Social network text target topic detection method based on sparse subspace clustering

Abstract

A target topic detection method for social network text based on sparse subspace clustering involves the technical fields of computer data processing in natural language processing and social network data mining. This method reduces interference from short text noise and event mismatches in subsequent analysis by writing text identifiers into the target platform's text content and interaction event records, followed by word segmentation, denoising, and vectorization. A social relationship graph is then constructed, and edge credibility weights are calculated to form graph regularization constraint parameters, reducing the weight of low-credibility, spamming edges and suppressing relationship noise caused by organized manipulation at its source. Next, sparse representation coefficients are solved, and a similarity matrix is ​​constructed to perform spectral clustering, enhancing the separability of samples with similar semantics but different propagation patterns. Finally, semantic cohesion and propagation deviation are calculated, and target topic clusters are determined using a dual threshold. Key text sets and propagation evidence sets are extracted to form verifiable target topic detection results, thereby improving the target topic detection rate and reducing false positives.

Description

Technical Field

[0001] This invention relates to the technical field of computer data processing in natural language processing and social network data mining, and in particular to a method for detecting target topics in social network text based on sparse subspace clustering. Background Technology

[0002] With the widespread adoption of social media and mobile internet, social network texts have become a crucial data source for public opinion analysis, content governance, and risk warning. Early detection of anomalous topics relied heavily on keyword statistics, rule triggers, or traditional clustering methods. Subsequently, vector space models, topic models, and deep semantic representations were gradually introduced, improving the semantic aggregatability of short texts. Simultaneously, interactive behaviors such as forwarding, commenting, and mentioning on social platforms are used to assist in propagation structure analysis, thereby enhancing the understanding of topic evolution and diffusion paths. However, in real-world scenarios, anomalous topics often share semantic similarities but exhibit different propagation patterns. Common characteristics, such as centralized forwarding due to organized manipulation, repeated triggering in a spamming manner, and cross-account collaborative diffusion, make it easy for clustering based solely on semantic similarity to mix abnormal propagation with normal discussion in the same cluster. On the other hand, methods based solely on social relationships or community segmentation are easily affected by noise edges, zombie accounts, and non-genuine interactions, resulting in cluster boundary drift and an increase in false alarms. In particular, when there are a large number of low-credibility events in the interaction relationships, the pull of the propagation structure on the clustering results may conflict with semantic similarity, making it difficult to stably separate abnormal topics and lacking verifiable propagation evidence output, which is not conducive to subsequent handling and engineering deployment.

[0003] CN103793489A discloses a method for discovering community topics in online social networks. It collects social network data and discovers topics based on community segmentation. However, it focuses on community structure and topic extraction process, and does not convert the forwarding sequence, repeated triggering and other behavioral statistics in interactive events into computable edge credibility weights. It also lacks a mechanism to introduce edge credibility weights into graph regularization constraints to suppress low credibility relationship noise. Therefore, in the context of organized manipulation and spamming scenarios with similar semantics but abnormal propagation, it is difficult to form stable abnormal topic cluster boundaries.

[0004] CN103514183A discloses an information retrieval method and system based on interactive document clustering. However, its interaction constraints mainly come from user feedback and feature word reconstruction. It does not construct a social relationship graph with text identifiers as nodes for social network interaction events, nor does it provide a method for generating edge trust weights based on interaction interval, repeated forwarding ratio and account interaction stability and entering graph regularization constraints. Therefore, its ability to constrain propagation noise and abnormal diffusion patterns is limited. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0006] Given the shortcomings of existing community topic discovery and document spectrum clustering techniques in terms of noise suppression, semantic and propagation coupling modeling, and evidence output, this invention addresses the problems of difficulty in separating abnormal topics due to semantic similarity but different propagation patterns, and the noise edges pulling the clustering boundaries due to spamming and collaborative manipulation. It proposes a social network text target topic detection method based on sparse subspace clustering. This method constructs a social relationship graph with text identifiers as nodes and calculates edge credibility weights to form graph regularization constraint parameters. Under graph regularization constraints, sparse representation coefficients are solved and combined with spectrum clustering to generate topic clusters. Finally, target topic clusters are determined using a dual threshold of semantic cohesion and propagation deviation, and key text and propagation evidence are extracted.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] A method for detecting target topics in social network text based on sparse subspace clustering includes the following steps:

[0009] S1. Collect the text content and interaction event records of the target platform, write a text identifier for each piece of text content, then perform word segmentation, noise reduction and vectorization processing on the text content, and merge the interaction event records according to the text identifier to obtain the text embedding sample set and the relationship record set.

[0010] S2. Construct a social relationship graph with text identifiers as nodes based on the relationship record set, and calculate the edge trust weights according to the interaction interval, repeated forwarding ratio and account interaction stability to obtain the graph regularization constraint parameters with edge trust weights.

[0011] S3. Solve for the sparse representation coefficients of the text embedding sample set according to the graph regularization constraint parameters, and construct a similarity matrix accordingly to perform spectral clustering to obtain a set of topic clusters;

[0012] S4. Calculate the semantic cohesion and propagation deviation of each topic cluster in the topic cluster set, and make a judgment based on the preset semantic cohesion threshold and the preset propagation deviation threshold to obtain the target topic cluster set. Then, extract the key text set and the propagation evidence set from the target topic cluster set to generate the target topic detection result.

[0013] As one implementation method, the text embedding sample set and the relation record set are obtained through the following sub-steps:

[0014] S1.1 The social network text processing server collects text content and interaction event records of the target platform within a preset sampling period, and writes a unique text identifier for each piece of text content to obtain a set of texts and a set of interaction events with text identifiers.

[0015] S1.2. Perform word segmentation and noise reduction on the text set with text tags in sequence to obtain a normalized text set;

[0016] S1.3. Vectorize the normalized text set and output the text embedding vectors that correspond one-to-one with the text identifiers to obtain the text embedding sample set.

[0017] S1.4 Merge the set of interactive events according to text identifiers, and perform a text identifier propagation chain continuity check on the merged result. Remove event records with missing text identifiers or discontinuous propagation chains to obtain a set of relationship records with the same text identifiers as the text embedding sample set.

[0018] As one implementation method, the graph regularization constraint parameters are obtained through the following sub-steps:

[0019] S2.1 The graph construction module extracts text nodes and text interaction edges based on the relationship record set. The text nodes are text identifiers, and the text interaction edges are text node pairs triggered by forwarding, commenting, or @, to obtain the initial node set and initial edge set of the social relationship graph.

[0020] S2.2. For each interactive edge in the initial edge set, calculate the interaction interval and the repeated forwarding ratio. The interaction interval is the time difference between the occurrence of the corresponding interactive event, and the repeated forwarding ratio is the proportion of the same text identifier being repeatedly triggered by the same account within a preset time window, thus obtaining the edge statistical feature set.

[0021] S2.3 Calculate the account interaction stability for each interaction edge in the initial edge set. The account interaction stability is jointly determined by the fluctuation of the interaction frequency between the same account identifier and the same target node pair within a preset time window and the dispersion of the interaction object, thus obtaining the edge stability feature set.

[0022] S2.4 Calculate the edge confidence weights based on the edge statistical feature set and the edge stable feature set, and write the edge confidence weights into the social relationship graph to form a weighted social relationship graph. Generate graph regularization constraint parameters from the weighted social relationship graph, where the graph regularization constraint parameters include the node adjacency relationship and the edge confidence weight matrix.

[0023] As one implementation method, the topic cluster set is obtained through the following sub-steps:

[0024] S3.1. Generate a sample index based on the text embedding sample set, and align the node adjacency relationship and edge confidence weight matrix in the graph regularization constraint parameters to the text embedding sample set according to the sample index, so as to obtain a graph constraint set with the same index as the text embedding sample set.

[0025] S3.2. Based on the graph constraint set, perform sparse representation solution on the text embedding sample set to obtain a sparse representation coefficient matrix that corresponds one-to-one with the sample index. The difference between the coefficients of adjacent nodes in the sparse representation coefficient matrix decreases as the edge confidence weight increases.

[0026] S3.3 Construct a symmetric similarity matrix based on the sparse representation coefficient matrix, and normalize the similarity matrix to obtain the spectral clustering input matrix;

[0027] S3.4 Perform eigenvalue decomposition on the spectral clustering input matrix and perform clustering partitioning, outputting topic cluster identifiers corresponding to sample indices to obtain a set of topic clusters.

[0028] As one implementation method, the generated target topic detection results are as follows:

[0029] S4.1 Extract the set of text embedding vectors and the set of interaction edges within each topic cluster in the topic cluster set to obtain a cluster-level sample set that corresponds one-to-one with the topic cluster.

[0030] S4.2 Calculate semantic cohesion based on cluster-level sample set, where semantic cohesion is jointly determined by the average distance between text embedding vectors within a cluster and the deviation of the cluster center vector, thus obtaining the semantic cohesion set of topic clusters;

[0031] S4.3 Calculate the propagation deviation based on the cluster-level sample set, where the propagation deviation is jointly determined by the concentration of the credibility weights within the cluster and the decay of the credibility weights outside the cluster, thus obtaining the propagation deviation set of the topic cluster.

[0032] S4.4. For the same topic cluster, compare the semantic cohesion with the preset semantic cohesion threshold and compare the propagation deviation with the preset propagation deviation threshold. When the semantic cohesion meets the preset semantic cohesion threshold and the propagation deviation meets the preset propagation deviation threshold, determine the topic cluster as the target topic cluster and merge the topic clusters determined as target topic clusters into the target topic cluster set.

[0033] S4.5. For each target topic cluster in the target topic cluster set, extract the set of text embedding vectors and the set of interaction edges within the cluster, and retain the one-to-one correspondence with the text identifiers to obtain the target cluster sample set; calculate the cluster center vector for each target topic cluster in the target cluster sample set, and calculate the distance between each text embedding vector within the cluster and the cluster center vector to obtain a representative score, and select a preset number of text identifiers from high to low according to the representative score, and merge them to obtain the key text set; sort the set of interaction edges within the cluster in the target cluster sample set from high to low according to the edge credibility weight, and filter the interaction edges according to the preset edge credibility threshold to obtain the candidate evidence edge set, and perform connectivity merging on the candidate evidence edge set according to the preset maximum hop count to obtain the propagation evidence set; associate the key text set and the propagation evidence set with the corresponding target topic cluster identifiers, write them into the target topic detection results, and output them.

[0034] As one implementation method, the calculation of edge reliability weights based on the edge statistical feature set and the edge stability feature set is as follows:

[0035] For each interactive edge in the edge statistical feature set, the interaction interval and repeated forwarding ratio are obtained, and mapped to interaction score and repeated score respectively according to a preset first tier threshold set. The interaction score increases monotonically with the increase of the interaction interval, and the repeated score decreases monotonically with the increase of the repeated forwarding ratio, thus obtaining the edge statistical score set. For each interactive edge in the edge stability feature set, the account interaction stability is obtained, and mapped to stability score according to a preset second tier threshold set. The stability score increases monotonically with the increase of the account interaction stability, thus obtaining the edge stability score set. For the same interactive edge, the edge statistical score set and the edge stability score set are weighted and fused, and upper and lower bound constraints are applied to the fusion result to obtain the edge trust weight.

[0036] As one implementation method, the mapping of interactive ratings and repeated ratings according to a preset first tier threshold set includes:

[0037] For each interaction edge, obtain the interaction interval and repeated forwarding ratio; compare the interaction interval with the preset first forwarding threshold set in sequence. The preset first forwarding threshold set is arranged from small to large and includes three interaction thresholds. Divide the interaction scores into 4 levels according to the comparison relationship with the three interaction thresholds.

[0038] The duplicate forwarding ratio is compared with the preset first duplicate tier threshold set in turn. The preset first duplicate tier threshold set is arranged from small to large and includes three duplicate thresholds. The duplicate scores are divided into 4 tiers based on the comparison relationship with these three duplicate thresholds.

[0039] The process of mapping the preset second-tier threshold set to a stable score includes: obtaining the account interaction stability for each interaction edge; comparing the account interaction stability with the preset second-tier threshold set in sequence, wherein the preset second-tier threshold set is arranged from small to large and includes three stable thresholds; and dividing the stable score into four tiers based on the comparison relationship with these three stable thresholds.

[0040] The beneficial effects of this invention are as follows: This invention obtains a text embedding sample set and a relationship record set by writing text identifiers into the target platform's text content and interaction event records, and then performing word segmentation, denoising, and vectorization processing. This reduces the interference of short text noise and event mismatches in subsequent analysis. Based on the relationship record set, a social relationship graph with text identifiers as nodes is constructed. The graph regularization constraint parameters are formed by calculating edge credibility weights based on interaction interval, repeated forwarding ratio, and account interaction stability. This reduces the weight of low-credibility spamming edges in the constraints, suppressing relationship noise caused by organized manipulation at the source. Subsequently, sparse representation coefficients are solved under the graph regularization constraints, and a similarity matrix is ​​constructed to perform spectral clustering. This enhances the separability of samples with similar semantics but different propagation patterns, resulting in a more stable set of topic clusters. Finally, semantic cohesion and propagation deviation are calculated, and target topic clusters are determined using dual thresholds. Key text sets and propagation evidence sets are further extracted to form verifiable target topic detection results, thereby improving the target topic detection rate and reducing false positives. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0042] Figure 1 This is a flowchart illustrating the social network text target topic detection method based on sparse subspace clustering as shown in this invention.

[0043] Figure 2 This is the weighted social relationship graph shown in this invention.

[0044] Figure 3 This is a schematic diagram of the clustering results after sparse solution as shown in this invention. Detailed Implementation

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0046] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0047] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0048] According to an embodiment of the present invention, in combination Figure 1 The flowchart shown illustrates a social network text target topic detection method based on sparse subspace clustering running in a computer system. The computer system includes at least a social network text processing server and a data storage unit. The social network text processing server includes a data acquisition module, a preprocessing module, a text embedding generation module, a graph construction and edge confidence weight calculation module, a sparse representation solution module, a spectral clustering module, an anomaly detection module, and an evidence output module. The data storage unit stores the text embedding sample set, relationship record set, weighted social relationship graph, graph regularization constraint parameters, sparse representation coefficient matrix, topic cluster set, and target topic detection results.

[0049] Specifically, this embodiment will be described in conjunction with S1 to S4:

[0050] S1. Collect text content and interaction event records from the target platform, write a text identifier for each piece of text content, then perform word segmentation, noise reduction, and vectorization processing on the text content, and merge the interaction event records according to the text identifiers to obtain a text embedding sample set and a relation record set. Note that the following points should be noted in this step:

[0051] S1.1 The social network text processing server collects text content and interaction event records of the target platform within a preset sampling period, and writes a unique text identifier for each piece of text content, thus obtaining a set of texts and a set of interaction events with text identifiers.

[0052] It should be noted that the data collection module retrieves text content and interaction event records through the target platform's public interface or authorized data interface; the text content includes at least the text body, the publication timestamp, and the publishing account identifier; the interaction event records include at least the event type, the event occurrence timestamp, the triggering account identifier, the source text identifier, and the target text identifier; the triggering account identifier and the publishing account identifier are processed by irreversible hashing to obtain an account desensitization identifier before being entered into the database, and are written into the data collection record along with the data collection batch number.

[0053] For example, the preset sampling period adopts a fixed duration rolling method, such as triggering a collection task every 10 minutes, and each task pulls newly added text content and interaction event records within the most recent 10 minutes; when there is a rate limit on the interface, it pulls by pagination cursor and timestamp breakpoint, and removes duplicate records by event unique key.

[0054] Furthermore, when writing a unique text identifier for each text content, the platform's original text number is preferably used as the primary key. When the platform's original text number is unavailable, a text fingerprint is generated by combining the text's normalized result, the publication timestamp, and the anonymized identifier of the publishing account. This text fingerprint is then used as the text identifier, such as P01-20250101-000123, where P01 is the platform number, 20250101 is the collection date, and 000123 is the sequence number for that day. The text identifier, along with the text's publication timestamp and the anonymized identifier of the publishing account, is written into a text set with the text identifier.

[0055] S1.2. Perform word segmentation and noise reduction on the text set with text identifiers in sequence to obtain a normalized text set.

[0056] In a preferred embodiment, the word segmentation process includes: performing Chinese word segmentation on Chinese text; segmenting English and number fragments by spaces and punctuation; and performing structured extraction on topic tags, links, and @mention fields and retaining them as independent tags.

[0057] In a preferred embodiment, the noise reduction process preferably includes: deleting high-frequency function words from a preset stop word list; merging consecutively repeated characters and consecutively repeated punctuation marks; compressing and recording phrases that appear more than a preset number of times in the same text; and writing noise tags into templated segments that have no semantic load and removing or reducing their weight in subsequent vectorization processing.

[0058] For example, with a threshold of 3 for consecutive repeated characters, text with a templated fragment hit rate of over 80% is written into noise markers.

[0059] S1.3. Vectorize the normalized text set and output the text embedding vectors that correspond one-to-one with the text identifiers to obtain the text embedding sample set.

[0060] It should be noted that the vectorization process in this embodiment adopts a semantic encoding process from word sequence to fixed-length vector: after mapping the normalized text to word sequence, it is input into the pre-trained semantic encoding model, outputting a fixed-length text embedding vector, and writing it into the text embedding sample set in the key-value structure of text identifier-text embedding vector.

[0061] Furthermore, to facilitate the reproduction of the implementation method, the text embedding vector dimension is fixed at 768 dimensions in this embodiment; the ultra-long text is truncated or segmented according to the preset maximum word length, and the segmentation and aggregation method is written into the vector generation record; for example, the text with a length of more than 512 words is segmented into segments of 256 words each, and the average value of the resulting segmented vectors is taken as the text embedding vector of the text.

[0062] S1.4 Merge the set of interactive events according to text identifiers, and perform a text identifier propagation chain continuity check on the merged result. Remove event records with missing text identifiers or discontinuous propagation chains to obtain a set of relationship records with the same text identifiers as the text embedding sample set.

[0063] In a preferred embodiment, when merging by text identifier, it is preferable to establish an event grouping table with the target text identifier as the primary key, and retain an event record sequence of source text identifier, target text identifier, event type, event occurrence timestamp, and triggering account desensitization identifier within each group; when the same event is retrieved multiple times, deduplication is performed using the combination of source text identifier, target text identifier, event type, event occurrence timestamp, and triggering account desensitization identifier as the event unique key.

[0064] It should also be noted that the text identifier propagation chain continuity verification includes at least existence constraints and order constraints: the existence constraint requires that both the source text identifier and the target text identifier in the verification event record can be found in the text set with text identifiers; event records that do not meet the existence constraint are judged as missing text identifier event records and are removed; the order constraint requires that for multiple event records of the same trigger account's desensitized identifier on the same propagation chain, the event occurrence timestamps must satisfy an increasing relationship; event records that do not satisfy the increasing relationship are judged as discontinuous propagation chain event records and are removed.

[0065] As an example, if there exists an event record where text identifier A is triggered by account anonymization identifier U to generate text identifier B, and simultaneously there exists an event record where text identifier B is triggered by account anonymization identifier U at an earlier timestamp to generate text identifier C, then the latter record does not satisfy the order constraint and enters the removal set; after removal, the continuous link event records are retained and enter the relation record set; thus, the relation record set and the text embedding sample set are consistent in the text identifier set, and mismatched edges caused by missing and temporal backfilling are no longer introduced in the subsequent graph construction stage.

[0066] Based on the above S1 process, the relation record set entering the subsequent calculation satisfies the constraints of text identifier existence and propagation chain continuity, thereby reducing the propagation of relation noise in the graph constraint and sparse representation solution stages.

[0067] S2. Construct a social relationship graph with text identifiers as nodes based on the relationship record set, and calculate edge trust weights according to interaction interval, repeated forwarding ratio, and account interaction stability to obtain graph regularization constraint parameters with edge trust weights. Note that the following should be noted in this step:

[0068] S2.1 The graph construction module extracts text nodes and text interaction edges based on the relationship record set. The text nodes are text identifiers, and the text interaction edges are text node pairs triggered by forwarding, commenting, or @, thus obtaining the initial node set and initial edge set of the social relationship graph.

[0069] It should be noted that, in this embodiment, during extraction, it is preferable to write all text identifiers in the text embedding sample set into the node table to form an initial node set, and extract directed pairs of source text identifiers and target text identifiers from the relation record set to form an initial edge set; when there is a record in the relation record where the source text identifier or the target text identifier does not appear in the node table, a missing marker field is written for the record, and the record is filtered out from the initial edge set according to the missing marker field.

[0070] As an example, the missing marker field includes at least the source missing marker, the target missing marker, and the double missing marker.

[0071] S2.2. For each interactive edge in the initial edge set, calculate the interaction interval and the repeated forwarding ratio. The interaction interval is the time difference between the occurrence of the corresponding interactive event, and the repeated forwarding ratio is the proportion of the same text identifier being repeatedly triggered by the same account within a preset time window, thus obtaining the edge statistical feature set.

[0072] It should be noted that when the interaction type is forwarding, the interaction interval is the time difference between forwarding and forwarding; when the interaction type is commenting or @, the interaction interval is the time difference between the timestamp of the interaction event and the timestamp of the source text it references, or the time difference between the timestamps of preceding events in the same chain.

[0073] S2.3 Calculate the account interaction stability for each interaction edge in the initial edge set. The account interaction stability is jointly determined by the fluctuation of the interaction frequency between the same account identifier and the same target node pair within a preset time window and the dispersion of the interaction object, thus obtaining the edge stability feature set.

[0074] It should be noted that the volatility of interaction frequency is preferably characterized by the coefficient of variation of the binning trigger count sequence, and the dispersion of interaction objects is preferably characterized by the normalized entropy of the interaction object distribution. The stability of account interaction is obtained by combining the two; the mathematical expression is as follows:

[0075]

[0076]

[0077]

[0078] in, For account interaction stability; The weighting coefficient for the fluctuation term; The coefficient of variation for the frequency fluctuation of interactions; Normalized entropy for the discreteness of interactive objects; The standard deviation of the sequence of bin trigger counts within the preset time window; The average of the number of times each bin is triggered within a preset time window; It is a smoothing constant; De-identify the account and the first The percentage of times each interactive object interacts; The number of interactive objects that de-identify the account within a preset time window.

[0079] For example, a preset time window of 24 hours is used, and the trigger count sequence is counted in 30-minute intervals; Take 0.6, Pick .

[0080] S2.4 Calculate the edge confidence weights based on the edge statistical feature set and the edge stable feature set, and write the edge confidence weights into the social relationship graph to form a weighted social relationship graph. Generate graph regularization constraint parameters from the weighted social relationship graph, where the graph regularization constraint parameters include the node adjacency relationship and the edge confidence weight matrix.

[0081] In an optional implementation, calculating the edge trust weight includes: obtaining the interaction interval and repeated forwarding ratio for each interaction edge in the edge statistical feature set, and mapping them to an interaction score and a repeated score respectively according to a preset first tier threshold set, wherein the interaction score monotonically increases with the increase of the interaction interval, and the repeated score monotonically decreases with the increase of the repeated forwarding ratio, thus obtaining an edge statistical score set; obtaining the account interaction stability for each interaction edge in the edge stability feature set, and mapping it to a stability score according to a preset second tier threshold set, wherein the stability score monotonically increases with the increase of the account interaction stability, thus obtaining an edge stability score set; and weightedly fusing the edge statistical score set and the edge stability score set for the same interaction edge, and applying upper and lower bound constraints to the fusion result to obtain the edge trust weight.

[0082] Furthermore, the preset first tier threshold set is mapped to interactive ratings and repeated ratings, including:

[0083] For each interaction edge, obtain the interaction interval and repeated forwarding ratio; compare the interaction interval with a preset first forwarding threshold set in sequence. The preset first forwarding threshold set is arranged in ascending order and includes the first interaction threshold, the second interaction threshold, and the third interaction threshold. When the interaction interval is less than the first interaction threshold, determine the first interaction level and assign the first interaction score. When the interaction interval is greater than or equal to the first interaction threshold and less than the second interaction threshold, determine the second interaction level and assign the second interaction score. When the interaction interval is greater than or equal to the second interaction threshold and less than the third interaction threshold, determine the third interaction level and assign the third interaction score. When the interaction interval is greater than or equal to the third interaction threshold, determine the fourth interaction level and assign the fourth interaction score. The first interaction score is less than the second interaction score, less than the third interaction score, and less than the fourth interaction score.

[0084] The duplicate forwarding ratio is compared sequentially with a preset first duplicate tier threshold set, which is arranged in ascending order and includes the first duplicate threshold, the second duplicate threshold, and the third duplicate threshold. When the duplicate forwarding ratio is less than the first duplicate threshold, the first duplicate tier is determined and a first duplicate score is assigned. When the duplicate forwarding ratio is greater than or equal to the first duplicate threshold and less than the second duplicate threshold, the second duplicate tier is determined and a second duplicate score is assigned. When the duplicate forwarding ratio is greater than or equal to the second duplicate threshold and less than the third duplicate threshold, the third duplicate tier is determined and a third duplicate score is assigned. When the duplicate forwarding ratio is greater than or equal to the third duplicate threshold, the fourth duplicate tier is determined and a fourth duplicate score is assigned. The first duplicate score is greater than the second duplicate score, which is greater than the third duplicate score, which is greater than the fourth duplicate score.

[0085] Furthermore, a stable score is mapped to a preset second-tier threshold set, including: obtaining the account interaction stability for each interaction edge; comparing the account interaction stability with the preset second-tier threshold set sequentially, the preset second-tier threshold set being arranged in ascending order and including a first stable threshold, a second stable threshold, and a third stable threshold; determining the first stable tier and assigning a first stable score when the account interaction stability is less than the first stable threshold; determining the second stable tier and assigning a second stable score when the account interaction stability is greater than or equal to the first stable threshold and less than the second stable threshold; determining the third stable tier and assigning a third stable score when the account interaction stability is greater than or equal to the second stable threshold and less than the third stable threshold; and determining the fourth stable tier and assigning a fourth stable score when the account interaction stability is greater than or equal to the third stable threshold, wherein the first stable score is less than the second stable score, less than the third stable score, and less than the fourth stable score.

[0086] As an example, the mathematical expression for edge confidence weight is:

[0087]

[0088] in, For nodes With nodes The trusted weights of the edges interacting with each other; This is the lower bound for the edge's trustworthy weight, such as 0.05; This represents the upper limit of the edge's trust weight, such as 0.95. The interaction rating is weighted, for example, by 0.35; The weight for duplicate ratings is set to 0.35. To stabilize the scoring, a weighting factor of 0.3 is used. Interaction scores obtained by mapping interaction intervals; The repeated score is obtained by mapping the repeated trigger ratio; The stability score is obtained by mapping the stability of account interactions.

[0089] It should also be noted that the preset first tier threshold set and the preset second tier threshold set are preferably determined according to the quantiles of historical data: the 25th, 50th and 75th percentile values ​​of the interaction interval, the repeated trigger ratio and the account interaction stability in the past 7 days are respectively used as candidate values ​​for the first threshold, the second threshold and the third threshold, and the candidate values ​​are then fixed into the tier threshold set after boundary pruning.

[0090] For example, the interaction interval thresholds are set to 30 seconds, 5 minutes, and 60 minutes, respectively, with corresponding interaction scores of 0.2, 0.5, 0.8, and 1.0; the repeat trigger ratio thresholds are set to 0.2, 0.5, and 0.8, respectively, with corresponding repeat scores of 1.0, 0.7, 0.4, and 0.1; and the stability thresholds are set to 0.3, 0.6, and 0.85, respectively, with corresponding stability scores of 0.2, 0.5, 0.8, and 1.0.

[0091] Reference Figure 2 , Figure 2 In this graph, w represents the edge trust weight, and t represents the event type code (e.g., 1 for forward, 2 for comment, 3 for @). When writing the edge trust weight into the social relationship graph, a sparse adjacency structure is preferred for storage. Specifically: using the node table order as the index, the adjacent node index, edge trust weight, and event type code are written into the adjacency table; at the same time, an edge trust weight matrix is ​​generated according to the node index; the node adjacency relationship includes the outgoing edge adjacency set and the incoming edge adjacency set of each node, and the edge trust weight matrix is ​​a sparse weight matrix arranged by node index, where non-zero elements correspond to the positions where there are interactive edges; thus, the graph regularization constraint parameters with edge trust weights are obtained.

[0092] Based on the above S2 process, the interaction interval, repeated trigger ratio and account interaction stability in the interaction event are converted into edge trust weights and entered into the graph regularization constraint parameters, so that the subsequent sparse representation solution stage has a structured suppression path for relation noise, reduces the traction strength of low trust edge pair coefficient structure, and reduces the repeated calculation and cache expansion caused by abnormal edge triggers in online operation.

[0093] S3. Solve for the sparse representation coefficients of the text embedding sample set based on the graph regularization constraint parameters, and construct a similarity matrix accordingly to perform spectral clustering to obtain a set of topic clusters. Note that the following points should be noted in this step:

[0094] S3.1. Generate a sample index based on the text embedding sample set, and align the node adjacency relationship and edge confidence weight matrix in the graph regularization constraint parameters to the text embedding sample set according to the sample index, so as to obtain a graph constraint set with the same index as the text embedding sample set.

[0095] It should be noted that when generating the sample index, this embodiment preferably rearranges the text embedding sample set according to the order of the text identifiers in the node table to obtain an index mapping table that maps the sample index to the text identifier. The sample index takes consecutive integers and is consistent with the row number of the text embedding vector matrix. During alignment, the interaction edges in the relation record set are mapped to row indexes and column indices using a dictionary that maps text identifiers to node indices. The sample index is then re-encoded to obtain a sparse weight structure with the same index system as the text embedding vector matrix, thereby obtaining the graph constraint set.

[0096] S3.2. Based on the graph constraint set, perform sparse representation solution on the text embedding sample set to obtain a sparse representation coefficient matrix that corresponds one-to-one with the sample index. The difference between the coefficients of adjacent nodes in the sparse representation coefficient matrix decreases as the edge confidence weight increases.

[0097] In a preferred embodiment, the sparse representation solution employs a joint objective of reconstruction error, sparse constraints, and graph regularization constraints, and applies zero constraints to the diagonal elements of the coefficient matrix; its mathematical expression is:

[0098]

[0099] in, The optimal sparse representation coefficient matrix obtained by solving for the sparse representation; Embed a vector matrix for the text; The coefficient matrix is ​​sparsely represented. It is the F-norm; It is the sum of the absolute values ​​of the elements; This is the sparsity constraint coefficient, such as 0.8; The graph regularization constraint coefficient is, for example, 0.2; The trace of the matrix; The weighted graph Laplacian matrix is ​​constructed from the edge confidence weight matrix; Let be the diagonal element vector of the sparse representation coefficient matrix C.

[0100] It should also be noted that the weighted graph Laplacian matrix is ​​calculated from the edge confidence weight matrix, thereby generating stronger coefficient consistency constraints in the graph regularization term for adjacency relationships with larger edge confidence weights. This will not be elaborated further in this embodiment.

[0101] S3.3 Construct a symmetric similarity matrix based on the optimal sparse representation coefficient matrix obtained from the solution, and normalize the similarity matrix to obtain the spectral clustering input matrix.

[0102] For example, the mathematical expressions for the symmetric similarity matrix and the normalized spectral clustering input matrix are as follows:

[0103]

[0104]

[0105]

[0106] in, It is a diagonal vector, which has the same meaning as the above formula. To symmetricize similarity matrices; C is a matrix whose element-wise absolute values ​​are taken; D is a degree matrix; Let W be the element in the j-th row and j-th column of the symmetric similarity matrix W; A is the spectral clustering input matrix. It is the negative half power of the degree matrix.

[0107] S3.4 Perform eigenvalue decomposition on the spectral clustering input matrix and perform clustering partitioning, outputting topic cluster identifiers corresponding to sample indices to obtain a set of topic clusters.

[0108] Reference Figure 3In the diagram, nodes represent text identifiers, and cluster regions represent topic clusters output by spectral clustering. Each node corresponds to a unique topic cluster identifier. Specifically, the spectral clustering input matrix undergoes eigenvalue decomposition and clustering, including: constructing a corresponding graph Laplacian matrix from the spectral clustering input matrix, and performing eigenvalue decomposition on the graph Laplacian matrix to obtain an eigenvalue sequence arranged in ascending order; determining the number of clusters based on the spectral gap size of the difference between adjacent eigenvalues ​​in the eigenvalue sequence and a preset spectral gap threshold, and selecting the first few eigenvectors corresponding to the number of clusters to form a feature representation matrix; normalizing the feature representation matrix row-wise to obtain a normalized feature representation matrix; and normalizing the feature representation matrix row-wise. The system performs clustering on the normalized feature representation matrix, extracting sample feature vectors row by row according to the sample index. The Euclidean distance between sample feature vectors is used as the distance metric between samples, and a preset number of clusters is used as the cluster number constraint. The system initializes the set of cluster center vectors. Based on the distance metric between samples, each sample feature vector is assigned to the cluster corresponding to the cluster center vector with the smallest distance, resulting in an initial cluster set. The system updates the set of cluster center vectors based on the initial cluster set and repeats the assignment until the sample set within adjacent clusters no longer changes. It outputs topic cluster identifiers that correspond one-to-one with the sample indices and merges samples with the same topic cluster identifier into the same topic cluster, forming a topic cluster set.

[0109] As an example, the mathematical expression for the Graph Laplace matrix is:

[0110]

[0111] Where L is the graph Laplacian matrix; I is the identity matrix; and A is the spectral clustering input matrix.

[0112] It should be noted that the preset spectral gap threshold in this embodiment is determined according to the quantile of the spectral gap distribution in historical data, such as 0.10. For example, if the first few terms of the feature value sequence are 0.02, 0.03, 0.035, 0.20, and 0.22, then the spectral gap between 0.035 and 0.20 is 0.165, which is greater than the preset spectral gap threshold, and the corresponding cluster number is 3 in this example.

[0113] Based on the above S3 process, the sparse representation coefficients form structured separability under graph regularization constraints. The spectral clustering input matrix is ​​obtained by normalizing the symmetric similarity matrix, thereby improving the stability of topic cluster partitioning in cases where the semantics are similar and the propagation patterns are different, and reducing cluster center drift and repeated iterations caused by relation noise during online operation.

[0114] S4. Calculate the semantic cohesion and propagation deviation for each topic cluster in the topic cluster set, and determine the target topic cluster set using preset semantic cohesion thresholds and preset propagation deviation thresholds. Then, extract the key text set and propagation evidence set from the target topic cluster set to generate the target topic detection result. Note that the following should be noted in this step:

[0115] S4.1 Extract the set of text embedding vectors and the set of interaction edges within each topic cluster in the topic cluster set to obtain a cluster-level sample set that corresponds one-to-one with the topic cluster.

[0116] During extraction, the list of text identifiers for each topic cluster is extracted from the text embedding sample set, and the corresponding text embedding vectors are extracted to form a set of text embedding vectors within the cluster. At the same time, interaction edges in the weighted social relationship graph where both the source text identifier and the target text identifier belong to the same cluster are extracted to form a set of interaction edges within the cluster, and the edge confidence weight and event type encoding of each interaction edge are retained.

[0117] S4.2 Calculate semantic cohesion based on cluster-level sample set, where semantic cohesion is jointly determined by the average distance between text embedding vectors within a cluster and the deviation of the cluster center vector, thus obtaining the semantic cohesion set of topic clusters.

[0118] For example, the mathematical expression for semantic cohesion is:

[0119]

[0120]

[0121]

[0122] in, For the first The semantic cohesion of a topic cluster; This is the combined weighting coefficient, such as 0.5; is the average Euclidean distance between the text embedding vectors within the cluster; This represents cluster center deviation. For the first The number of texts per topic cluster; For the first The first topic cluster The text embedding vector of each text; For the first The first topic cluster The text embedding vector of each text; For the first Cluster center vectors of topic clusters; It is the Euclidean norm.

[0123] S4.3 Calculate the propagation deviation based on the cluster-level sample set, where the propagation deviation is jointly determined by the concentration of the credibility weights within the cluster and the decay of the credibility weights outside the cluster, thus obtaining the propagation deviation set of the topic cluster.

[0124] For example, the mathematical expression for propagation deviation is:

[0125]

[0126]

[0127]

[0128] in, For the first The propagation deviation of each topic cluster; This is the combined weighting coefficient in the propagation deviation, such as 0.6; For the first The set of intra-cluster interaction edges for each topic cluster; For the first A set of interaction edges connecting a topic cluster to nodes outside the cluster; For interactive edges The edge reliability weights; For example, a smoothing constant, ; The entropy of the edge weight distribution within the cluster; The normalized proportion of edge weights within the cluster; This represents the number of interactive edges within the cluster.

[0129] S4.4. For the same topic cluster, compare the semantic cohesion with the preset semantic cohesion threshold and compare the propagation deviation with the preset propagation deviation threshold. When the semantic cohesion meets the preset semantic cohesion threshold and the propagation deviation meets the preset propagation deviation threshold, determine the topic cluster as the target topic cluster and merge the topic clusters determined as target topic clusters into the target topic cluster set.

[0130] Specifically, the preset semantic cohesion thresholds are defined as a first semantic cohesion threshold and a second semantic cohesion threshold, with the first semantic cohesion threshold being less than the second semantic cohesion threshold. The semantic cohesion of topic clusters is compared step by step. When the semantic cohesion is less than the first semantic cohesion threshold, the semantic cohesion determination flag is determined to be no. When the semantic cohesion is greater than or equal to the second semantic cohesion threshold, the semantic cohesion determination flag is determined to be yes. When the semantic cohesion is greater than or equal to the first semantic cohesion threshold and less than the second semantic cohesion threshold, the semantic cohesion determination flag of the previous sampling period is used in non-initial sampling periods, and the semantic cohesion determination flag is determined to be no in the initial sampling period.

[0131] The preset propagation deviation thresholds are defined as a first propagation deviation threshold and a second propagation deviation threshold, with the first propagation deviation threshold being less than the second propagation deviation threshold. The propagation deviation degree of the topic cluster is compared step by step. When the propagation deviation degree is less than the first propagation deviation threshold, the propagation deviation degree determination flag is determined to be no. When the propagation deviation degree is greater than or equal to the second propagation deviation threshold, the propagation deviation degree determination flag is determined to be yes. When the propagation deviation degree is greater than or equal to the first propagation deviation threshold and less than the second propagation deviation threshold, the propagation deviation degree determination flag of the previous sampling period is used in non-initial sampling periods, and the propagation deviation degree determination flag is determined to be no in the initial sampling period.

[0132] When the semantic cohesion determination flag is yes and the propagation deviation determination flag is yes, the final determination flag of the topic cluster is determined to be yes; when the semantic cohesion determination flag is no or the propagation deviation determination flag is no, the final determination flag of the topic cluster is determined to be no; the topic clusters with the final determination flag being yes are written into the target topic cluster set.

[0133] It should be noted that the threshold selection rules in this embodiment are determined based on the quantiles of the historical normal topic cluster distribution: the first semantic cohesion threshold is the 60th quantile of the semantic cohesion over the past 7 days, and the second semantic cohesion threshold is the 80th quantile of the semantic cohesion over the past 7 days; the first propagation deviation threshold is the 70th quantile of the propagation deviation over the past 7 days, and the second propagation deviation threshold is the 90th quantile of the propagation deviation over the past 7 days, and boundary clipping is performed on extreme values; for example, the first semantic cohesion threshold is 0.55, the second semantic cohesion threshold is 0.70; the first propagation deviation threshold is 0.60, and the second propagation deviation threshold is 0.80.

[0134] When the semantic cohesion or propagation deviation falls within the range between the two thresholds, the corresponding judgment identifier of the previous sampling period is used in non-initial sampling periods. To ensure that the retained objects are traceable, this embodiment performs cluster alignment on the topic cluster identifier in each sampling period: calculate the cluster center vector for each topic cluster in the current sampling period and calculate the similarity with the center vectors of each topic cluster in the previous sampling period. When the maximum similarity is greater than the preset alignment threshold (e.g., 0.85), the current topic cluster is mapped to the topic cluster identifier of the previous sampling period and its historical judgment identifier is used. When the alignment threshold is not met, the judgment identifier is determined according to the rules of the initial sampling period.

[0135] S4.5. For each target topic cluster in the target topic cluster set, extract the set of text embedding vectors and the set of interaction edges within the cluster, and retain the one-to-one correspondence with the text identifiers to obtain the target cluster sample set; calculate the cluster center vector for each target topic cluster in the target cluster sample set, and calculate the distance between each text embedding vector within the cluster and the cluster center vector to obtain a representative score, and select a preset number (e.g., 10) of text identifiers from high to low representative scores, and merge them to obtain the key text set; sort the set of interaction edges within the cluster in the target cluster sample set from high to low according to the edge credibility weight, and filter the interaction edges according to the preset edge credibility threshold to obtain the candidate evidence edge set, and perform connectivity merging on the candidate evidence edge set according to the preset maximum hop count to obtain the propagation evidence set; associate the key text set and the propagation evidence set with the corresponding target topic cluster identifiers and write them into the target topic detection results and output them.

[0136] As an example, the mathematical expression for a representative score is:

[0137]

[0138] in, For the first The first of the target topic clusters A representative score for each text; For the first The first of the target topic clusters The text embedding vector of each text; For the first Cluster center vectors of target topic clusters; This is the distance scaling factor, which is the median of the cluster center deviation within the target topic cluster; This is the smoothing constant.

[0139] Furthermore, in this embodiment, the method for setting the edge credibility threshold is based on the upper limit of the evidence edge size: first, set an upper limit on the number of candidate evidence edges that can be retained for each target topic cluster, such as 200; then, select the threshold corresponding to the quantile that makes the number of selected edges not exceed the upper limit on the edge credibility weight distribution of the intra-cluster interaction edge set of the target topic cluster, and solidify it as the preset edge credibility threshold, such as 0.75.

[0140] It should be noted that the method for setting the maximum number of hops in this embodiment is determined based on a trade-off between interpretability and path noise, such as setting the maximum number of hops to 3.

[0141] It should also be noted that during connectivity merging, the candidate evidence edge set is expanded hierarchically according to the edge connection relationship, starting from the node corresponding to the key text set. Edges whose shortest path length does not exceed the preset maximum number of hops are limited to entering the same evidence connectivity component. For each evidence connectivity component, the node sequence, edge confidence weight sequence, and event type encoding sequence are output and associated with the target topic cluster identifier and written into the target topic detection result. For example, the propagation evidence set contains several evidence connectivity components, and each evidence connectivity component contains a propagation path segment with no more than 3 hops, which facilitates subsequent review and coordinated handling.

[0142] Preferably, based on the above S4 process, the target topic cluster is obtained by a dual threshold determination of semantic cohesion and propagation deviation, and the key text set and propagation evidence set are solidified into an auditable output structure; on the computer system side, the number of candidate evidence edges is constrained by a preset edge credibility threshold and upper limit, and the connectivity merging is constrained by a preset maximum number of hops, so that the cache size and number of calculations in the evidence construction stage fall within a controllable range, thereby reducing the frequency of secondary calculations triggered by false alarm writes and the peak fluctuation of memory during online operation.

[0143] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A social network text target topic detection method based on sparse subspace clustering, characterized in that, Includes the following steps: S1. Collect the text content and interaction event records of the target platform, write a text identifier for each piece of text content, then perform word segmentation, noise reduction and vectorization processing on the text content, and merge the interaction event records according to the text identifier to obtain the text embedding sample set and the relationship record set. S2. Construct a social relationship graph with text identifiers as nodes based on the relationship record set, and calculate the edge trust weights according to the interaction interval, repeated forwarding ratio and account interaction stability to obtain the graph regularization constraint parameters with edge trust weights. S3. Solve for the sparse representation coefficients of the text embedding sample set according to the graph regularization constraint parameters, and construct a similarity matrix accordingly to perform spectral clustering to obtain a set of topic clusters; S4. Calculate the semantic cohesion and propagation deviation for each topic cluster in the topic cluster set, and make a judgment based on the preset semantic cohesion threshold and the preset propagation deviation threshold to obtain the target topic cluster set. Then, extract the key text set and the propagation evidence set from the target topic cluster set to generate the target topic detection result. The specific results of the generated target topic detection are as follows: S4.1 Extract the set of text embedding vectors and the set of interaction edges within each topic cluster in the topic cluster set to obtain a cluster-level sample set that corresponds one-to-one with the topic cluster. S4.2 Calculate semantic cohesion based on cluster-level sample set, where semantic cohesion is jointly determined by the average distance between text embedding vectors within a cluster and the deviation of the cluster center vector, thus obtaining the semantic cohesion set of topic clusters; S4.3 Calculate the propagation deviation based on the cluster-level sample set, where the propagation deviation is jointly determined by the concentration of the credibility weights within the cluster and the decay of the credibility weights outside the cluster, thus obtaining the propagation deviation set of the topic cluster. S4.

4. For the same topic cluster, compare the semantic cohesion with the preset semantic cohesion threshold and compare the propagation deviation with the preset propagation deviation threshold. When the semantic cohesion meets the preset semantic cohesion threshold and the propagation deviation meets the preset propagation deviation threshold, determine the topic cluster as the target topic cluster and merge the topic clusters determined as target topic clusters into the target topic cluster set. S4.

5. For each target topic cluster in the target topic cluster set, extract the set of text embedding vectors and the set of interaction edges within the cluster, and retain the one-to-one correspondence with the text identifiers to obtain the target cluster sample set; calculate the cluster center vector for each target topic cluster in the target cluster sample set, and calculate the distance between each text embedding vector within the cluster and the cluster center vector to obtain a representative score, and select a preset number of text identifiers from high to low according to the representative score, and merge them to obtain the key text set; sort the set of interaction edges within the cluster in the target cluster sample set from high to low according to the edge credibility weight, and filter the interaction edges according to the preset edge credibility threshold to obtain the candidate evidence edge set, and perform connectivity merging on the candidate evidence edge set according to the preset maximum hop count to obtain the propagation evidence set; associate the key text set and the propagation evidence set with the corresponding target topic cluster identifiers, write them into the target topic detection results, and output them.

2. The method of claim 1, wherein, The text embedding sample set and relation record set are obtained through the following sub-steps: S1.1 The social network text processing server collects text content and interaction event records of the target platform within a preset sampling period, and writes a unique text identifier for each piece of text content to obtain a set of texts and a set of interaction events with text identifiers. S1.

2. Perform word segmentation and noise reduction on the text set with text tags in sequence to obtain a normalized text set; S1.

3. Vectorize the normalized text set and output the text embedding vectors that correspond one-to-one with the text identifiers to obtain the text embedding sample set. S1.4 Merge the set of interactive events according to text identifiers, and perform a text identifier propagation chain continuity check on the merged result. Remove event records with missing text identifiers or discontinuous propagation chains to obtain a set of relationship records with the same text identifiers as the text embedding sample set.

3. The method for detecting target topics in social network text based on sparse subspace clustering according to claim 2, characterized in that, The graph regularization constraint parameters are obtained through the following sub-steps: S2.1 The graph construction module extracts text nodes and text interaction edges based on the relationship record set. The text nodes are text identifiers, and the text interaction edges are text node pairs triggered by forwarding, commenting, or @, to obtain the initial node set and initial edge set of the social relationship graph. S2.

2. For each interactive edge in the initial edge set, calculate the interaction interval and the repeated forwarding ratio. The interaction interval is the time difference between the occurrence of the corresponding interactive event, and the repeated forwarding ratio is the proportion of the same text identifier being repeatedly triggered by the same account within a preset time window, thus obtaining the edge statistical feature set. S2.3 Calculate the account interaction stability for each interaction edge in the initial edge set. The account interaction stability is jointly determined by the fluctuation of the interaction frequency between the same account identifier and the same target node pair within a preset time window and the dispersion of the interaction object, thus obtaining the edge stability feature set. S2.4 Calculate the edge confidence weights based on the edge statistical feature set and the edge stable feature set, and write the edge confidence weights into the social relationship graph to form a weighted social relationship graph. Generate graph regularization constraint parameters from the weighted social relationship graph, where the graph regularization constraint parameters include the node adjacency relationship and the edge confidence weight matrix.

4. The method for detecting target topics in social network text based on sparse subspace clustering according to claim 3, characterized in that, The topic cluster set is obtained through the following sub-steps: S3.

1. Generate a sample index based on the text embedding sample set, and align the node adjacency relationship and edge confidence weight matrix in the graph regularization constraint parameters to the text embedding sample set according to the sample index, so as to obtain a graph constraint set with the same index as the text embedding sample set. S3.

2. Based on the graph constraint set, perform sparse representation solution on the text embedding sample set to obtain a sparse representation coefficient matrix that corresponds one-to-one with the sample index. The difference between the coefficients of adjacent nodes in the sparse representation coefficient matrix decreases as the edge confidence weight increases. S3.3 Construct a symmetric similarity matrix based on the sparse representation coefficient matrix, and normalize the similarity matrix to obtain the spectral clustering input matrix; S3.4 Perform eigenvalue decomposition on the spectral clustering input matrix and perform clustering partitioning, outputting topic cluster identifiers corresponding to sample indices to obtain a set of topic clusters.

5. The method for detecting target topics in social network text based on sparse subspace clustering according to claim 4, characterized in that, The specific steps for calculating the edge confidence weight based on the edge statistical feature set and the edge stable feature set are as follows: For each interactive edge in the edge statistical feature set, the interaction interval and repeated forwarding ratio are obtained, and mapped to interaction score and repeated score respectively according to a preset first tier threshold set. The interaction score increases monotonically with the increase of the interaction interval, and the repeated score decreases monotonically with the increase of the repeated forwarding ratio, thus obtaining the edge statistical score set. For each interactive edge in the edge stability feature set, the account interaction stability is obtained, and mapped to stability score according to a preset second tier threshold set. The stability score increases monotonically with the increase of the account interaction stability, thus obtaining the edge stability score set. For the same interactive edge, the edge statistical score set and the edge stability score set are weighted and fused, and upper and lower bound constraints are applied to the fusion result to obtain the edge trust weight.

6. The method for detecting target topics in social network text based on sparse subspace clustering according to claim 5, characterized in that, The step of performing eigenvalue decomposition and clustering on the spectral clustering input matrix includes: constructing a corresponding graph Laplacian matrix for the spectral clustering input matrix, and performing eigenvalue decomposition on the graph Laplacian matrix to obtain an eigenvalue sequence arranged in ascending order; determining the number of clusters based on the spectral gap size of the difference between adjacent eigenvalues ​​in the eigenvalue sequence and a preset spectral gap threshold, and selecting the first few eigenvectors corresponding to the number of clusters to form a feature representation matrix; normalizing the feature representation matrix row-wise to obtain a normalized feature representation matrix; and performing clustering on the normalized feature representation matrix, wherein the normalized feature representation... The matrix extracts sample feature vectors row by row according to the sample index; the Euclidean distance between sample feature vectors is used as the distance metric between samples, and the preset number of clusters is used as the cluster number constraint; the set of cluster center vectors is initialized; each sample feature vector is assigned to the cluster corresponding to the cluster center vector with the smallest distance according to the distance metric between samples to obtain the initial cluster set; the set of cluster center vectors is updated based on the initial cluster set and the assignment is repeated until the sample set in adjacent clusters no longer changes, the topic cluster identifier corresponding to the sample index is output, and samples with the same topic cluster identifier are merged into the same topic cluster to form a topic cluster set.

7. The method for detecting target topics in social network text based on sparse subspace clustering according to claim 6, characterized in that, The mapping of interactive ratings and repeated ratings according to the preset first tier threshold set includes: For each interaction edge, obtain the interaction interval and repeated forwarding ratio; compare the interaction interval with the preset first forwarding threshold set in sequence. The preset first forwarding threshold set is arranged from small to large and includes three interaction thresholds. Divide the interaction scores into 4 levels according to the comparison relationship with the three interaction thresholds. The duplicate forwarding ratio is compared with the preset first duplicate tier threshold set in turn. The preset first duplicate tier threshold set is arranged from small to large and includes three duplicate thresholds. The duplicate scores are divided into 4 tiers based on the comparison relationship with these three duplicate thresholds. The process of mapping the preset second-tier threshold set to a stable score includes: obtaining the account interaction stability for each interaction edge; comparing the account interaction stability with the preset second-tier threshold set in sequence, wherein the preset second-tier threshold set is arranged from small to large and includes three stable thresholds; and dividing the stable score into four tiers based on the comparison relationship with these three stable thresholds.

8. The method for detecting target topics in social network text based on sparse subspace clustering according to claim 7, characterized in that, The target topic cluster set is obtained by judging based on a preset semantic cohesion threshold and a preset propagation deviation threshold, as follows: The preset semantic cohesion thresholds are defined as a first semantic cohesion threshold and a second semantic cohesion threshold, with the first semantic cohesion threshold being less than the second semantic cohesion threshold. The semantic cohesion of topic clusters is compared step by step. When the semantic cohesion is less than the first semantic cohesion threshold, the semantic cohesion determination flag is determined to be no. When the semantic cohesion is greater than or equal to the second semantic cohesion threshold, the semantic cohesion determination flag is determined to be yes. When the semantic cohesion is greater than or equal to the first semantic cohesion threshold and less than the second semantic cohesion threshold, the semantic cohesion determination flag of the previous sampling period is used in non-initial sampling periods, and the semantic cohesion determination flag is determined to be no in the initial sampling period. The preset propagation deviation threshold is defined as a first propagation deviation threshold and a second propagation deviation threshold, wherein the first propagation deviation threshold is less than the second propagation deviation threshold; The propagation deviation of the topic cluster is compared step by step. When the propagation deviation is less than the first propagation deviation threshold, the propagation deviation judgment mark is determined to be no. When the propagation deviation is greater than or equal to the second propagation deviation threshold, the propagation deviation judgment mark is determined to be yes. When the propagation deviation is greater than or equal to the first propagation deviation threshold and less than the second propagation deviation threshold, the propagation deviation judgment mark of the previous sampling period is used in non-initial sampling periods, and the propagation deviation judgment mark is determined to be no in the initial sampling period. When both the semantic cohesion determination flag and the propagation deviation determination flag are yes, the final determination flag for identifying the topic cluster is yes. When the semantic cohesion determination flag is negative or the propagation deviation determination flag is negative, the final determination flag for the topic cluster is negative. Write the topic clusters that are finally identified as "yes" into the target topic cluster set.

Images