Collect user interaction data in the service system, including user ID collection , item number collection and a check-in triplet containing a timestamp , first according to the sign-in triple set Filter out the bipartite graph of user-item interactions ,in is a set of equal-weight edges. Based on the check-in triple set, the interaction sequence prefix set is sampled for each user by time. , where all interaction item pairs are , the remaining pairs are The prefix of For users The total number of items interacted with.

Images