Decoupling evolution modeling and service recommendation method oriented to user interest state distribution

Abstract

The invention relates to a user interest state distribution-oriented decoupling evolution modeling and service recommendation method, which comprises the following steps of: firstly, collecting interaction data of a user in a service system, and constructing a user-article-time triple sequence; performing initial static embedding learning on the user and the article based on a graph convolution algorithm, and extracting a cooperative signal in the user-article interaction graph; constructing a decoupling state space model, representing a user interest state as a distribution variable composed of an orthogonal rotation matrix and a diagonal scaling matrix, and respectively capturing interest direction evolution and diversity change characteristics; generating a user interest precision matrix according to rotation-stretching transformation, and measuring the matching degree between a user and an article by adopting an anisotropic mahalanobis distance; and constructing a recommendation list according to the distance scores, and taking the first K optimal articles as recommendation results. According to the method, static embedding limitation of user interest modeling is broken through, invariant keeping and potential preference extrapolation in a dynamic evolution structure are achieved, and the expression ability and generalization performance of a recommendation system are improved.

Description

Technical Field

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

[0002] The dynamic evolution modeling of user interests is the key to improving the responsiveness and intelligence level of personalized recommendation systems. Its core lies in capturing the fine-grained changes of user preferences over time to support more accurate and interpretable behavior prediction. In practical applications, the recommendation system needs to simultaneously perceive the intensity and directionality of interests and depict their stable evolution paths during the interaction process. However, existing methods have limitations in two key aspects: recommendation retrieval and interest evolution modeling. On the one hand, the user interest state is usually modeled as a static high-dimensional embedded point, which is difficult to express its time-varying and distribution characteristics, resulting in the lack of description of the dynamic evolution and uncertainty of interests in the recommendation process. On the other hand, the update of the interest state mostly uses additive or aggregative transformations of sequences, which are limited within the convex hull formed by historical interaction contents, making it difficult to express potential preference extrapolation, and also difficult to retain the structural invariants or capture the diversity changes during the interest evolution process. In view of the above problems, the present invention proposes a paradigm optimization strategy that combines telescopic rotation evolution and interest distribution modeling. On the one hand, the user interest state is extended from a static embedded point to a high-dimensional random variable to reflect its dynamic evolution attributes and potential distribution structure. On the other hand, the interest evolution process is decoupled into orthogonal rotation transformation and telescopic transformation. The former is used as an isometric transformation to maintain the invariant structure during the interest direction migration process, and the latter depicts the dynamic changes in the diversity of user interests. Although interest distribution modeling is introduced to reflect the uncertainty of evolution, the recommendation retrieval itself is still carried out within a deterministic framework, relying on the joint modeling of the orthogonal rotation group and the group of positive definite diagonal matrices, and combining the anisotropic Mahalanobis distance to achieve accurate measurement based on structural alignment, thereby breaking through the historical convex hull limit and improving the modeling ability and generalization effect of the recommendation system for complex interest dynamics. Summary of the Invention

[0003] In order to overcome the deficiencies of the prior art, the present invention provides a decoupled evolution modeling and service recommendation method for user interest state distribution, aiming to break through the static embedding limit of user interest modeling, achieve the preservation of invariants and potential preference extrapolation in the dynamic evolution structure, and improve the expression ability and generalization performance of the recommendation system.

[0004] The decoupled evolution 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 that the interest distribution is , and the true interest distribution is obtained by statistically analyzing historical interaction data.

[0016] The training objective of the model at 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 ability, a pointwise feedforward network can be used to stack two linear state space models respectively to increase the network depth.

[0019] For a user 's interaction sequence , two state space modules are constructed, and the outputs at the last sequence position of each module are taken as the prediction results, denoted as , ,

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

[0021] According to the user interest drift generator vector , an anti-symmetric matrix is constructed element by element, and then the corresponding orthogonal rotation matrix is obtained through the exponential mapping.

[0022] Here, it is assumed that the Lsu modeled by the model is the anti-symmetric matrix of the implicit evolution time step. This is because it is difficult to directly estimate the time of the prediction target, and the anti-symmetric matrix, as the generator of the orthogonal rotation matrix, requires a scalar to parameterize the rotation angle, and the anti-symmetric matrix itself only indicates a rotation axis and direction.

[0023] Among them, the specific method of constructing the anti-symmetric matrix element by element is to use to fill the lower triangular part of the matrix with diagonal elements being 0, and symmetrically fill the opposite numbers in the upper triangular part.

[0024] The obtained user interest state precision matrix is , which is used as the bilinear form for measuring the distance between the user interest state and the static representation of the item in the follow-up.

[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] , , where is the hidden state of the scale vector at the -th step, is the embedding vector of the -th item in the interaction sequence and can be appended with the sine encoding of the corresponding timestamp, is the scale vector of the user interest state modeled at the -th step, is the learnable parameter matrix, is the extended dimension of the hidden state. To ensure that the resulting diagonal matrix is positive definite, a Softplus activation function is appended to the output layer of this network .

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

[0029] , , where is the rotation hidden state at the -th step, is the embedding vector of the -th item in the interaction sequence and can be appended with the sine encoding of the corresponding timestamp, is the user interest drift generator vector modeled at the -th step, is the learnable parameter matrix, is the extended dimension of the hidden state, which is used to construct an anti-symmetric matrix, i.e., the Lie algebra of the orthogonal rotation group.

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

[0031] For a certain user 's sequence its recommendation list is defined as , where the Mahalanobis distance , i.e., the anisotropic bilinear form corresponding to the above user interest state precision matrix .

[0032] The network optimization objective is to minimize the cross-entropy loss , where is the set of all prefixes of the interaction sequence of user , is 's corresponding positive sample item label.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[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 interaction data of users in the service system is collected, including the set of user IDs , the set of item IDs and the check-in triples containing timestamps . First, according to the set of check-in triples , the user-item interaction bipartite graph is filtered out, where is the set of equal-weight edges.

[0041] Secondly, according to the set of check-in triples, the prefix set of interaction sequences is sampled for each user by time , where the set of all interaction item pairs is , and the remaining pairs are the prefixes of , is the total number of items interacted by user . After the dataset is divided for each user's sequence according to a ratio, sampling is performed according to the above method.

[0042] Step 2: The initial embedding vectors of items and users are pre-trained through the graph convolution algorithm, and the lightweight graph convolution clustering method is used 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-like 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, capturing the evolution of interest direction and the characteristics of diversity change respectively; the training objective at this stage is to minimize the divergence between the predicted distribution and the true distribution.

[0043] Specifically, it includes the following steps:

[0044] S2.1: Embedding layer

[0045] The initial embedding vectors are assigned to users and items by sampling the standard multivariate Gaussian distribution , .

[0046] S2.2: Graph convolution update layer

[0047] For user and the items with which it has interacted, the graph convolution update formulas for the corresponding embedding vectors are defined as follows:

[0048] , , where are the user and item embedding vectors of the -th layer respectively, and are the set of neighbor nodes of this node and the degree of this node respectively.

[0049] S2.3: Average pooling layer

[0050] The final embeddings of the user and the item are represented by average pooling as:

[0051] .

[0052] S2.4: Network Optimization Layer

[0053] The model predicts the interest distribution as , and the true interest distribution , where is the number of interactions between user and item .

[0054] The training objective of the model at 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 user interest state rotation and evolution state space model. For the historical interaction item embedding sequence, use two independent content-aware state space models to separately model the diagonal matrix describing the user interest diversity and the rotation matrix describing the user interest rotation and evolution in this sequence state;

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

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

[0058] [[ID=3C]], ,

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

[0060] According to the solution of this ordinary differential equation and discretize the above parametric equations with a time step , the following discretized equations are obtained

[0061] , ,

[0062] where , and the trainable parameters of the model are .

[0063] The parameters required for modeling using the content-aware state space model include the following five steps:

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

[0065] For the input where the superscripts represent the batch size and the maximum sequence length used for training, respectively. First, the dimension is expanded using a linear projection to obtain:

[0066] , and the two are equal, where is reserved as the input for the subsequent residual connection. That is, for , first use a one-dimensional convolution to capture the relationship between neighboring item embedding vectors, and then use the SiLU activation function to capture the non-linear relationship to obtain as the basis for the subsequent input content-aware 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 constructs independent state space parameters for the feature dimensions of the input, and only stores diagonal elements to represent the matrix on ; for the parameter , it comes from two different linear projections respectively, and the information sources are the above .

[0069] S3.2: Discretize the parameter matrix

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

[0071] According to the original parameter matrix obtained above and the content-aware discretization step parameter to construct the discretized parameter:

[0072] , ,

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

[0074] Take as the input and apply the transfer equation:

[0075] ,

[0076] to obtain the hidden state tensor of the selected state space model output , and use the above reserved to perform a residual connection to obtain:

[0077] , finally obtain the network output through 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 means of this embodiment.

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

[0080] Among them, the first state - space model, that is, the user interest diversity evolution module, is defined as:

[0081] , , where is the hidden state of the scale vector at the th step, is the embedding vector of the th item in the interaction sequence and can be appended with the sine encoding of the corresponding timestamp, is the scale vector of the user interest state modeled at the th step, is the learnable parameter matrix, is the hidden state expansion dimension. To ensure that the obtained diagonal matrix is positive - definite, a Softplus activation function is added to the output layer of this network

[0082] Among them, the second state - space model, that is, the user interest rotation evolution module, is defined as:

[0083] , , where is the hidden state of rotation at the th step, is the embedding vector of the th item in the interaction sequence and can be appended with the sine encoding of the corresponding timestamp, is the drift generator vector of the user interest modeled at the th step, is the learnable parameter matrix, is the hidden state expansion dimension. is used to construct an anti - symmetric matrix, that is, the Lie algebra of the orthogonal rotation group.

[0084] To enhance the network modeling ability, a point - wise feed - forward network can be used to stack the two linear state - space models respectively to increase the network depth.

[0085] For a user 's interaction sequence , take the output of the last sequence position of the two modules as the prediction result, denoted as and .

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

[0087] According to the user interest drift generator vector construct an anti-symmetric matrix element by element , and then obtain the corresponding orthogonal rotation matrix through the exponential map . Here, it is assumed that the Lsu modeled by the model is the anti-symmetric matrix of the implicit evolution time step. This is because it is difficult to directly estimate the time of the prediction target, and the anti-symmetric matrix, as the generator of the orthogonal rotation matrix, requires a scalar to parameterize the rotation angle, and the anti-symmetric matrix itself only indicates a rotation axis and direction. Among them, constructing the anti-symmetric matrix element by element specifically uses to fill the lower triangular part of the matrix with diagonal elements being 0, and symmetrically fill the opposite numbers in the upper triangular part.

[0088] The obtained user interest status precision matrix is , which is used as the bilinear form for measuring the distance between the user interest status and the static representation of the item in the following.

[0089] Step 4: Construct an anisotropic recommendation retriever. Construct the user interest status precision matrix through the diagonal matrix and the rotation matrix obtained in Step 3, use the Mahalanobis distance as the recommendation metric, and take the top K items with the closest distances as the recommendation list.

[0090] For a certain user 's sequence its recommendation list is defined as , where the Mahalanobis distance , that is, the anisotropic bilinear form corresponding to the above user interest status precision matrix .

[0091] The network optimization objective is to minimize the cross-entropy loss , where is the set of all interaction sequence prefixes of user , is the corresponding positive sample 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. Decoupled evolution modeling and service recommendation method for user interest state distribution, characterized in that: It includes the following steps: Step 1: Collect the interaction data of users in the service system, construct an interaction sequence of triples of user ID - item ID - interaction time, and partition the data set; Step 2: Pre-train the initial embedding vectors of items and users through the graph convolution algorithm, and adopt a 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-like distribution; Step 3: Construct a user interest state rotation and evolution state space model. For the historical interaction item embedding sequence, use two independent content-aware state space models to respectively model the diagonal matrix describing the user interest diversity and the rotation matrix describing the user interest rotation and evolution in this sequence state; Step 4: Construct an anisotropic recommendation retriever. Construct a user interest state precision matrix through the diagonal matrix and rotation matrix obtained in Step 3. Take the Mahalanobis distance as the recommendation metric, and select the top K items with the closest distances as the recommendation list.

2. The decoupled evolution modeling and service recommendation method for user interest state distribution according to claim 1, wherein: The graph convolution clustering algorithm is defined as follows: For a given set of users , a set of items and their historical interaction pairs to construct an equal-weighted undirected bipartite graph , and assign initial embedding vectors to users and items by sampling from a standard multivariate Gaussian distribution ; For the user and the items they have interacted with , the corresponding embedded vector graph convolution update formulas are defined as follows: ， , where are the -th layer user and item embedding vectors, are the set of neighbor nodes of this node and the degree of this node, respectively; The user and item final embeddings are represented by average pooling as follows: , The model predicts that the interest distribution is , and the true interest distribution is obtained by statistically analyzing historical interaction data. The training objective of this stage model is to fit the real distribution, and the loss function is defined as the KL divergence between the real distribution and the predicted distribution: .

3. The decoupled evolution modeling and service recommendation method for user interest state distribution according to claim 1, characterized in that: The user interest state rotation and evolution state space model is defined as follows: For a user of an interaction sequence two state space modules are constructed and the output of the last sequence position of each module is taken as the prediction result, denoted as and , According to the user interest state scale vector Construct a positive definite diagonal matrix , Generate a meta-vector based on user interest drift Construct an anti-symmetric matrix element by element , and then obtain the corresponding orthogonal rotation matrix through exponential mapping , where the construction of the anti-symmetric matrix element by element specifically uses to fill the lower triangular part of the upper matrix with diagonal elements being 0 , and symmetrically fill the opposite numbers in the upper triangular part The obtained user interest state precision matrix is , which serves as a bilinear form for subsequently measuring the distance between the user interest state and the static representation of the item.

4. The decoupled evolution modeling and service recommendation method for user interest state distribution according to claim 3, characterized in that: The user interest state rotation and evolution state space model includes a user interest diversity evolution module and a user interest rotation and evolution module.

5. The decoupled evolution modeling and service recommendation method for user interest state distribution according to claim 4, characterized in that: The user interest diversity evolution module is defined as: , , where is the hidden state of the -th step scale vector, is the embedding vector of the -th item in the interaction sequence and can be appended with the sine encoding of the corresponding timestamp, is the scale vector of the -th step user interest state modeled, is a learnable parameter matrix, is the extended dimension of the hidden state. To ensure that the resulting diagonal matrix is positive definite, a Softplus activation function is appended to the output layer of this network .

6. The decoupled evolutionary modeling and service recommendation method for user interest state distribution according to claim 4, characterized in that: The user interest rotation and evolution module is defined as: , , where is the step rotation hidden state, is the embedding vector of the th item in the interaction sequence and can be appended with the sine encoding of the corresponding timestamp, is the step user interest drift generator vector modeled, is the learnable parameter matrix, is the hidden state expansion dimension, used to construct an anti-symmetric matrix, i.e., the Lie algebra of the orthogonal rotation group.

7. The decoupled evolutionary modeling and service recommendation method for user interest state distribution according to claim 1, characterized in that: The anisotropic recommendation retriever specifically includes: For a certain user sequence its recommended list is defined as , where the Mahalanobis distance , that is, the anisotropic bilinear form corresponding to the above user interest state precision matrix ​ The network optimization objective is to minimize the cross-entropy loss , where is the set of all interaction sequence prefixes of users , and is the corresponding positive sample item label 8. The decoupled evolution modeling and service recommendation method for user interest state distribution according to claim 1, characterized in that: Step 1 specifically includes: Collect the interaction data of users in the service system, including the set of user IDs , the set of item IDs and the check-in triples containing timestamps . First, filter out the user-item interaction bipartite graph from the set of check-in triples , where is the set of equal-weight edges. Sample the prefix set of interaction sequences for each user according to the set of check-in triples , where the set of all interaction item pairs is , and the remaining pairs are the prefixes of . is the total number of items interacted by user .

Images