A group clue discovery method based on multi-feature fusion

Abstract

The present disclosure provides a group clue discovery method based on multi-feature fusion. It comprises two parts: a key group identification module based on multi-feature fusion and a key group clue discovery module. The key group identification module, based on multi-feature fusion, inputs the positionally encoded word embedding sequence into a multi-layer Transformer network, where it uses a self-attention mechanism for deep feature extraction. A multi-expert hybrid model is then designed to perform feature fusion and discrimination on code value features, text feature representations, and statistical features after dimensionality reduction. Specifically, a gating mechanism is designed to analyze the input features to generate weighting coefficients for each expert network. The key group clue discovery module receives the computational results of the multi-expert hybrid model and uses a large-model-based text clue discovery and sliding window temporal clue discovery method to screen key groups and obtain group screening results.

Description

Technical Field

[0001] The present disclosure relates to the field of computers, and more particularly, to a group clue discovery method based on multi-feature fusion. Background Art

[0002] Due to the complexity, diversity, and real-time nature of platform data, the media used for streaming media information dissemination on the platform are diverse. Key information data is sparsely distributed, has a very low proportion, and is discrete and hidden. This increases the difficulty of identifying groups and obtaining group clues on the platform. Furthermore, clue group data on the platform usually appears in the form of multi-dimensional features, such as text, code values, and statistical information. Most existing methods interactively represent various features and extract key clues from multi-dimensional feature data based on a unified representation. However, due to semantic differences between feature data of different categories, existing methods find it difficult to learn multi-dimensional features with strong correlation attributes, and are unable to utilize the intrinsic connections and dependencies between features of different categories to assist in the discovery of key clues, resulting in a poorer performance of the model in extracting multi-dimensional group features. Summary of the Invention

[0003] The purpose of the embodiments of the present disclosure is to provide a group clue discovery method based on multi-feature fusion. The main research goal of the present invention is to identify and analyze key groups more efficiently.

[0004] First, the present invention identifies key groups by comprehensively analyzing multiple features, including text, coded values, and statistical information. These features include, but are not limited to, text semantics, numerical coding features, and statistical features extracted from historical data. This multi-dimensional information analysis method avoids the limitations of relying solely on a single feature, significantly improving the accuracy and reliability of key group identification.

[0005] After identifying key groups, the present invention further combines large-scale models with time-series information to conduct in-depth mining of the data within the group to uncover hidden important clues. The application of large-scale models enables the present invention to process and understand complex patterns and relationships, while the analysis of time-series information captures the dynamic changes in data over time, thereby identifying clues that stand out at specific points in time or within time periods. This approach not only enables the present invention to discover potentially important information but also to predict and explain the evolving trends of this information, providing strong support for subsequent decision-making.

[0006] In a general aspect, a group clue discovery method based on multi-feature fusion is provided, comprising two parts: a key group identification module based on multi-feature fusion and a key group clue discovery module;

[0007] The multi-feature fusion-based key group identification module first uses the GPT model to extract group text data on the social platform, and pre-processes the text data to convert it into word embedding representation;

[0008] Next, the word embedding sequence after position encoding is input into the multi-layer Transformer network to use the self-attention mechanism for deep feature extraction. The calculation formula is:

[0009]

[0010] Where Q represents the query matrix, which indicates the degree of attention the current word pays to other words. Each word position generates a query vector, which is used to find words related to it. K represents the key matrix, which represents the characteristics or content of the word. Each word position generates a key vector, which is used to match the query vector (Q) to determine the relevance. V represents the value matrix, which represents the value or output information of the word. Each word position generates a value vector, which represents the content information of the word, which will be weighted and summarized into the output. k Represents the dimension of the key vector K, which is a normalization factor used to narrow the value range of the dot product result to prevent excessively large values ​​from causing the softmax function to approach saturation and thus lose discrimination. T represents the transposed matrix of the key matrix K; the deep features include code value features, text features, and statistical features;

[0011] The multi-expert hybrid model first establishes independent expert networks for text features, code value features and statistical features respectively. Each expert network is a multi-layer perceptron that generates representations for text features and reduces the dimensionality of statistical features. Then, a multi-expert hybrid model is designed to perform feature fusion and discrimination on the code value features, the representations of text features, and the statistical features after dimensionality reduction. Specifically, a gating mechanism is designed to generate weighted coefficients for each expert network by analyzing the input features, and dynamically adjust the output weights of each expert network according to the importance of the input features. Finally, the outputs of all expert networks are fused by weighted summation to generate a comprehensive feature representation. The fusion formula is as follows:

[0012]

[0013] Where n is the number of expert networks, F final is the final feature representation after fusion;

[0014] The key group clue discovery module receives the operation results of the multi-expert hybrid model, and screens key groups through text clue discovery based on large model prompts and sliding window time series clue discovery methods, and obtains group screening results.

[0015] The analysis method of the code value features is: focusing on code values ​​that appear repeatedly in multiple groups or have obvious directionality;

[0016] The statistical feature analysis method is to use a multi-layer perceptron model to fuse all statistical features and generate the final statistical feature representation. The calculation formula is as follows:

[0017] F stat =σ(W×X+b)

[0018] Among them, X is the statistical feature vector of the input, W is the weight matrix, b is the bias, and σ is the activation function.

[0019] The text clue discovery based on large model prompts utilizes a pre-trained language model and designs appropriate prompts. When processing prompts, the model comprehensively considers the word frequency, semantic association, and emotional expression in the text to generate output containing key information, thereby mining text clues closely related to group behavior, including the group's charter, emotional tendencies, common topics, hot keywords, and communication style.

[0020] The sliding window temporal clue discovery method dynamically analyzes the temporal dimension of group activities on social platforms to identify specific behaviors. Specifically, the data is divided into overlapping time periods. Within each time period, the group's speech frequency, keyword usage, and user activity indicators are independently analyzed to capture sudden behaviors and hot spots within the group.

[0021] The similarity of the activity patterns of two groups in different time periods is judged. The dynamic time rule method is introduced to measure the time series similarity of the two groups. By calculating the minimum distance between the activity patterns of the groups, the groups with similar activity characteristics in different time periods are identified. The calculation formula is:

[0022]

[0023] Among them, X[i] and Y[j] represent the values ​​of two time series X and Y at the i-th and j-th time points respectively, and n is the length of the time series

[0024] The innovation of the embodiment of the present invention lies in:

[0025] This invention improves the accuracy and responsiveness of group identification across multiple dimensions. By combining text, code values, and statistical features, it enables comprehensive group analysis, effectively overcoming the limitations of single-feature analysis in existing technologies. Furthermore, the time series analysis module can promptly capture group activity during hot events, enhancing both hotness and real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The above and other objects and features of the present disclosure will become more apparent from the following description in conjunction with the accompanying drawings.

[0027] Figure 1 is a schematic diagram illustrating a multi-expert mixture model according to an embodiment of the present disclosure;

[0028] Figure 2 1 is a schematic diagram illustrating an architecture diagram of an open source code repository summary automatic generation model according to an embodiment of the present disclosure;

[0029] Figure 3 is a schematic diagram illustrating dynamic time warping according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0030] The following detailed description is provided to help the reader gain a comprehensive understanding of the methods, devices and / or systems described herein. However, various changes, modifications and equivalents of the methods, devices and / or systems described herein will be clear after understanding the disclosure of the present application. For example, the order of operations described herein is merely an example and is not limited to those orders set forth herein, but can be changed as will be clear after understanding the disclosure of the present application, except for operations that must occur in a specific order. In addition, for greater clarity and conciseness, descriptions of features known in the art may be omitted.

[0031] The features described herein can be implemented in different forms and should not be construed as limited to the examples described herein. Rather, the examples described herein are provided to illustrate only some of the many possible ways to implement the methods, devices, and / or systems described herein, which will become clear after understanding the disclosure of this application.

[0032] As used herein, the term "and / or" includes any one of the associated listed items and any combination of any two or more.

[0033] Although terms such as "first," "second," and "third" may be used herein to describe various members, components, regions, layers, or portions, these members, components, regions, layers, or portions should not be limited by these terms. Instead, these terms are used solely to distinguish one member, component, region, layer, or portion from another member, component, region, layer, or portion. Thus, what is referred to as a first member, first component, first region, first layer, or first portion in the examples described herein may also be referred to as a second member, second component, second region, second layer, or second portion without departing from the teachings of the examples.

[0034] In the specification, when an element (such as a layer, region, or substrate) is described as being “on,” “connected to,” or “coupled to” another element, the element may be directly “on,” “connected to,” or “coupled to” the other element, or one or more other elements may be present therebetween. Conversely, when an element is described as being “directly on,” “directly connected to,” or “directly coupled to” another element, there may be no other elements present therebetween.

[0035] The terms used herein are intended only to describe various examples and are not intended to limit the disclosure. Unless the context clearly indicates otherwise, the singular is intended to include the plural. The terms "comprise," "include," and "have" indicate the presence of the recited features, quantities, operations, components, elements, and / or combinations thereof, but do not preclude the presence or addition of one or more other features, quantities, operations, components, elements, and / or combinations thereof.

[0036] Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure pertains after understanding the present disclosure. Unless expressly defined otherwise herein, terms (such as those defined in general dictionaries) should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and the present disclosure, and should not be interpreted in an idealized or overly formal manner.

[0037] Furthermore, in describing the examples, when it is deemed that a detailed description of well-known related structures or functions would cause ambiguous interpretation of the present disclosure, such detailed description will be omitted.

[0038] Figure 1 3 is a schematic diagram illustrating a group clue discovery method based on multi-feature fusion according to an embodiment of the present disclosure.

[0039] First, the present invention identifies key groups by comprehensively analyzing multiple features, including text, coded values, and statistical information. These features include, but are not limited to, text semantics, numerical coding features, and statistical features extracted from historical data. This multi-dimensional information analysis method avoids the limitations of relying solely on a single feature, significantly improving the accuracy and reliability of key group identification.

[0040] After identifying key groups, the present invention further combines large-scale models with time-series information to conduct in-depth mining of the data within the group to uncover hidden important clues. The application of large-scale models enables the present invention to process and understand complex patterns and relationships, while the analysis of time-series information captures the dynamic changes in data over time, thereby identifying clues that stand out at specific points in time or within time periods. This approach not only enables the present invention to discover potentially important information but also to predict and explain the evolving trends of this information, providing strong support for subsequent decision-making.

[0041] Key group identification based on multi-feature fusion

[0042] In this paper, the GPT model is used to extract features from group text information on social platforms. This text information includes the group name, description, chat content, and more. Extracting text features is a crucial step in group analysis, helping to reveal communication patterns and potential interactions between group members. To ensure data consistency and model accuracy, the text data is first preprocessed, including removing irrelevant characters, standardizing punctuation, and addressing spelling errors. This step ensures a uniform data format for input into the model, reducing the impact of noise on model performance.

[0043] The preprocessed text data is converted into word embedding representations. Next, the positionally encoded word embedding sequence is fed into a multi-layer Transformer network for deep feature extraction. The core mechanism of the Transformer network is the self-attention mechanism, which dynamically focuses on other parts of the text as it processes each word, thereby capturing long-range lexical dependencies. The calculation formula for the self-attention mechanism is:

[0044]

[0045] In this formula, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d k Represents the dimension of the key vector. The self-attention mechanism can effectively integrate the correlation between each word, thereby generating a feature representation rich in contextual information.

[0046] To extract and analyze code value features, we first need to identify all possible code value information from the group's text data. This information may include URL links, hash values, encoded strings, and more. Rule matching and pattern recognition techniques can effectively extract these key code values. The extracted code value information is then further analyzed to identify potential relationships between them. When analyzing code value features, we pay special attention to code values ​​that appear repeatedly in multiple groups or have clear directionality. This type of information is often the key nodes of interaction between groups. By analyzing these nodes, we can more accurately describe the structure of the group network and identify clues to hot activities.

[0047] Statistical features are an important aspect of identifying group behavior patterns. Statistical features cover user behavior data within the group, such as the number of users, speaking frequency, message length, IP distribution, etc. These features can reflect the activity of the group, user composition, and interaction patterns, and help identify hot activities. In order to integrate these statistical information, the present invention uses a multi-layer perceptron (MLP) model to fuse all statistical features and generate the final statistical feature representation. Its core calculation formula is as follows:

[0048] F stat =σ(W×X+b)

[0049] Where X is the input statistical feature vector, W is the weight matrix, b is the bias, and σ is the activation function. This formula can be used to fuse multiple statistical features into a high-dimensional representation for subsequent group analysis.

[0050] To fully utilize the multi-dimensional information of text features, code value features, and statistical features, this paper designs a multi-expert hybrid model for feature fusion and discrimination. This model can flexibly select and combine the most appropriate expert networks based on the characteristics of different features to improve the recognition accuracy of key groups.

[0051] The core idea of ​​the multi-expert hybrid model is to assign features to different expert networks, each of which focuses on processing a specific type of feature, such as text features, code value features, or statistical features. The outputs of each expert network are then weighted and fused through a gating mechanism to obtain a comprehensive feature representation. This approach can effectively cope with the heterogeneity of multi-dimensional feature data and fully explore the potential information of various features, such as Figure 2 shown.

[0052] First, the text features, code value features, and statistical features are processed by independent expert networks. Each expert network is a multi-layer perceptron (MLP). In this way, each expert network can independently extract a deep representation of the input features.

[0053] Next, the gating mechanism dynamically adjusts the output weights of each expert network based on the importance of the input features. The gating network analyzes the input features and generates weighted coefficients for each expert network. These coefficients determine the contribution of each expert network to the final output.

[0054] Finally, the outputs of all expert networks are fused through weighted summation to generate a comprehensive feature representation. The fused feature representation can express the multi-dimensional characteristics of the group, has stronger discriminative power, and can help the model more accurately identify key groups. The fusion formula is as follows:

[0055]

[0056] Where n is the number of expert networks, F final It is the final feature representation after fusion, and the expert model is composed of a stack of multi-layer perceptrons.

[0057] Furthermore, the multi-expert hybrid model is highly scalable. The number of expert networks can be increased or decreased based on actual application requirements, and the structure of each expert network can be optimized for specific applications. This makes the invention widely applicable to data analysis across various social platforms, demonstrating strong robustness and adaptability when dealing with complex and diverse group information.

[0058] Key group clues discovered

[0059] Discovering textual clues based on large-scale model prompts is a key step in group analysis. By leveraging pre-trained language models (such as GPT) and designing appropriate prompts, we can effectively mine textual clues closely related to group behavior. These clues include group charters, sentiment, common topics, hot keywords, and communication styles, helping to identify potential key groups.

[0060] In large language model applications, prompt design directly impacts model performance. Prompts are crucial for guiding the language model to generate the target output. They require careful design based on the specific requirements of the task to ensure the model accurately captures key information. In this paper, we designed a series of highly targeted prompts to extract important text clues from groups.

[0061] Here’s a prompt design for identifying the core topic of a group:

[0062] "Please analyze the following text and list the three most frequently discussed topics, as well as the possible emotional tendencies behind these topics:\n\n'{Group chat history}'\n\nPlease describe each topic and its related keywords in detail."

[0063] This prompt guides the model to analyze the text, not only identifying the main topics but also inferring the sentiment behind them based on the context. This approach allows for a better understanding of the overall group atmosphere and potential risks. When processing this prompt, the model comprehensively considers word frequency, semantic relevance, and emotional expression in the text to generate an output containing key information. For example, in a group discussing a hot topic, the model might identify sentiment related to the event and flag sensitive keywords, providing crucial insight for subsequent identification of key groups.

[0064] In addition, in order to more comprehensively capture the communication style of the group, Prompt is designed to analyze conversation characteristics:

[0065] Please outline common expressions in the following conversations, including any bias or slanting towards specific events:\n\n'{Group chat log}'\n\nAnd indicate any recurring slogans or specific language patterns.

[0066] Using this prompt, the model can identify language patterns, recurring slogans, and potential biases in conversations. This analysis can reveal whether a group's communication style has specific tendencies or other behavioral characteristics, helping to identify potential focus groups in advance.

[0067] The sliding window time series clue discovery method identifies specific behaviors by dynamically analyzing the time dimension of group activities on social platforms. The sliding window technology divides the data into overlapping time periods, and independently analyzes key indicators such as the group's speech frequency, keyword usage, and user activity within each time period, thereby promptly capturing sudden behaviors and hot activities within the group. By setting the time granularity of the sliding window to 1 hour or 30 minutes, a rapid response to hot events is ensured. Within each time window, by comparing the changes in the current indicator with the previous window, hot time periods can be effectively identified. For example, a sudden surge in the number of speeches within a certain time window may indicate that the group is focusing on discussing hot events or planning hot activities.

[0068] Next, we introduce some advanced time series analysis and pattern recognition methods to determine the similarity of the activity patterns of the two groups in different time periods. In particular, the Dynamic Time Warping (DTW) method is used to measure the similarity of the time series of the two groups. DTW is a powerful time series analysis tool that can align and compare time series of different lengths to identify the similarities between them. In the present invention, DTW can effectively identify groups with similar activity characteristics in different time periods by calculating the minimum distance between the activity patterns of the groups. The process is as follows: Figure 3shown.

[0069] The calculation formula for DTW is as follows:

[0070]

[0071] Where X[i] and Y[j] represent the values ​​of two time series X and Y at time points i and j, respectively, and n is the length of the time series. DTW aligns the time series and calculates their similarity by calculating the Euclidean distance of all possible pairs between the two time series and finding the pairing path that minimizes the total distance.

[0072] Combining kurtosis analysis and multi-feature fusion strategy, the present invention can accurately locate high-risk time periods and reveal the temporal patterns of group behavior, providing strong support for the comprehensive analysis of key groups.

[0073] In a more specific embodiment, the present invention provides:

[0074] Key group identification module based on multi-feature fusion

[0075] This paper uses the GPT model to extract features from group text information on social platforms. This text information includes the group name, description, chat content, and more. Extracting text features is a crucial step in group analysis, revealing communication patterns and potential specific behaviors among group members. To ensure data consistency and model accuracy, the text data is first preprocessed, including removing irrelevant characters, standardizing punctuation, and addressing spelling errors. This step ensures a uniform data format for input into the model, reducing the impact of noise on model performance.

[0076] The preprocessed text data is converted into word embedding representations. Next, the positionally encoded word embedding sequence is fed into a multi-layer Transformer network for deep feature extraction. The core mechanism of the Transformer network is the self-attention mechanism, which dynamically focuses on other parts of the text as it processes each word, thereby capturing long-range lexical dependencies. The calculation formula for the self-attention mechanism is:

[0077]

[0078] In this formula, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and \(d_k\) represents the dimension of the key vector. The self-attention mechanism can effectively integrate the correlation between each word to generate a feature representation rich in contextual information.

[0079] If the vector dimension is specified as 512, then after GPT text preprocessing it will be converted into:

[0080] w1=[0.02133,0.10713,.............0.00159]

[0081] In group analysis on social platforms, in order to extract and analyze code value features, it is first necessary to identify all possible code value information from the group's text data. This information may include URL links, hash values, encoded strings, etc. Through rule matching and pattern recognition technology, these key code values ​​can be effectively extracted. The extracted code value information is then further analyzed to identify the potential relationships between them. In the process of parsing code value features, special attention is paid to code values ​​that appear repeatedly in multiple groups or have obvious directionality. This type of information is often the key node of interaction between groups. By analyzing these nodes, the structure of the group network can be more accurately described and clues of hot activities can be identified. Through analysis of the 'China XX Network' group, it was found that there was a link jumping to "https: / / weibo.com / u / 2377587254"

[0082] Statistical features are crucial for identifying group behavior patterns. They encompass user behavior data within a group, such as the number of users, speaking frequency, message length, and IP distribution. These features can reflect a group's activity, user composition, and interaction patterns, helping to identify hot spots. To synthesize this statistical information, the present invention uses a multi-layer perceptron (MLP) model to fuse all statistical features and generate a final statistical feature representation.

[0083] F stat =σ(W×X+b)

[0084] Where X is the input statistical feature vector, W is the weight matrix, b is the bias, and σ is the activation function. This formula can be used to fuse multiple statistical features into a high-dimensional representation for subsequent group analysis.

[0085] To fully utilize the multi-dimensional information of text features, code value features, and statistical features, this paper designs a multi-expert hybrid model for feature fusion and discrimination. This model can flexibly select and combine the most appropriate expert networks based on the characteristics of different features to improve the recognition accuracy of key groups.

[0086] The core idea of ​​the multi-expert hybrid model is to assign features to different expert networks, each of which specializes in processing a specific type of feature, such as text features, code value features, or statistical features. The outputs of each expert network are then weighted and fused using a gating mechanism to produce a comprehensive feature representation. This approach effectively addresses the heterogeneity of multi-dimensional feature data and fully exploits the potential information of each feature.

[0087] First, the text features, code value features, and statistical features are processed by independent expert networks. Each expert network is a multi-layer perceptron (MLP). In this way, each expert network can independently extract a deep representation of the input features.

[0088] Next, the gating mechanism dynamically adjusts the output weights of each expert network based on the importance of the input features. The gating network analyzes the input features and generates weighted coefficients for each expert network. These coefficients determine the contribution of each expert network to the final output.

[0089] Finally, the outputs of all expert networks are fused through weighted summation to generate a comprehensive feature representation. The fused feature representation can express the multi-dimensional characteristics of the group, has stronger discriminative power, and can help the model more accurately identify key groups. The fusion formula is as follows:

[0090]

[0091] Where n is the number of expert networks, F final It is the final feature representation after fusion.

[0092] Furthermore, the multi-expert hybrid model is highly scalable. The number of expert networks can be increased or decreased based on actual application requirements, and the structure of each expert network can be optimized for specific applications. This makes the invention widely applicable to data analysis across various social platforms, demonstrating strong robustness and adaptability when dealing with complex and diverse group information.

[0093] Key group clues discovered

[0094] Discovering textual clues based on large-scale model prompts is a key step in group analysis. By leveraging pre-trained language models (such as GPT) and designing appropriate prompts, we can effectively mine textual clues closely related to group behavior. These clues include group charters, sentiment, common topics, hot keywords, and communication styles, helping to identify potential key groups.

[0095] In large language model applications, prompt design directly impacts model performance. Prompts are key to guiding the language model to generate the target output and require careful design based on the specific requirements of the task to ensure the model accurately captures key information. In this paper, we designed a series of highly targeted prompts to extract important text clues from groups.

[0096] Here’s a prompt design for identifying the core topic of a group:

[0097] "Please analyze the following text and list the three most frequently discussed topics, as well as the possible emotional tendencies behind these topics:\n\n'{Group chat history}'\n\nPlease describe each topic and its related keywords in detail."

[0098] Regarding the chat content on China XX website, we obtained the following clues:

[0099] Topics: Domestic circulation, employment rate, countervailing duty

[0100] Emotional tendency: rational and objective

[0101] Keywords: Internal circulation: [international environment, anti-globalization, going overseas], employment rate: [graduation, civil service exam, takeout], anti-subsidy: [new energy vehicles, dairy products, EU],

[0102] In addition, to more comprehensively capture the communication style of the group, we also designed Prompt to analyze conversation characteristics:

[0103] Please summarize the common expressions used in the following conversations, including any bias towards specific events:\n\n'{Group chat history}'\n\nAnd indicate any recurring content or specific language patterns.

[0104] An analysis of the comments on China's XX website yielded the following clues:

[0105] Is there a bias towards a specific event? "Support for independent research and development of electric vehicles"

[0106] By designing and applying these prompts, we can extract highly representative information from group text content and construct detailed text clue maps of the group. These maps include not only key topics and keywords, but also deeper features such as sentiment and communication style, providing a key basis for subsequent group analysis.

[0107] The sliding window temporal clue discovery method dynamically analyzes the temporal dimension of group activity on social platforms to identify specific behaviors. This sliding window technique divides data into overlapping time periods, independently analyzing key indicators such as group chat frequency, keyword usage, and user activity within each time period to promptly capture emergent behaviors within the group.

[0108] We typically set the sliding window granularity to 1 hour or 30 minutes to ensure a rapid response to hot topics. Within each time window, comparing the changes in the current metric with those in the previous window can effectively identify hot periods. For example, a sudden surge in the number of posts within a certain time window may indicate that the group is focusing on discussing a hot topic or planning a hot event.

[0109] Taking the frequency of speech on China XX website as an example, we get the following clues:

[0110] Suppose we are analyzing the speeches on China XX website on March 11 and June 12, 2024.

[0111] The data shows that there was a sudden increase in the number of speeches during the following two time periods:

[0112] - March 11, 2024, 1:00 PM - 2:00 PM

[0113] -10:00 PM, June 12, 2024 - 1:00 AM, June 13, 2024

[0114] Let the speech time series on March 11 be X, and the speech time series from June 12 to 13 be Y. Assume that X = [100, 150, 200, 50] represents the number of speeches per hour from 10:00 to 14:00 on March 11, and Y = [80, 90, 300, 250, 60] represents the number of speeches per hour from 21:00 on June 12 to 1:00 on June 13. Because the lengths of X and Y are different (X has 4 time points and Y has 5 time points), direct comparison may lose some information. At this time, DTW is used to align the two time series and calculate the minimum distance between them to evaluate the similarity of their activity patterns. The calculation formula of DTW is as follows:

[0115]

[0116] Here, X[i] and Y[j] represent the values ​​of two time series, X and Y, at time points i and j, respectively, and n is the length of the time series. DTW aligns the time series and calculates their similarity by calculating the Euclidean distance of all possible pairs between them and finding the pairing path that minimizes the total distance. The results show the similarity in the frequency of posts on the Chinese XX website between March 11 and June 12-13.

[0117] DTW results can be further combined with other features (such as sentiment analysis of speech content and keyword extraction) for multi-feature fusion analysis to more accurately determine whether these time periods are high-risk periods (for example, whether there are discussions of potential market fluctuations or sudden policy interpretations). Through this analysis, we can provide early warning information to managers of China XX Network, helping them to take timely countermeasures when hot events occur.

[0118] While some embodiments of the present disclosure have been shown and described, it will be appreciated by those skilled in the art that changes may be made to these embodiments without departing from the principles and spirit of the disclosure, the scope of which is defined by the claims and their equivalents.

Claims

1. A group clue discovery method based on multi-feature fusion, characterized in that: It consists of two parts: a key group identification module based on multi-feature fusion and a key group clue discovery module; The multi-feature fusion-based key group identification module first uses the GPT model to extract group text data on the social platform, and pre-processes the text data to convert it into word embedding representation; Next, the word embedding sequence after position encoding is input into the multi-layer Transformer network to use the self-attention mechanism for deep feature extraction. The calculation formula is: Where Q represents the query matrix, which indicates the degree of attention of the current word to other words. A query vector is generated for each word position to find related words. K represents the key matrix, which indicates the characteristics or content of the word. A key vector is generated for each word position to match with the query vector (Q) to determine the relevance. V represents the value matrix, which indicates the value or output information of the word. A value vector is generated for each word position to indicate the content information of the word, which will be weighted and summarized into the output. k represents the dimension of the key vector K, which is a normalization factor used to narrow the value range of the dot product result to prevent excessively large values ​​from causing the softmax function to approach saturation and thus lose discrimination. T represents the transposed matrix of the key matrix K. The deep features include code value features, text features, and statistical features. Then, independent expert networks are established for text features, code value features, and statistical features. Each expert network is a multi-layer perceptron that generates representations for text features and reduces the dimensionality of statistical features. A multi-expert hybrid model is then designed to perform feature fusion and discrimination on the code value features, text feature representations, and statistical features after dimensionality reduction. Specifically, a gating mechanism is designed to analyze the input features to generate weighted coefficients for each expert network, and dynamically adjust the output weights of each expert network based on the importance of the input features. Finally, the outputs of all expert networks are fused through weighted summation to generate a comprehensive feature representation. The fusion formula is as follows: Where n is the number of expert networks, F final is the final feature representation after fusion; The key group clue discovery module receives the operation results of the multi-expert hybrid model, and screens key groups through text clue discovery based on large model prompts and sliding window time series clue discovery methods, and obtains group screening results.

2. The method for discovering group clues based on multi-feature fusion according to claim 1, characterized in that: The analysis method of the code value features is: focusing on code values ​​that appear repeatedly in multiple groups or have obvious directionality; The statistical feature analysis method is to use a multi-layer perceptron model to fuse all statistical features and generate the final statistical feature representation. The calculation formula is as follows: F stat =σ(W×X+b) Among them, X is the statistical feature vector of the input, W is the weight matrix, b is the bias, and σ is the activation function.

3. The method for discovering group clues based on multi-feature fusion according to claim 2, characterized in that: The text clue discovery based on large model prompts utilizes a pre-trained language model and designs appropriate prompts. When processing prompts, the model comprehensively considers the word frequency, semantic association, and emotional expression in the text to generate output containing key information, thereby mining text clues closely related to group behavior, including the group's charter, emotional tendencies, common topics, hot keywords, and communication style.

4. The method for discovering group clues based on multi-feature fusion according to claim 3, characterized in that: The sliding window temporal clue discovery method dynamically analyzes the temporal dimension of group activities on social platforms to identify specific behaviors. Specifically, the data is divided into overlapping time periods. Within each time period, the group's speech frequency, keyword usage, and user activity indicators are independently analyzed to capture sudden behaviors and hot spots within the group. The similarity of the activity patterns of two groups in different time periods is judged. The dynamic time rule method is introduced to measure the time series similarity of the two groups. By calculating the minimum distance between the activity patterns of the groups, the groups with similar activity characteristics in different time periods are identified. The calculation formula is: Where X[i] and Y[j] represent the values ​​of the two time series X and Y at the i-th and j-th time points respectively, and n is the length of the time series.

Images