A multi-source information clustering method, device, equipment and storage medium

Abstract

The application provides a multi-source information clustering method and device, equipment and a storage medium, which does not solidify a relation graph before clustering, sets learnable fusion weights for each background knowledge matrix to obtain a comprehensive relation matrix, modifies the matrix according to the edge confidence between samples according to feature data evaluation, and obtains an optimized graph structure through graph structure optimization; after feature coding and clustering based on the graph, the fusion weights, edge confidence evaluation, feature coding and clustering are jointly trained and the graph structure is iteratively updated through a unified target function containing a clustering target, so that the selection of the relation is corrected with the clustering result, the unreliable part is adaptively down-weighted instead of being fixed, and thus a sample clustering result that is not interfered by a biased part is obtained.

Description

Technical Field

[0001] This invention relates to the field of data clustering and graph neural network technology, and particularly to a multi-source information clustering method, apparatus, device and storage medium. Background Technology

[0002] Clustering is a fundamental task in data analysis, aiming to divide samples into groups based on their similarity. Traditional clustering methods (such as K-means, spectral clustering, and time series clustering based on dynamic time warping) typically calculate similarity based solely on the original characteristics of the samples and group them accordingly.

[0003] In practical scenarios such as enterprise profiling, user behavior analysis, intelligence analysis, and text topic mining, in addition to the original features of the samples themselves, there is often supplementary information describing the relationships between samples (referred to as background knowledge in this paper). Examples include business relationships, social attention relationships, text co-occurrence or citation relationships, user-product interaction relationships, and prior constraints from professional fields. This background knowledge contains sample relationships that are difficult to observe directly from the original features alone, and it has reference value for revealing the true and interpretable grouping structure. To utilize this background knowledge, existing methods such as constrained clustering and semi-supervised clustering typically transform the background knowledge into a deterministic relationship graph (such as an adjacency matrix) before clustering begins. When the background knowledge comes from multiple sources, the multi-source background knowledge is first merged and solidified into a deterministic relationship graph, which is then used as a given prior in the subsequent clustering process.

[0004] However, background knowledge typically comes from multiple independent sources, and the accuracy of the relationships reflected by each source varies. Furthermore, these sources may contain missing, erroneous, or contradictory relationships. More importantly, the role a relationship should play in grouping depends on its fit with the true grouping structure of the samples—a structure that remains unknown during the clustering process. The aforementioned constrained clustering and semi-supervised clustering methods solidify multi-source background knowledge into a defined relationship graph before clustering begins. This essentially predetermines the selection and strength of relationships without knowing the true grouping structure. If any part of the relationship graph deviates from the true grouping structure, that part is fixed and persists throughout the entire clustering process, making it difficult to correct later. This leads to biased grouping results, sometimes even worse than using no background knowledge at all.

[0005] In view of the above, this application is hereby submitted. Summary of the Invention

[0006] This invention discloses a multi-source information clustering method, apparatus, device, and storage medium, aiming to solve the technical problem of how to obtain sample clustering results that are not affected by the biased parts when the reliability of multi-source background knowledge is unknown and its relationship graph cannot be correctly determined before clustering.

[0007] The first embodiment of the present invention provides a multi-source information clustering method, including: Acquire feature data of multiple samples and multiple background knowledge matrices, wherein the background knowledge matrices are used to characterize the correlation between samples; Learnable fusion weights are set for each of the background knowledge matrices, and weighted fusion is performed to obtain a comprehensive relationship matrix; the edge confidence between samples is evaluated based on the feature data; the comprehensive relationship matrix is ​​corrected using the edge confidence to obtain a corrected relationship matrix, and the corrected relationship matrix is ​​then optimized using a graph structure to obtain an optimized graph structure. Based on the optimized graph structure, feature encoding is performed on the feature data to obtain a sample embedding representation; Clustering is performed based on the sample embedding representation to obtain the sample clustering results; Using a unified objective function that includes the clustering objective, end-to-end joint training is performed on the learning of the fusion weights, the evaluation of the edge confidence, the feature encoding, and the clustering; when the iteration termination condition is met, the sample clustering result is output.

[0008] The second embodiment of the present invention provides a multi-source information clustering device, comprising: The data acquisition module is used to acquire feature data of multiple samples and multiple background knowledge matrices, wherein the background knowledge matrices are used to characterize the correlation between samples; An adaptive graph learning module is used to set learnable fusion weights for each of the background knowledge matrices and perform weighted fusion to obtain a comprehensive relation matrix; evaluate the edge confidence between samples based on the feature data; correct the comprehensive relation matrix using the edge confidence to obtain a corrected relation matrix; and optimize the graph structure of the corrected relation matrix to obtain an optimized graph structure. The feature encoding module is used to perform feature encoding on the feature data based on the optimized graph structure to obtain the sample embedding representation; The clustering module is used to perform clustering based on the sample embedding representation to obtain sample clustering results; The joint training module is used to perform end-to-end joint training on the learning of the fusion weights, the evaluation of the edge confidence, the feature encoding, and the clustering using a unified objective function that includes the clustering objective; and outputs the sample clustering results when the iteration termination condition is met.

[0009] The third embodiment of the present invention provides a multi-source information clustering device, including a memory and a processor. The memory stores a computer program, which can be executed by the processor to implement a multi-source information clustering method as described in any of the above embodiments.

[0010] The fourth embodiment of the present invention provides a computer-readable storage medium storing a computer program, which can be executed by a processor of the device in which the computer-readable storage medium is located, to implement a multi-source information clustering method as described in any of the above embodiments.

[0011] This invention provides a multi-source information clustering method, apparatus, device, and storage medium. Instead of fixing the relationship graph before clustering, it first sets learnable fusion weights for each background knowledge matrix to obtain a comprehensive relationship matrix. Then, it evaluates the edge confidence between samples based on feature data and corrects the matrix accordingly. An optimized graph structure is obtained through graph structure optimization. After feature encoding and clustering based on this graph, a unified objective function containing the clustering objective is used to jointly train and iteratively update the graph structure end-to-end for the fusion weights, edge confidence evaluation, feature encoding, and clustering. This allows the strength of relationship selection and the clustering results to mutually correct each other, and unreliable parts are adaptively weighted rather than fixed, thereby obtaining sample clustering results unaffected by biased parts. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating a multi-source information clustering method provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a multi-source information clustering device provided in the second embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] To better understand the technical solution of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0015] This invention discloses a multi-source information clustering method, apparatus, device, and storage medium, aiming to solve the technical problem of how to obtain sample clustering results that are not affected by the biased parts when the reliability of multi-source background knowledge is unknown and its relationship graph cannot be correctly determined before clustering.

[0016] Please see Figure 1The first embodiment of the present invention provides a multi-source information clustering method, which can be executed by a multi-source information clustering device (hereinafter referred to as clustering device or system), and in particular, by one or more processors within the clustering device, to at least implement the following steps: S101, acquire feature data of multiple samples and multiple background knowledge matrices, wherein the background knowledge matrices are used to characterize the correlation between samples; In this embodiment, multidimensional time series feature data of multiple samples and multiple background knowledge matrices are obtained, and both are preprocessed.

[0017] Specifically, the feature data is multidimensional time series feature data. It records the characteristic changes of each sample at consecutive time points, which can be represented as ,in For the sample size, For the length of time, The feature dimensions corresponding to each time point; taking enterprise profile analysis as an example, each sample corresponds to one enterprise. For continuous data collection time points (such as monthly or quarterly). These are the characteristic dimensions reflecting the business status of enterprises at each point in time (such as transaction volume, employment scale, tax indicators, etc.), thereby using this multidimensional time series characteristic data to characterize the dynamic characteristics of each sample as it changes over time.

[0018] The background knowledge matrix is ​​as follows: Background knowledge matrix ,in The number of background knowledge matrices, Indicates the first Background knowledge matrix, usually The square array, its first Line number The elements of the column are used to characterize the sample. With sample The strength or similarity of the association between them in this type of relationship; the association includes, but is not limited to, business association, co-occurrence relationship, text similarity relationship, geographical proximity relationship, historical interaction relationship or expert prior relationship, different Sourced from different information channels, each of them portrays the potential relationships between samples that are difficult to observe directly from the original features alone, thus providing a basis for subsequent multi-source fusion based on relationships from different channels.

[0019] Considering that the aforementioned multidimensional time series feature data and background knowledge matrices come from different sources, have different dimensions, and varying degrees of reliability, further preprocessing is performed to ensure that data from different sources can participate in subsequent graph structure learning and feature encoding at a unified scale. The preprocessing of the multidimensional time series feature data may include one or more of the following: missing value handling, outlier handling, time alignment, standardization, or normalization, to eliminate missing values ​​and jumps, align time axes, and eliminate dimensional differences. The preprocessing of the background knowledge matrices may include one or more of the following: matrix normalization, matrix symmetry, missing relation completion, anomalous relation weakening, or self-connection, to regulate the range of relation strength values, complete sparse relations, and suppress obviously anomalous relation edges. If the input feature data and background knowledge matrices already meet the subsequent modeling requirements, the above preprocessing may not be performed. The preprocessed multidimensional time series feature data and background knowledge matrices are then used as input for evaluating edge confidence and weighted fusion to obtain the comprehensive relation matrix.

[0020] S102, set learnable fusion weights for each of the background knowledge matrices, and perform weighted fusion to obtain a comprehensive relationship matrix; evaluate the edge confidence between samples based on the feature data; use the edge confidence to correct the comprehensive relationship matrix to obtain a corrected relationship matrix, and perform graph structure optimization on the corrected relationship matrix to obtain an optimized graph structure; In this embodiment, based on multidimensional time series feature data and various background knowledge matrices, multi-source background knowledge is integrated on the one hand, and the edge confidence between samples is evaluated on the other hand. Based on this, the two are combined to construct an optimized graph structure.

[0021] First, a weighted fusion of multi-source background knowledge is performed. This results in the acquisition of a background knowledge matrix for each category. Set a learnable fusion weight This allows the model to automatically adjust the importance of different background knowledge sources during training, and according to... Weighted fusion yields a comprehensive relationship matrix. ,in This represents the number of background knowledge matrices. Because... These are parameters that are updated along with the clustering objective during subsequent end-to-end training. Reliable background knowledge sources that are highly relevant to the clustering objective are given greater weights to increase their contribution, while background knowledge sources that are noisy or less relevant to the clustering objective are given smaller weights to reduce their impact. This achieves adaptive selection of multi-source background knowledge at the information source level.

[0022] Secondly, the edge confidence between samples is evaluated based on multidimensional time series feature data. For any sample pair... Represent its corresponding feature in the feature data. , (e.g., for the sample) ,sample The vector representation of the time series features (after flattening or aggregation) is input into a multilayer perceptron and activated by a Sigmoid function to obtain the edge confidence of the relationship between the sample pairs. , ,in It is a multilayer perceptron. Use the Sigmoid activation function. Side confidence. The larger the value, the more reliable the judgment of the relationship between sample pairs based on the characteristics of the samples themselves; The smaller the value, the more likely the relationship is to be weak, redundant, or noisy. Unlike fusion weights, which distinguish reliability at the information source level, edge confidence is a judgment made on the reliability of each relationship based on the original features at the specific sample pair level. The two complement each other.

[0023] Subsequently, the comprehensive relation matrix is ​​corrected using the aforementioned edge confidence, resulting in the corrected relation matrix. Specifically, let the comprehensive relation matrix be denoted as... ,according to Element-wise multiplication, where , Each is a matrix , The Middle Line number Column elements, For the sample With sample The edge confidence between them. The resulting corrected relation matrix. It not only reflects the strength of prior relations given by background knowledge This further reflects the reliability of the relationship at the sample feature level. For relationships with high prior strength but low confidence at the feature level, their... The values ​​in the model are suppressed accordingly, thus preventing unreliable relationships from being directly introduced as reliable relationships into subsequent modeling.

[0024] Finally, the corrected relation matrix By sequentially performing sparsification and normalization processes, the graph structure optimization is completed, resulting in the optimized adjacency matrix corresponding to the optimized graph structure. The sparsity processing is used to retain important relation edges and remove weak, redundant, or noisy edges, and can employ at least one of the following methods: threshold filtering, Top-k retention, or differentiable edge selection. Normalization processing ensures numerical stability in subsequent graph convolution or graph relation modeling processes. The resulting optimized adjacency matrix... That is, it serves as the graph structure input used for feature encoding of the multidimensional time series feature data, thereby passing the adaptively learned relational structure in this step to the subsequent feature encoding and clustering stages.

[0025] S103, Based on the optimized graph structure, feature encoding is performed on the feature data to obtain a sample embedding representation; In this embodiment, based on an optimized graph structure, spatiotemporal feature encoding is performed on the multidimensional time series feature data to obtain a sample embedding representation. The spatiotemporal feature encoding extracts the dynamic change pattern of each sample in the time dimension, and aggregates the information of neighboring samples along the optimized graph structure in the sample relationship dimension. The combination of the two results in a sample representation that includes both temporal dynamics and a reliable relationship structure.

[0026] Specifically, firstly, temporally gated convolution is used to process multidimensional time-series feature data. Encoding is performed to extract the dynamic change patterns of the samples over time. The calculation process can be represented as follows: as well as ,in This refers to the temporal convolution operation performed along the time dimension. For temporal convolution kernels, For bias terms, , These are the two components of the output from the temporal convolution. For element-wise multiplication, It is the Sigmoid activation function. This represents the obtained time characteristics. In this process, It acts as a gating mechanism: the closer its value is to 1, the stronger the component... The more time-related dynamic information is preserved, the more the value is suppressed, allowing the model to selectively retain time-varying patterns that are meaningful for clustering, thus obtaining a temporal feature representation. .

[0027] Subsequently, the above time characteristics are represented With the optimized adjacency matrix Common input graph relation modeling, resulting in sample representations after graph relation modeling. Its calculation method can be expressed as ,in , These are the learnable weights and biases in graph relationship modeling. It is a non-linear activation function. Graph relationship modeling can use a graph convolution structure, derived from the optimized adjacency matrix. This determines how information is propagated and aggregated between samples; due to It is an optimized graph structure obtained after edge confidence correction, sparsification, and normalization. In this process, each sample mainly aggregates information from its neighboring samples with strong relationships and high confidence, while neighboring samples corresponding to weak, redundant, or noisy relationships contribute very little, thus avoiding the introduction of noise into the sample representation by unreliable relationships. Therefore, the temporal feature representation... The time dynamic information carried and The reliable relationship structure depicted in The elements are combined to form the spatiotemporal feature code.

[0028] Finally, the sample representation after modeling the graph relations. Flattening yields the sample embedding representation. ,in Indicates the first The sample embedding representation of each sample. This represents the number of samples. The sample embedding is... It encodes the dynamic features of each sample over time and integrates neighbor information aggregated along a reliable relational structure, which is then used as input for clustering. This allows the spatiotemporal features obtained in this step to be passed on to subsequent clustering and joint training stages.

[0029] S104, perform clustering based on the sample embedding representation to obtain the sample clustering results; In this embodiment, clustering optimization is performed based on the sample embedding representation to obtain the current sample clustering result.

[0030] Specifically, based on the sample embedding representation Clustering multiple samples, for example using K-means clustering: This involves representing the embeddings of each sample. It is assigned to the nearest cluster based on its distance from each cluster center, and the cluster centers are iteratively updated until convergence, thereby... Each sample is divided into several clusters, and the cluster to which each sample belongs constitutes the current sample clustering result. It should be noted that the clustering result obtained here is the clustering result under the current sample embedding representation, which will be continuously optimized as the sample embedding representation is updated in subsequent end-to-end joint training, hence it is called the current sample clustering result.

[0031] Based on this, a clustering loss function is constructed according to the clustering results. This ensures that the embeddings of samples within the same cluster are close to each other, while the embeddings of samples from different clusters are separate. The clustering loss function can be calculated by summing the squared L2 distances between each sample embedding and the cluster centers of its respective cluster, and then averaging the sums over all samples. The formula can be expressed as follows: ,in For the sample size, For the first The sample embedding representation of each sample. For the sample The index of the cluster to which it belongs. For the sample The cluster center of the cluster to which it belongs. It is an L2 norm. The smaller the value of , the closer the embedding representation of each sample within the same cluster is to its cluster center, resulting in a more compact cluster structure and better separability; therefore, minimizing This enables the model to learn sample embeddings that are more conducive to clustering.

[0032] The clustering loss function As the clustering objective, in step S105, it is weighted and combined with the edge confidence constraint, graph sparsity constraint, and fusion weight constraint to form a unified objective function, and the parameters in the aforementioned steps are jointly trained end-to-end accordingly; thus, the sample clustering result not only depends on the sample embedding representation, but also, conversely, through... The guidance sample embedding representation and even the upstream fusion weights, edge confidence and graph structure learning are updated in a direction that is more conducive to clustering, thereby forming a collaborative optimization between steps S102 and S104.

[0033] S105, using a unified objective function that includes the clustering objective, perform end-to-end joint training on the learning of the fusion weights, the evaluation of the edge confidence, the feature encoding, and the clustering; when the iteration termination condition is met, output the sample clustering result.

[0034] In this embodiment, the total loss function for joint training is first constructed. Let the unified objective function be denoted as the total loss function. Due to clustering loss It is obtained by weighting and combining the edge confidence constraint, graph sparsity constraint, and fusion weight constraint. Its calculation formula can be expressed as follows: ,in , , These are weighting parameters used to balance the contributions of each constraint term. Among them, the edge confidence constraint term... To constrain the consistency between edge confidence and sample clustering results: Based on the current sample clustering results, let the corresponding edge confidence be set to... Sample pairs between different clusters The constraint term is constructed using the binary cross-entropy method, and its calculation formula can be expressed as follows: ,in For the sample With sample The edge confidence between them The summation is a logarithmic function that iterates through the sample pairs involved in the constraints. By minimizing This ensures that sample relationships within the same cluster tend to have higher confidence, while sample relationships between different clusters tend to have lower confidence, thus making the edge confidence consistent with the clustering results; graph sparsity constraint term. To control the number of edges or the size of edge weights in the optimized graph structure and avoid an overly dense graph, the L1 norm of the optimized adjacency matrix can be used. Its calculation formula can be expressed as follows: ,in The L1 norm of the matrix is ​​used to encourage the model to retain fewer but more important sample relationship edges; weight constraint terms are incorporated. The fusion weights are used to constrain the fusion of multi-source background knowledge and avoid the model from over-relying on a single source of background knowledge. In the case of normalization, at least one of the following can be used: a weight squared penalty, a negative entropy constraint, or a distance constraint between the fusion weights and the uniform weight distribution.

[0035] Subsequently, based on the total loss function All parameters in steps S102 to S104 are updated, including the fusion weights of multi-source background knowledge. The parameters of the multilayer perceptron used for edge confidence evaluation, the relevant parameters of graph structure optimization, the parameters of temporal convolution kernel and graph relationship modeling weights in spatiotemporal feature encoding, and the parameters of clustering enable graph structure learning, feature encoding and clustering optimization to be optimized collaboratively under the same objective function, rather than solved step by step in isolation, thus avoiding suboptimal solutions caused by step-by-step optimization.

[0036] After completing a parameter update, determine whether the iteration termination condition is met. The iteration termination condition may include reaching a preset maximum number of iterations, the total loss function, etc. The following conditions must be met: the change in clustering weights is less than a preset threshold; the change in clustering results (clustering labels) is less than a preset threshold in several consecutive iterations; and the optimized graph structure and sample embedding representation tend to stabilize. If these conditions are not met, return to step S102, evaluate the optimized graph structure again with the updated fusion weights and edge confidence, and continue to steps S103 to S105; if these conditions are met, end the iteration and output the sample clustering results.

[0037] Through the aforementioned end-to-end joint training, the graph structure learning in step S102, the spatiotemporal feature encoding in step S103, and the clustering optimization in step S104 form a closed-loop iteration, enabling the multi-source background knowledge fusion weights, edge confidence, sample embedding representations, and sample clustering results to be updated collaboratively under a unified optimization objective: on the one hand, the sample clustering results are determined by the optimized graph structure and sample embedding representations; on the other hand, the sample clustering results are further updated through… Conversely, adjust the side confidence and via The update of the fusion weights and graph structure is driven by adaptively reducing the weights of relationships that do not conform to the true grouping structure, rather than fixing them before clustering. When outputting the sample clustering results, corresponding interpretable information is also output, including the optimized graph structure (i.e., the optimized adjacency matrix). ), the side confidence between samples and the fusion weight of multi-source background knowledge This helps users understand which background information contributes more to the current clustering and which is more likely to be noise, thus improving the interpretability of the clustering results.

[0038] Please see Figure 2 The second embodiment of the present invention provides a multi-source information clustering device, comprising: Data acquisition module 201 is used to acquire feature data of multiple samples and multiple background knowledge matrices, wherein the background knowledge matrices are used to characterize the correlation between samples; The adaptive graph learning module 202 is used to set learnable fusion weights for each of the background knowledge matrices and perform weighted fusion to obtain a comprehensive relation matrix; evaluate the edge confidence between samples based on the feature data; correct the comprehensive relation matrix using the edge confidence to obtain a corrected relation matrix; and optimize the graph structure of the corrected relation matrix to obtain an optimized graph structure. Feature encoding module 203 is used to perform feature encoding on the feature data based on the optimized graph structure to obtain a sample embedding representation; Clustering module 204 is used to perform clustering based on the sample embedding representation to obtain sample clustering results; The joint training module 205 is used to perform end-to-end joint training on the learning of the fusion weights, the evaluation of the edge confidence, the feature encoding, and the clustering using a unified objective function that includes the clustering objective; and outputs the sample clustering result when the iteration termination condition is met.

[0039] The third embodiment of the present invention provides a multi-source information clustering device, including a memory and a processor. The memory stores a computer program, which can be executed by the processor to implement a multi-source information clustering method as described in any of the above embodiments.

[0040] The fourth embodiment of the present invention provides a computer-readable storage medium storing a computer program, which can be executed by a processor of the device in which the computer-readable storage medium is located, to implement a multi-source information clustering method as described in any of the above embodiments.

[0041] This invention provides a multi-source information clustering method, apparatus, device, and storage medium. Instead of fixing the relationship graph before clustering, it first sets learnable fusion weights for each background knowledge matrix to obtain a comprehensive relationship matrix. Then, it evaluates the edge confidence between samples based on feature data and corrects the matrix accordingly. An optimized graph structure is obtained through graph structure optimization. After feature encoding and clustering based on this graph, a unified objective function containing the clustering objective is used to jointly train and iteratively update the graph structure end-to-end for the fusion weights, edge confidence evaluation, feature encoding, and clustering. This allows the strength of relationship selection and the clustering results to mutually correct each other, and unreliable parts are adaptively weighted rather than fixed, thereby obtaining sample clustering results unaffected by biased parts.

[0042] Exemplary examples show that the computer program described in the third and fourth embodiments of the present invention can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in implementing a multi-source information clustering device. For example, the apparatus described in the second embodiment of the present invention.

[0043] The processor referred to can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the multi-source information clustering method, connecting the various parts of the method through various interfaces and lines.

[0044] The memory can be used to store the computer program and / or modules. The processor implements various functions of a multi-source information clustering method by running or executing the computer program and / or modules stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, text conversion function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, text message data, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0045] If the implemented module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0046] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0047] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-source information clustering method, characterized in that, include: Acquire feature data of multiple samples and multiple background knowledge matrices, wherein the background knowledge matrices are used to characterize the correlation between samples; Learnable fusion weights are set for each of the background knowledge matrices, and weighted fusion is performed to obtain a comprehensive relationship matrix; the edge confidence between samples is evaluated based on the feature data; The comprehensive relation matrix is ​​corrected using the edge confidence to obtain a corrected relation matrix, and the graph structure of the corrected relation matrix is ​​optimized to obtain an optimized graph structure. Based on the optimized graph structure, feature encoding is performed on the feature data to obtain a sample embedding representation; Clustering is performed based on the sample embedding representation to obtain the sample clustering results; Using a unified objective function that includes the clustering objective, end-to-end joint training is performed on the learning of the fusion weights, the evaluation of the edge confidence, the feature encoding, and the clustering; when the iteration termination condition is met, the sample clustering result is output.

2. The multi-source information clustering method according to claim 1, characterized in that, The feature data is multidimensional time series feature data. , ,in For sample size, For time length, Feature dimensions for each time point; The background knowledge matrix is ​​as follows: Background knowledge matrix ,in The number of background knowledge matrices, Indicates the first A background knowledge matrix is ​​used to characterize a type of association between samples, wherein the association includes at least one of business association, co-occurrence relationship, text similarity relationship, geographical proximity relationship, historical interaction relationship or expert prior relationship.

3. The multi-source information clustering method according to claim 2, characterized in that, The weighted fusion yields a comprehensive relationship matrix, the calculation formula of which is: ,in The comprehensive relationship matrix is ​​as follows. For the first Class background knowledge matrix The corresponding fusion weights, This represents the number of background knowledge matrices.

4. The multi-source information clustering method according to claim 2, characterized in that, The formula for calculating the edge confidence between samples based on the feature data is as follows: , ,in, For the sample With sample The edge confidence between them , These are the multidimensional time series feature data. medium sample ,sample The corresponding feature representation, It is a multilayer perceptron. This is the Sigmoid activation function.

5. The multi-source information clustering method according to claim 4, characterized in that, The edge confidence is used to analyze the comprehensive relationship matrix. After making corrections, the corrected relation matrix is ​​obtained, and its calculation formula is as follows: ； in, The corrected relation matrix is... , Each is a matrix , The Middle Line 1 Column elements, For the sample With sample The edge confidence between them; The graph structure optimization of the modified relation matrix includes optimizing the modified relation matrix. By sequentially performing sparsification and normalization processes, the optimized adjacency matrix corresponding to the optimized graph structure is obtained. The sparsification process employs at least one of threshold filtering, Top-k retention, or differentiable edge selection.

6. The multi-source information clustering method according to claim 5, characterized in that, The step of performing feature encoding on the feature data based on the optimized graph structure to obtain the sample embedding representation is specifically as follows: Temporally gated convolution is used to process the multidimensional time series feature data. Encode to obtain time feature representation The calculation formula is as follows: ， ； in For temporal convolution operations, For temporal convolution kernels, For bias terms, , These are the two components of the output from the temporal convolution. For element-wise multiplication, For the Sigmoid activation function, Represented by time characteristics; Representing the time features With the optimized adjacency matrix Input graph relation modeling, obtain sample representations after graph relation modeling The calculation formula is as follows: ； in , These represent the learnable weights and biases for graph relationship modeling. It is a non-linear activation function; The sample representation Flattening yields the sample embedding representation, calculated using the following formula: ,in For the first The sample embedding representation of each sample. This represents the number of samples.

7. The multi-source information clustering method according to claim 6, characterized in that, The step of clustering based on the sample embedding representation to obtain the sample clustering results is as follows: According to the sample embedding representation K-means clustering is performed on multiple samples to obtain several clusters; and a clustering loss function is constructed. The clustering loss function is calculated to ensure that the embeddings of samples within the same cluster are close to each other, while the embeddings of samples from different clusters are separate from each other. ,in For the sample size, For the first The sample embedding representation of each sample. For the sample The index of the cluster to which it belongs. For the sample The cluster center of the cluster to which it belongs. It is an L2 norm.

8. The multi-source information clustering method according to claim 7, characterized in that, The unified objective function is the total loss function. The total loss function is obtained by weighting the clustering loss function with edge confidence constraints, graph sparsity constraints, and fusion weight constraints. Its calculation formula is as follows: ,in , , These are the weighting parameters used to balance the various constraint terms; The edge confidence constraint term To constrain the consistency between the edge confidence and the sample clustering results, a binary classification cross-entropy form is used, and its calculation formula is as follows: ； in For the sample With sample The edge confidence between them For tags and Indicates sample With sample Belonging to the same cluster This indicates that they belong to different clusters. The summation is a logarithmic function that iterates through the sample pairs involved in the constraints. ; The graph sparsity constraint term The L1 norm of the optimized adjacency matrix is ​​calculated using the following formula: ； in The optimized adjacency matrix is... Let L1 be the L1 norm of the matrix; The fusion weight constraint term The fusion weights are used to constrain the multi-source background knowledge. The weighted squared penalty, negative entropy constraint, or distance constraint between the uniform distribution and the weighted squared penalty is at least one of the following: weighted squared penalty, negative entropy constraint, or distance constraint between the weighted squared penalty and the uniform distribution.

9. A multi-source information clustering device, characterized in that, include: The data acquisition module is used to acquire feature data of multiple samples and multiple background knowledge matrices, wherein the background knowledge matrices are used to characterize the correlation between samples; An adaptive graph learning module is used to set learnable fusion weights for each of the background knowledge matrices, and to perform weighted fusion to obtain a comprehensive relationship matrix; and to evaluate the edge confidence between samples based on the feature data. The comprehensive relation matrix is ​​corrected using the edge confidence to obtain a corrected relation matrix, and the graph structure of the corrected relation matrix is ​​optimized to obtain an optimized graph structure. The feature encoding module is used to perform feature encoding on the feature data based on the optimized graph structure to obtain the sample embedding representation; The clustering module is used to perform clustering based on the sample embedding representation to obtain sample clustering results; The joint training module is used to perform end-to-end joint training on the learning of the fusion weights, the evaluation of the edge confidence, the feature encoding, and the clustering using a unified objective function that includes the clustering objective; and outputs the sample clustering results when the iteration termination condition is met.

10. A multi-source information clustering device, characterized in that, The system includes a memory and a processor, wherein the memory stores a computer program that can be executed by the processor to implement a multi-source information clustering method as described in any one of claims 1 to 8.

Images