Short text clustering and fuzzy recognition algorithm based on large-scale network online subgraph sampling

Abstract

This invention provides a short text clustering and fuzzy recognition algorithm for large-scale online subgraph sampling, comprising the following steps: Step S1, extraction and preprocessing of training samples; Step S2, construction of the neural network; Step S3, overall clustering prediction; Step S4, fuzzy sample recognition; Step S5, retraining of the neural network. This invention combines short text clustering with a large language model, which not only improves clustering accuracy but also enables the handling of clustering tasks with different themes and classification requirements, significantly reducing the manual cost of data annotation. Furthermore, this invention can annotate fuzzy samples for classification, using K-nearest neighbors combined with minimum spanning trees to assist subgraph sampling in the selection scheme. This utilizes sparse structure to reduce computational costs and exposes the fluctuations of boundary samples through spectral clustering, providing a more comprehensive perspective for fuzzy sample selection and improving the accuracy and interpretability of the clustering results.

Description

Technical Field

[0001] This invention relates to a short text clustering and fuzzy recognition algorithm for large-scale online subgraph sampling of networks, belonging to the field of machine learning technology. Background Technology

[0002] Text clustering is an unsupervised learning technique that aims to automatically group large amounts of unlabeled text, ensuring that texts within the same group have high semantic or thematic similarity, while texts between different groups show significant differences. This technique is widely used in text mining, information retrieval, public opinion analysis, content recommendation, and knowledge organization to uncover the structural features and potential information of unstructured text data.

[0003] Text clustering tasks typically rely on feature representation methods such as word frequency statistics and feature weighting, combined with classic clustering algorithms such as K-means clustering and hierarchical clustering to group texts. Traditional text clustering tasks usually require word frequency statistics of the text, combining the frequency of word occurrences in the complete document set to capture the latent information features of the text, and then further clustering multiple texts based on this.

[0004] In recent years, with the development of natural language processing technology, clustering of short texts has become an important research topic. Semantic representation methods based on text embedding vectors and pre-trained language models have been gradually introduced into text clustering tasks. By obtaining and applying higher quality word embedding vectors, the consistency of clustering results at the semantic level has been improved, and the adaptability of clustering tasks in the short text domain has been enhanced to some extent.

[0005] However, these methods are often limited to longer, semantically complete text paragraphs. For short texts of less than 10 characters, they cannot effectively solve the problems of polysemy and homosemansis. They lack an understanding of the context and deep semantic relationships of words, resulting in poor performance of text clustering in short text clustering scenarios with semantic leaps.

[0006] Furthermore, in real-world applications, traditional text clustering techniques require human intervention with some understanding of the relevant domain. For example, stop word filtering and stemming operations in text preprocessing, as well as the selection and setting of the final number of clusters, vary significantly in different text scenarios. Often, complex manual design and testing are needed to achieve high accuracy, which requires substantial manpower and costs.

[0007] Therefore, there is a need in this field for a low-cost, high-accuracy method for short text clustering and fuzzy sample recognition. Summary of the Invention

[0008] The purpose of this invention is to provide a low-cost, high-accuracy, and time-efficient method for short text clustering and fuzzy sample identification. This method combines traditional neural network text clustering with large language models and outlier identification techniques, thereby providing support for feature information mining and text content analysis of large amounts of short text. While ensuring high clustering accuracy for short texts, it further reduces the model's operating cost compared to traditional methods, greatly improving the efficiency of clustering tasks and enhancing the model's stability and interpretability.

[0009] The technical solution of the present invention is as follows.

[0010] A short text clustering and fuzzy recognition algorithm for large-scale online subgraph sampling of networks includes the following steps:

[0011] Step S1, Data preprocessing and training sample labeling; Representative samples are randomly selected from the overall data without replacement. Short texts were used as training samples; a large language model and a text embedding vector generation model were invoked to obtain text embedding vectors based on the sample descriptions. and clustering tags ;

[0012] Step S2, building the neural network; embedding text into vectors. The clustering labels of short texts are used as input feature vectors for the neural network. As the output of the neural network, a neural network for predicting short text clustering is constructed;

[0013] Step S3, overall cluster prediction; the extracted clusters... The text embedding vectors of the training samples obtained in advance and clustering tags The constructed neural network is input for training, and the optimized neural network parameters are obtained. The trained neural network is then used to obtain the clustering labels of the short texts. ;

[0014] Step S4, fuzzy sample recognition; calculate the text embedding vector of the sample. Nearest neighbor graph, construct minimum spanning tree and by Nearest neighbor-assisted subgraph sampling is performed; the randomly sampled samples are subjected to spectral clustering, and the Hungarian algorithm is used to align the spectral clustering results with the predicted labels; the sample normalization entropy is calculated based on the results of random subgraph sampling, and fuzzy sample labeling is performed.

[0015] Preferably, the algorithm further includes the following steps:

[0016] Step S5: Retrain the neural network; retrain the neural network with the labels after fuzzy sample annotation to obtain a classifier that can perform clustering and fuzzy sample annotation simultaneously through text embedding vectors.

[0017] Preferably, the training samples in step S1 , scenario-based construction semantic clusters ( ), and specifies the minimum sampling density for each class as .

[0018] Preferably, step S2 uses a four-layer fully connected feedforward network as the classifier of the neural network.

[0019] Preferably, in step S3, the Adam optimizer is used to train the neural network, and the learning rate is set to... The number of iterations is greater than 500;

[0020] The network loss function uses normalized exponential loss plus cross-entropy loss:

[0021]

[0022] in The sample obtained by the normalized exponential function belongs to the first The probability of a class.

[0023] Preferably, the calculation in step S4 is... The process of the nearest neighbor graph further includes:

[0024] Let the embedding vector matrix of all short text samples be... Calculate the pairwise distance between sample pairs using cosine distance. The similarity map of the samples is obtained. ,

[0025]

[0026] Smaller cosine distances correspond to stronger semantic proximity; a sparse distance graph is obtained by preserving local geometric structure using a K-nearest neighbor sparsification strategy. .

[0027] Preferably, the process of constructing the minimum spanning tree in step S4 further includes:

[0028] With diagram Given a fully connected graph, find an acyclic subgraph that connects all vertices and minimizes the total distance of the edges. .

[0029] Preferably, the subgraph sampling and spectral clustering process in step S4 further includes:

[0030] Will Distance of the edge Considered as weight According to the probability matrix ,

[0031]

[0032] right Image cut For each edge, a connected sample set with a suitable sample size is randomly selected as the sampling subset. Obtain a subset Similarity graph Sparse distance graph And perform normalized spectral clustering;

[0033] right Compute the symmetric normalized Laplace.

[0034]

[0035] Pick smallest The feature vectors are concatenated column by column. Normalize each row, then normalize the matrix. row vectors Clustering yields subset prediction result clusters of spectral clustering.

[0036] Preferably, the process of predicting labels in step S4 further includes:

[0037] Constructing a confusion matrix ,in The result of this spectral clustering is The actual result is The number of samples is used to perform linear regression on the confusion matrix using the Hungarian algorithm, and the results are aligned using the hit maximization rule, mapping the subset prediction result clusters of spectral clustering to the original cluster labels.

[0038] Preferably, the process of calculating the normalized entropy of the samples and performing fuzzy sample labeling in step S4 further includes:

[0039] Repeat the subgraph sampling and spectral clustering process 100 times, the first time... The overall prediction result obtained from training based on random subsampled samples ;

[0040] The samples selected in each subgraph sampling The label that was judged against Counting, using samples Judged to label The total number of times the sample is divided by the total number of times it is selected to obtain the sample. For tags The soft probabilities are then used to obtain the soft probability matrix. ;

[0041] The sample normalized entropy is obtained from soft probability. ,

[0042]

[0043] in The closer a value is to 1, the closer the sample is to the edge of the class, and the easier it is to be classified into different classes. The more unstable the classification, the more likely it is to be a fuzzy sample.

[0044] Through the above technical solutions, this invention constructs a technical system for short text clustering in multiple scenarios. It achieves comprehensive innovation through three key mechanisms: generative semantic enhancement, large-scale network noise reduction enhancement, and fuzzy recognition of subgraph sampling. This provides a high-precision, low-cost solution for clustering various types of short texts that can be applied on a large scale in real time. Specifically, this invention can achieve the following beneficial technical effects.

[0045] First, this invention provides generative semantic completion for the semantic structure deficiencies in short texts, enabling the originally semantically sparse short texts to obtain a more complete and high-dimensional semantic expression, significantly improving clustering accuracy and reducing manual annotation costs, and significantly enhancing the adaptability of the technology under different themes and classification requirements.

[0046] Secondly, this invention constructs a lightweight yet information-rich large-scale semantic network based on K-nearest neighbors and minimum spanning trees. It effectively preserves key semantic structures through sparsification and noise reduction enhancement mechanisms, while reducing the computational cost of network construction and spectral clustering, thus providing an efficient structural foundation for large-scale short text processing.

[0047] Finally, this invention proposes a fuzzy recognition mechanism based on similarity-driven minimum spanning tree cutting and multiple subgraph sampling. By structural perturbation and probabilistic consensus, fuzzy samples are stably captured, enabling boundary samples such as polysemy and semantic jumps to be accurately identified in multiple rounds of sampling, thereby improving the stability and interpretability of clustering results. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the algorithm for short text clustering and fuzzy recognition based on large-scale network online subgraph sampling according to the present invention.

[0049] Figure 2 This is a schematic diagram of the algorithm flow for short text clustering and fuzzy recognition based on large-scale online subgraph sampling of networks with real-time requirements, according to the present invention.

[0050] Figure 3 yes Figure 1 and Figure 2 The diagram illustrates step four, which is the method for amplifying fuzzy samples in short text clustering based on minimum spanning tree perturbation. Detailed Implementation

[0051] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0052] In fields such as text mining, information retrieval, public opinion analysis, content recommendation, and knowledge organization, a large number of short texts with fewer than 10 characters often coexist. Due to their limited length, these texts often appear in the form of phrases or even words, and their semantics lack the coherence and completeness of longer paragraphs, with inconsistent and incomplete grammatical functions. As an unsupervised learning clustering technique, short text clustering aims to automatically classify these types of short texts according to their different semantic categories or themes, grouping texts with similar meanings, belonging to the same category, or sharing the same theme into the same class based on specific needs. The short text accurate clustering technology of this invention can further improve the accuracy and interpretability of classification at low cost, enhancing the quality of classification results and facilitating subsequent information extraction and situation analysis.

[0053] The short text clustering task scenario involved in this invention is as follows: texts with fewer than 10 characters are defined as short texts, and there are n short texts (i.e., Chinese texts with fewer than 10 characters) of data. Suppose these short texts can be divided into several categories based on some specific classification criteria (different semantic categories or themes of expression, etc.). This technology aims to obtain the category of short text at low cost while ensuring classification accuracy. Clustering tags During the classification process, it is necessary to obtain... of 3D text embedding vector ,Right now Due to factors such as polysemy, homonymy, and text length limitations, some short texts simultaneously possess characteristics of two categories, making it impossible to definitively classify them into either category. These short texts are defined as fuzzy samples. This invention will obtain clustering labels... Under the premise of further normalized entropy Perform fuzzy sample identification.

[0054] The short text clustering algorithm in this invention combines Large Language Model (LLM) with deep learning, using data annotation based on the LLM to build a neural network. In practical applications, single text data of less than 10 characters is called "short text," and the clustering scheme of this invention can... short text Classified according to the required semantic category or expressive theme The class, and soon Labeling clusters ,in In addition, this invention calculates the stability score of each short text annotation result based on random subsampling and the minimum spanning tree model, and selects short texts with ambiguous classification for annotation, thereby further improving the stability of the classification results.

[0055] Example 1

[0056] like Figure 1 As shown, this embodiment provides a short text clustering and fuzzy recognition algorithm for large-scale online subgraph sampling, including the following steps.

[0057] The first step is data preprocessing and training sample labeling. This involves randomly selecting representative samples from the overall dataset without replacement. A short text is used as a training sample. A large language model and a text embedding vector generation model are invoked to obtain text embedding vectors based on the sample description. and clustering tags .

[0058] The second step is to build the neural network. This involves embedding the text into vectors. Using the short text's clustering labels as the input feature vector to the neural network, a neural network classifier for predicting short text clusters is constructed. To enhance clustering accuracy, in a preferred embodiment, the neural network classifier employs a four-layer fully connected feedforward network.

[0059] The third step is overall cluster prediction. This involves extracting... The neural network is trained using pre-acquired text embedding vectors and cluster labels from training samples, ultimately obtaining the optimized neural network parameters. The trained neural network is then used to obtain cluster labels for short texts. .

[0060] The fourth step is fuzzy sample recognition. This involves calculating the text embedding vector of the sample. Nearest neighbor graph, construct minimum spanning tree and by Nearest neighbor-assisted subgraph sampling is performed. Randomly sampled samples are then subjected to spectral clustering, and the Hungarian algorithm is used to align the spectral clustering results with the predicted labels. The results of random subgraph sampling are used to calculate the sample normalized entropy, which reflects the clustering stability of each sample, and fuzzy sample labeling is performed.

[0061] Example 2

[0062] The difference between this embodiment and Embodiment 1 is that, to meet the real-time requirements, the following steps are added to Embodiment 1; all other steps are exactly the same as in Embodiment 1, such as... Figure 2 As shown.

[0063] The fifth step is to retrain the neural network. If real-time requirements are needed, the neural network is retrained using the labels after fuzzy sample annotation to obtain a classifier that can perform clustering and fuzzy sample annotation simultaneously using text embedding vectors.

[0064] Example 3

[0065] This embodiment provides a short text clustering and fuzzy recognition algorithm for large-scale online subgraph sampling, including the following steps.

[0066] Step 1: Data preprocessing and training sample labeling.

[0067] In this technical framework, the construction and preprocessing of training samples are regarded as the basic steps of the semantic clustering system. Its core goal is to ensure that the input samples received by the neural network classifier not only have statistical representativeness, but also present sufficient density, coverage and discriminability at the semantic level, thereby laying a high-quality semantic space foundation for subsequent short text clustering tasks.

[0068] (1) Representative sampling mechanism

[0069] Whether the training samples are representative of the population and whether they comprehensively include samples from various categories largely determines the training and prediction performance of the neural network.

[0070] To ensure the macroscopic statistical stability of the training data, this invention implements randomized sampling without replacement from the overall corpus. The training set size is set. To maintain the robustness of global feature distribution, scene-based construction semantic clusters ( ), and specifies the minimum sampling density for each class as This is to prevent certain semantic clusters from degenerating into low-sample regions during the training phase.

[0071] (2) A short text enhancement framework for semantic completion

[0072] To compensate for the lack of complete semantic structure in short texts due to length limitations, this invention constructs a two-layer collaborative framework for short texts: "text semantic enhancement—embedding generation." After obtaining basic training samples, this invention introduces a semantic enhancement module driven by a large language model to generate multi-level semantic annotations for the training samples.

[0073] Specifically, this invention generates high-quality semantic inference clustering labels by restating and completing the semantics of each text sample using a large language model within a specified semantic context, combined with the current application scenario and specific clustering requirements. l i Short text summaries with extended semantic completion and text embedding vectors x i .

[0074] By injecting this enhancement mechanism for semantic completion of short texts, the originally highly fragmented short texts can be restored to a higher semantic expression density, forming a more stable embedded expression in the semantic space, enabling the neural network to learn a more continuous and content-oriented semantic structure.

[0075] Step 2: Building the neural network.

[0076] Based on the extraction and preprocessing of training samples, and according to the text embedding vector To enhance clustering accuracy, this invention employs a four-layer fully connected feedforward network as the classifier, using the text embedding vectors of training samples and pre-labeled clustering labels as training data, and the text embedding vectors obtained based on sample descriptions as input, outputting predicted clustering labels. Let... The number of categories is The number of neurons in each layer is The random inactivation rate is The table below describes the details of each layer of the network and its main components.

[0077]

[0078] Step 3: Overall cluster prediction.

[0079] Extracted The neural network is trained using pre-acquired text embedding vectors and cluster labels from training samples, ultimately obtaining the optimized neural network parameters. The trained neural network is then used to obtain cluster labels for short texts. .

[0080] In a preferred embodiment, the optimizer of the neural network is the Adam optimizer, and the learning rate is... The number of iterations is greater than 500.

[0081] In a preferred embodiment, the loss function of the neural network uses normalized exponential loss plus cross-entropy loss:

[0082]

[0083] in The sample obtained by the normalized exponential function (softmax function) belongs to the first... The probability of a class.

[0084] Step 4: Identification of fuzzy samples.

[0085] The Chinese language contains many instances of polysemy, and short texts, limited by their length, often exhibit irregular grammatical functions and inconsistent semantic expression. Therefore, when classifying short texts, some may simultaneously contain features of more than one category. Furthermore, real-world text data is highly complex, and many short texts that do not fall within the classification target's scope or lack feature information cannot be effectively filtered out in the early stages of the classification task.

[0086] Therefore, to effectively address the challenges posed by the fragmented language structure of short texts, this invention constructs a fuzzy sample recognition system based on "graph structure reconstruction—random subgraph sampling—entropy-driven uncertainty quantification." This system aims to identify samples with unstable semantic boundaries and ambiguous category classifications in large-scale short text clustering tasks, thereby improving the overall robustness and stability of the clustering model. The specific steps are further described below, such as... Figure 3 As shown.

[0087] (1) Semantic network reconstruction based on similarity graph

[0088] Let the embedding vector matrix of all short text samples be... Calculate the pairwise distance between sample pairs using cosine distance. The similarity map of the samples is obtained. ,

[0089]

[0090] A smaller cosine distance corresponds to a stronger semantic proximity.

[0091] To reduce graph size and computational burden, this invention utilizes a K-nearest neighbor sparsification strategy to preserve local geometry, resulting in a sparse distance graph. A preferred strategy is to select the K nearest points for each sample, and only retain the distances to these K points, thus obtaining a sparse distance map. This operation ensures the fidelity of the local semantic graph structure while avoiding the exponential complexity of full-graph computation, significantly optimizing the engineering cost of building large-scale networks.

[0092] (2) Sparsity semantic denoising enhancement based on minimum spanning tree

[0093] In the figure Find the minimum spanning tree, i.e., the graph. Given a fully connected graph, find an acyclic subgraph that connects all vertices and minimizes the total distance of the edges. Its introduction has a dual purpose:

[0094] 1. Enhance effective semantic paths: Extract core semantic relationships by selecting the "most necessary" edges in the entire graph;

[0095] 2. Suppress noise interference: Unnecessary high-noise edges are automatically eliminated, forming a natural noise reduction mechanism.

[0096] Will Distance of the edge Considered as weight According to the probability matrix ,

[0097]

[0098] (3) Spectral clustering based on minimum spanning tree cutting and subgraph sampling

[0099] According to the probability matrix right Figure Secondary edge cutting, randomly selecting a connected sample set of moderate size as the sampling subgraph. Obtain a subset Similarity graph Sparse distance graph And perform normalized spectral clustering. Compute the symmetric normalized Laplace.

[0100]

[0101] Pick smallest The feature vectors are concatenated column by column. Normalize each row, then normalize the matrix. row vectors Clustering yields subset prediction result clusters of spectral clustering.

[0102] Constructing a confusion matrix ,in The result of this spectral clustering is The actual result is Given the number of samples, we perform linear regression on the confusion matrix using the Hungarian algorithm, align the results using the hit maximization rule, and map the subset prediction clusters of spectral clustering onto the original cluster labels.

[0103] Repeat the above subgraph sampling and spectral clustering process 100 times, the first time... The overall prediction result obtained from training based on random subsampled samples .

[0104] (4) Soft probability matrix and entropy-driven fuzziness measurement

[0105] The samples selected in each subgraph sampling The label that was judged against Counting, using samples Judged to label The total number of times the sample is divided by the total number of times it is selected to obtain the sample. For tags The soft probabilities are then obtained. Finally, the soft probability matrix is ​​obtained. The normalized entropy of the sample is obtained based on soft probability. ,

[0106]

[0107] in The closer the entropy is to 1, the closer the sample is to the class boundary, the less stable the classification, and the more likely it is to be a fuzzy sample. Therefore, the samples are sorted from high to low normalized entropy, and a certain proportion of high entropy samples are sorted according to the fuzzy sample threshold required in practice. The sample is labeled as a fuzzy sample.

[0108] Statistical theory shows that, under certain assumptions, by normalizing entropy... Fuzzy samples and core samples within semantic clusters can be distinguished statistically, as guaranteed by the following theorem.

[0109] Under the assumptions that the semantic distribution of the samples is compact within clusters, segregated between clusters, and that random sampling coverage holds,

[0110] For any core sample in the core sample set T within a real semantic cluster With the number of sampling rounds As it approaches infinity, its soft probability matrix It converges to a unit vector, and thus we have .

[0111] For any fuzzy sample set F, the fuzzy samples Its soft probability matrix Having positive probabilistic quality in at least two distinct categories, the normalized entropy of fuzzy samples has a non-zero lower bound. That is, as the number of sampling rounds increases Approaching infinity, making .

[0112] Therefore, it can be concluded that a threshold exists. This allows for a sufficiently large number of sampling rounds to have... Approaching 1, meaning that core samples and fuzzy samples within a semantic cluster are statistically separable on the normalized entropy index.

[0113] Therefore, the fuzzy sample recognition method proposed in this invention, under the condition of a sufficiently large number of sampling rounds, can achieve statistical separability between semantically fuzzy samples and core samples within semantic clusters through minimum spanning tree structure and probabilistic random cutting, providing a theoretical guarantee for the feasibility of the method and the increase of clustering stability.

[0114] Example 4

[0115] The difference between this embodiment and embodiment 3 is that, in order to meet the requirements of real-time performance, the following steps are added to the basis of embodiment 3, while the other steps are exactly the same as those in embodiment 3.

[0116] Step 5: Retrain the neural network. If real-time requirements exist, retrain the neural network using the labels after fuzzy sample annotation to obtain a classifier that can perform clustering and fuzzy sample annotation simultaneously using text embedding vectors.

[0117] This invention proposes a short text clustering and fuzzy sample recognition algorithm that further integrates the contextual semantic modeling capabilities of large language models on top of the traditional neural network clustering framework. It also introduces an anomaly sample detection mechanism based on distance distribution sampling to achieve accurate recognition of texts with ambiguous semantics and multiple category features. This algorithm not only improves the accuracy and stability of short text clustering but also significantly reduces the reliance on manual parameter tuning and domain knowledge. It achieves high-hit-rate automatic short text clustering while maintaining low computational costs, providing a more interpretable and engineering-valued technical solution for scenarios such as consumption analysis, payment risk control monitoring, user profiling, platform recommendation, and content analysis.

Claims

1. A short text clustering and fuzzy recognition algorithm for large scale network online subgraph sampling, characterized in that, Includes the following steps: Step S1, Data preprocessing and training sample labeling; Representative samples are randomly selected from the overall data without replacement. Short texts were used as training samples; The large language model and the text embedding vector generation model are used to obtain text embedding vectors based on the sample description. and clustering tags ; Step S2, building the neural network; embedding text into vectors. The clustering labels of short texts are used as input feature vectors for the neural network. As the output of the neural network, a neural network for predicting short text clustering is constructed; Step S3, overall cluster prediction; the extracted clusters... The text embedding vectors of the training samples obtained in advance and clustering tags The constructed neural network is input for training, and the optimized neural network parameters are obtained. The trained neural network is then used to obtain the clustering labels of the short texts. ; Step S4, fuzzy sample recognition; calculate the text embedding vector of the sample. Nearest neighbor graph, construct minimum spanning tree and by Nearest neighbor-assisted subgraph sampling; Randomly sampled samples are subjected to spectral clustering, and the Hungarian algorithm is used to align the spectral clustering results with the predicted labels. The normalized entropy of the samples is calculated based on the results of random subgraph sampling, and fuzzy sample labeling is performed. The calculation in step S4 The process of creating a nearest neighbor graph includes: Let the embedding vector matrix of all short text samples be... Calculate the pairwise distance between sample pairs using cosine distance. The similarity map of the samples is obtained. , The K-nearest neighbor sparsification strategy is used to preserve local geometric structure, resulting in a sparse distance graph. ; The process of constructing the minimum spanning tree in step S4 includes: With diagram Given a fully connected graph, find an acyclic subgraph that connects all vertices and minimizes the total distance of the edges. ; The subgraph sampling and spectral clustering process in step S4 includes: Will Distance of the edge Considered as weight According to the probability matrix , right Image cut For each edge, a connected sample set with a suitable sample size is randomly selected as the sampling subset. Obtain a subset Similarity graph Sparse distance graph And perform normalized spectral clustering; right Compute the symmetric normalized Laplace. Pick smallest The feature vectors are concatenated column by column. Normalize each row, then normalize the matrix. row vectors Clustering yields subset prediction result clusters of spectral clustering.

2. The short text clustering and fuzzy recognition algorithm for large-scale network online subgraph sampling according to claim 1, characterized in that, It also includes the following steps: Step S5: Retrain the neural network; retrain the neural network with the labels after fuzzy sample annotation to obtain a classifier that can perform clustering and fuzzy sample annotation simultaneously through text embedding vectors.

3. The short text clustering and fuzzy recognition algorithm for large-scale network online subgraph sampling according to claim 1 or 2, characterized in that, The training samples in step S1 , scenario-based construction Class semantic clusters, in which And specify the minimum sampling density for each class as .

4. The short text clustering and fuzzy recognition algorithm for large-scale network online subgraph sampling according to claim 3, characterized in that, Step S2 employs a four-layer fully connected feedforward network as the classifier of the neural network.

5. The short text clustering and fuzzy recognition algorithm for large-scale network online subgraph sampling according to claim 4, characterized in that, In step S3, the neural network is trained using the Adam optimizer, with a learning rate of [value missing]. The number of iterations is greater than 500; The network's loss function uses normalized exponential and cross-entropy loss: in The sample obtained by the normalized exponential function belongs to the first The probability of a class.

6. The short text clustering and fuzzy recognition algorithm for large-scale network online subgraph sampling according to claim 1, characterized in that, The process of predicting labels in step S4 further includes: Constructing a confusion matrix ,in The result of this spectral clustering is The actual result is The number of samples is used to perform linear regression on the confusion matrix using the Hungarian algorithm, and the results are aligned using the hit maximization rule, mapping the subset prediction result clusters of spectral clustering to the original cluster labels.

7. The short text clustering and fuzzy recognition algorithm for large-scale network online subgraph sampling according to claim 6, characterized in that, The process of calculating the normalized entropy of the samples and performing fuzzy sample labeling in step S4 further includes: Repeat the subgraph sampling and spectral clustering process 100 times, the first time... The overall prediction result obtained from training based on random subsampled samples ; The samples selected in each subgraph sampling The label that was judged against Counting, using samples Judged to label The total number of times the sample is divided by the total number of times it is selected to obtain the sample. For tags The soft probabilities are then used to obtain the soft probability matrix. ; The sample normalized entropy is obtained from soft probability. , in .

Images