A social media data mining method based on user behavior analysis

Abstract

The present application belongs to the technical field of data mining, and particularly relates to a social media data mining method based on user behavior analysis. Real-time collection of user multi-source behavior data, through natural language processing to analyze unstructured data, extract key word groups and map to structured fields; after data cleaning to eliminate abnormal data, behavior sequences are constructed within the time window, time-aware embedding vectors are introduced to generate time sequence feature representation, and BERT is used to extract behavior content semantic vector, and unified content representation vector is generated by fusing behavior type, intensity and other features; finally, the two types of features are fused through BiGRU-Attention network to predict user future behavior probability value to drive personalized recommendation. The method solves the problems of insufficient capture of behavior semantic motivation, difficulty in processing data heterogeneity and weak dynamic modeling capability in the prior art, and improves the social media data mining precision.

Description

Technical Field

[0001] This invention belongs to the field of data mining technology, and in particular relates to a social media data mining method based on user behavior analysis. Background Technology

[0002] With the widespread adoption of social media, users generate a wealth of behavioral data on these platforms, such as browsing, liking, commenting, and sharing. These behaviors reveal potential interests and social relationships, forming a crucial foundation for recommendation systems and sentiment analysis. Existing data mining methods largely rely on content keywords or historical tags for modeling, making it difficult to capture the semantic motivations behind these behaviors, leading to biases in interest comprehension. Furthermore, social behavior data suffers from strong heterogeneity and inconsistent modal distributions, making it difficult for traditional methods to unify modeling and integrate analysis. More importantly, user interests exhibit significant dynamic evolutionary characteristics, and existing methods, mostly static models, struggle to adapt to short-term behavioral fluctuations and relationship mutations, resulting in recommendation lags and analysis failures. Therefore, there is an urgent need for a user behavior analysis method that supports multimodal fusion and dynamic modeling to improve the accuracy of social media data mining. Summary of the Invention

[0003] To address the technical problems existing in the background art described above, this invention proposes a social media data mining method based on user behavior analysis.

[0004] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0005] S1. User Data Collection: Real-time acquisition of multi-source behavioral data generated by user behavior from social media platforms;

[0006] S2. Data Transformation: Parse unstructured data, extract keyword groups from multi-source behavioral data through natural language processing, and map the extracted keyword groups to preset structured fields;

[0007] S3. Data Preprocessing: Clean the data after data transformation to remove abnormal data;

[0008] S4. Constructing a temporal representation of behavior: Taking users as the unit, constructing behavior sequences in chronological order within a set time window, introducing time-aware embedding vectors to form a temporal feature representation of the dynamic changes in user behavior;

[0009] S5. Content Feature Extraction and Encoding: Taking users as units, extract perceptual features of behaviors based on content within a set time window, and convert the behavioral content features into a content representation vector with a unified dimension.

[0010] S6, user intention recognition: input the time sequence feature representation vector and the content representation vector into the behavior intention recognition model, predict the future behavior probability value of the user, and drive personalized recommendation.

[0011] Preferably, the multi-source behavior data in step S1 includes text comments, like records, forwarding trajectories, and browsing dwell time.

[0012] Preferably, the implementation of data conversion in step S2 is:

[0013] Context completion modeling is performed on the collected user behavior data, and a pre-trained language model is used to expand the context information of the text segment;

[0014] Based on the context attention mechanism, a candidate set of key word groups is extracted, and multiple feature factors such as word frequency, part of speech, dependency relationship, and context semantic density are fused to perform multi-dimensional scoring on the candidate word groups, and the key word groups with user intention expression characteristics are selected;

[0015] Performing syntactic analysis and dependency structure labeling on the candidate key word groups, and identifying the key word group boundary through graph structure traversal;

[0016] Generate confidence score for the semantic mapping relationship between the key word group and the structured field, and set a confidence threshold. If the threshold is exceeded, it means that the mapping structure is successful.

[0017] Preferably, the data cleaning of the data after the data conversion operation in step S3 is implemented as follows:

[0018] First, calculate the semantic cosine similarity between the behavior and the user's historical behavior through BERT to obtain the semantic consistency measure S c ;

[0019] According to the behavior occurrence time point, calculate whether the occurrence time interval falls within a reasonable fluctuation range to obtain the time sequence rationalization index T c , Where t i represents the timestamp of the current behavior, and w is a preset time fluctuation window;

[0020] Combine the semantic consistency measure and the time sequence rationalization index to obtain the behavior confidence score. If the behavior confidence score is less than a preset threshold, the behavior is determined to be an abnormal sample behavior and is deleted.

[0021] Preferably, the specific implementation of forming the time sequence feature representation of the user dynamic behavior change in step S4 is:

[0022] In the time window [t-w, t], arrange the behavior set B = {b1, b2...bn}, where each action b i Extract behavior type, behavior intensity, and behavior interval features; wherein, the behavior type xl i Converted to vector e via hot encoding xl Behavioral intensity xq i xq is obtained through normalization. i ′, behavioral interval Δt i =t i -t i-1 ;

[0023] Calculation behavior b i Relative position within the time window Where p i Represents relative position, t i For behavior b i The event stamp, t is the end point of the time window, i.e., the current time point; a time-aware vector is generated using sine and cosine functions. Where d is the embedded dimension index and D is the total dimension;

[0024] e xl 、xq i ′、Δt i With TPE(p i The concatenation forms a composite feature vector X. i And through a recurrent neural network, the sequence X = [X1, X2, ..., X...] is processed. n Model the hidden state sequence H and output it.

[0025] The hidden state sequence H is averaged and pooled to obtain the user temporal feature representation vector S. time .

[0026] Preferably, the specific implementation of the unified dimension content representation vector in step S5 is as follows:

[0027] Within the time window [tw,t], BERT is used to extract the text semantic vector for the behavior content of each behavior.

[0028] The behavior type vector e xl and behavioral intensity xq i ′, combined with text semantic vector Concatenate into multiple feature vectors

[0029] Interactive features are generated through linear transformation and activation functions. Among them W int Represented as a weight matrix, b int Represented as a bias vector;

[0030] The interaction features of each behavior are projected onto a unified dimension. Among them, W proj Represented as a projection matrix, b proj It is the bias vector;

[0031] Finally, all behaviors within the window... Aggregate to generate user-level content representation vectors:

[0032] Preferably, the intent recognition model in step S6 is implemented using a BiGRU-Attention network structure.

[0033] Preferably, the intention recognition model outputs the behavior probability value as follows:

[0034] First, the temporal feature representation vector and the content representation vector are respectively input into the dual-channel encoder structure to extract the corresponding deep semantic features;

[0035] The output features of the two channels are concatenated in a high-dimensional space and input into a fully connected fusion layer. A unified behavior state representation vector is generated through feature compression and reconstruction.

[0036] Next, the behavioral state representation vector is input into the intent discrimination module, which includes an attention gating mechanism. This module adaptively adjusts the response intensity to different behavioral factors based on the feature weights.

[0037] Finally, the Softmax activation function is used to output the probability values ​​of each action of the user within a predefined time window in the future.

[0038] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0039] 1. In the data transformation stage, this invention introduces the BERT language model to achieve context completion modeling, combines dependency structure graphs to mine keyword phrase boundaries, and designs a confidence scoring mechanism to automatically map unstructured data to structured fields. This solution is more robust and transferable than traditional manual rule extraction or static mapping based on TF-IDF, improving data adaptation efficiency.

[0040] 2. This paper proposes a method that concatenates user behavior information encoded in dual channels and inputs it into a fusion layer before feeding it into a BiGRU-Attention structure for behavior intent discrimination. Compared to a single RNN structure, this method possesses bidirectional memory capabilities and, combined with an attention mechanism, can dynamically adjust the perceptual weights for different behaviors, enhancing the model's responsiveness to key behavioral segments. This structure effectively improves the sensitivity and accuracy of user behavior prediction, solving the problem of traditional models' inability to respond promptly to short-term interests.

[0041] 3. A bidirectional modeling method is used to represent the behavior time sequence and the content semantic representation, decoupling the time trend of user behavior and the semantic cause behind the behavior, and fusing into a unified vector input intent recognition model. This structure is different from the traditional modeling method which only relies on content or behavior label, and can capture the dual association of behavior occurrence, and improve the modeling ability of short-term interest fluctuation and long-term behavior habit of users. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0043] Figure 1 The overall structure flowchart of a social media data mining method based on user behavior analysis. DETAILED DESCRIPTION

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0045] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the present application is not limited to the specific embodiments disclosed in the following description.

[0046] Embodiment, with the wide application of social media, users generate massive behavior data, covering browsing, liking, commenting, forwarding, etc. These data contain interest tendency and social relationship, which are important basis for recommendation system and public opinion analysis. However, the existing mining method mostly relies on content keywords or historical label modeling, which is difficult to grasp the behavior semantic cause, resulting in interest understanding deviation. At the same time, social behavior data is highly heterogeneous and modal distribution is not uniform, and traditional methods are difficult to model and analyze uniformly. Moreover, the dynamic evolution characteristics of user interest are significant, and the existing methods are mostly static modeling, which is difficult to adapt to short-term behavior fluctuation and relationship mutation, resulting in lagging recommendation and analysis failure. Therefore, the present application proposes a social media data mining method based on user behavior analysis, and the specific implementation steps are as shown in Figure 1

[0047] ​Firstly, in order to achieve comprehensive perception of user multi-source behavior, the method first acquires multi-source behavior data generated by the user on the platform in real time from a social media platform. The multi-source behavior data includes text comments, like records, forwarding trajectories, and browsing dwell time. In order to ensure the timeliness of the data, the behavior events are collected in real time, and the behavior is preliminarily cached and marked based on the event timestamp, providing time synchronization support for subsequent behavior sequence modeling.

[0048] Then, in order to convert the unstructured data in user behavior into processable structured semantic expression, a natural language processing technology based on a pre-trained language model such as BERT is used to extract key word groups from the original text and map them to structured fields. Specifically, first, the context completion modeling of the collected user text behavior data is performed. The pre-trained language model is used to predict the semantic context of the original text segment, fill in the implicit information in the text, and generate a complete semantic path. Through the context representation after completion, the real intention of the user in the limited number of words can be more accurately understood. Secondly, based on the context representation, a context attention mechanism is constructed. Taking each word in the sentence as the attention center, the attention weight of each word with other words in the context is calculated to capture the important semantic association within the sentence. Combined with the context attention score, a candidate set of key word groups is extracted. Then, each candidate word group is scored by multiple feature factors. Including: word frequency statistics, identifying high-frequency concepts; part-of-speech identification, preferentially retaining nouns and verb phrases; syntactic dependency relationship analysis, identifying core relationships such as subject-verb and verb-object; context semantic density calculation, measuring the semantic concentration of candidate words in the context through word embedding space. All features are combined to build a comprehensive scoring function according to the set weight, and the candidate key word groups are sorted and filtered. Subsequently, syntactic analysis and dependency structure labeling are performed on the filtered key word groups. Based on the syntactic dependency relationship graph, graph structure traversal is performed to identify the semantic boundaries of the word groups, preventing key words from being fragmented or missegmented. Finally, a semantic matching model between the key word groups and the preset structured fields is established. The cosine similarity between the word embedding representation and the structured field template is calculated to generate a matching confidence score; by setting a confidence threshold, if the score exceeds the threshold, it is determined that the key word group is successfully mapped to the structured field.

[0049] Then, in order to eliminate fake, abnormal and noisy behavior data and improve modeling accuracy, the data is cleaned to remove abnormal data in the following steps. Specifically, first, the semantic cosine similarity between the behavior and the user's historical behavior is calculated by BERT to obtain a semantic consistency measure S c, the [CLS] position output vector of each text is input into the encoding layer of the BERT model, and the output vector is taken as the global representation of the whole sentence semantics. Then, the cosine similarity between the current behavior semantic vector and the semantic vector of each behavior in the historical behavior is calculated, and the average or weighted average of the similarity is taken as the final semantic consistency measurement value; whether the time interval falls within a reasonable fluctuation range is calculated according to the behavior occurrence time point to obtain a time sequence rationalization index T c , where t i represents the time stamp of the current behavior, and w is a preset time fluctuation window; the behavior confidence score is obtained by weighted fusion of the semantic consistency measurement and the time sequence rationalization index; if the behavior confidence score is less than a preset threshold, the behavior is determined as an abnormal sample behavior, and the behavior is determined as an invalid behavior and is eliminated. The mechanism guarantees the semantic continuity and behavior rationality of the input sequence, and effectively improves the robustness and generalization ability of subsequent modeling.

[0050] In order to fully express the dynamic evolution characteristics of user behavior, the application constructs the behavior sequence in time sequence in the set time window for each user, introduces the time perception embedding vector, and forms the time sequence feature representation of the dynamic behavior change of the user. Specifically, in the time window [t-w, t], the behavior set B={b1, b2...b n} is arranged in ascending order according to the time stamp, wherein each behavior b i extracts the behavior type, the behavior intensity and the behavior interval feature; wherein the behavior type xl i is converted into a vector e xl through hot encoding, the behavior intensity xq i is obtained through normalization xq i ', and the behavior interval Δt i =t i -t i-1 ; the relative position of the behavior b i in the time window is calculated where p i represents the relative position, t i is the occurrence event stamp of the behavior b i , and t is the end point of the time window, that is, the current time point; the time perception vector TPE(p i ) is generated through the sine cosine function where d is the embedding dimension index, and D is the total dimension; e xl , xq i ', Δt i and TPE(p i ) are spliced into a composite feature vector X i , and the sequence X=[X1, X2,..., X nModeling, output the hidden state sequence H; finally, the mean pooling is performed on the hidden state sequence H to obtain a user timing feature representation vector S time The core role of this step is to model the behavior sequence of the user in time, and extract the evolution law of the user's behavior in a set time window. The user's behavior on the social media has significant time dependence, for example, high-frequency interaction in a short time may represent an interest outbreak, and intermittent browsing may represent a potential interest change. Therefore, the behavior events are constructed into a behavior sequence in chronological order, and a time-aware embedding vector is introduced, which can model the timing structure of the behavior and capture the rhythmicity, inertia and evolution trend of the user behavior, providing dynamic context support for subsequent user intention prediction and content recommendation.

[0051] Then, in order to extract the semantic features directly related to the user's interest from the behavior content, the application extracts the content features of the behavior in a set time window for each user, and converts the behavior content features into a content representation vector in a unified dimension. Specifically, in the time window [t-w, t], the text semantic vector of the behavior content of each behavior is extracted using BERT The behavior type vector e xl and the behavior intensity xq i ′, combined with the text semantic vector are spliced into a multi-feature vector The interaction feature is generated through linear transformation and activation function where W int is a weight matrix, and b int is a bias vector; the interaction feature of each behavior is projected into a unified dimension where W proj is a projection matrix, and b proj is a bias vector; finally, the of all behaviors in the window is aggregated to generate a user-level content representation vector: The core role of this step is to understand and model the content involved in the user's behavior at the semantic level. The behavior object of the user on the social platform is usually high-dimensional content, and these contents themselves carry the theme direction and emotional tendency of the user's interest. Therefore, the semantic features of the behavior content are extracted and encoded into a content representation vector in a unified dimension, which is a key link to build a user interest portrait and understand the user behavior motivation. This step converts different types of behavior content into structured and numerical expressions through a modal perception encoding mechanism, which serves as a key input for subsequent behavior fusion and intention recognition.

[0052] Finally, in order to accurately identify the user's potential behavior intention and drive real-time recommendation optimization, the time sequence feature representation vector and the content representation vector are input into a behavior intention identification model in the last step to predict the user's future behavior probability value. The intention identification model is realized by using a BiGRU-Attention network structure. First, the time sequence feature representation vector and the content representation vector are respectively input into a double-channel encoder structure to extract corresponding deep semantic features; then the output features of the two channels are spliced in a high-dimensional space and input into a fully connected fusion layer to generate a unified behavior state representation vector through feature compression and reconstruction; then the behavior state representation vector is input into an intention discrimination module containing an attention gate mechanism, which adaptively adjusts the response strength of different behavior factors according to the feature weight; finally, the probability value of each behavior of the user in the future predefined time window is output through the Softmax activation function. Specifically, the obtained time sequence feature representation vector and content representation vector are first input into two parallel encoder channels. In each channel, a bidirectional gated recurrent unit (BiGRU) is used for feature extraction. The BiGRU structure models the sequence information in both forward and backward directions, which can effectively capture the long-term dependence and context semantics of user behavior in the time dimension, enhancing the expression ability of the behavior sequence; at the same time, the content representation vector is modeled, which helps to understand the semantic reasons and preference patterns behind the behavior. Subsequently, the feature representations output by the two channels are spliced in a high-dimensional space to construct a joint feature vector containing time evolution information and content semantics. In order to control the feature dimension and prevent redundant interference, the spliced vector is input into a fully connected fusion layer to generate a unified dimension behavior state representation vector through feature compression and reconstruction operations. The fusion layer completes the structural unification of multi-modal information on the basis of preserving the main behavior semantics, so that the downstream discrimination module can more stably receive the fused feature representation. Next, the behavior state representation vector is sent to an intention discrimination module containing an attention gate mechanism. The module automatically identifies the key behavior feature dimensions in the current state through a self-attention structure and assigns different response weights. The attention gate mechanism dynamically adjusts the influence degree according to the intensity, suddenness and time position of the user's recent behavior, improves the model's perception ability of short-term interest changes and key content changes, and significantly improves the sensitivity and accuracy of user intention prediction. Finally, the model maps the high-dimensional vector to a normalized probability value based on the feature representation output by the intention discrimination module, representing the probability value of each type of behavior that the user may take in the future predefined time window. The invention can interface the prediction output result with the platform recommendation engine to realize real-time distribution of personalized content. The system dynamically adjusts the recommendation list according to the user's current behavior state, potential interest intention and platform recommendation strategy, improves the click-through rate and platform stickiness.

[0053] The above merely describes preferred embodiments of the present application, but is not intended to limit the present application to other forms, any person skilled in the art can make changes or modifications to the above disclosed technical contents into equivalent embodiments with equivalent changes, and apply to other fields, but any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application without departing from the technical solution content of the present application still belongs to the protection scope of the present application technical solution.

Claims

1. A social media data mining method based on user behavior analysis, characterized in that, The method comprises the following steps: S1, user data collection: acquiring multi-source behavior data generated by user behavior from a social media platform in real time; S2, data conversion: parsing unstructured data, extracting key word groups from multi-source behavior data through natural language processing, and mapping the extracted key word groups to preset structured fields; S3, data preprocessing: data cleaning of data after data conversion operation to remove abnormal data; S4, constructing behavior time sequence representation: taking a user as a unit, constructing behaviors in a time window into a behavior sequence in time sequence, introducing a time-aware embedding vector, and forming a time sequence feature representation of user dynamic behavior change; S5, content feature extraction and coding: taking a user as a unit, extracting behavior content features in a set time window, and converting behavior content features into a unified dimension content representation vector; S6, user intent recognition: inputting the time sequence feature representation vector and the content representation vector into a behavior intent recognition model to predict the future behavior probability value of the user and drive personalized recommendation; The specific implementation of the time sequence feature representation of the user dynamic behavior change formed in step S4 is: In a time window , arrange the behavior set in ascending order of timestamp , wherein each behavior extract behavior type, behavior intensity and behavior interval features; wherein the behavior type converted into a vector by thermal encoding , behavior intensity obtained by normalization , behavior interval ; Computing behavior Relative position within a time window wherein representing a relative position, is a behavior of an occurrence event stamp, t is a time window end point, i.e., a current time point; a time-aware vector is generated by a sine-cosine function wherein d is an embedding dimension index, and D is a total dimension; Concatenate , , with to a composite feature vector , and model the sequence through a recurrent neural network, outputting a sequence of hidden states H; performing mean-pooling on the hidden state sequence H to obtain a user time sequence feature representation vector ; The specific implementation of the content representation vector of the unified dimension in step S5 is: In the time window BERT is used to extract text semantic vectors for the behavior content of each behavior ; concatenate the behavior type vector and the behavior intensity , and the text semantic vector to form a multi-feature vector ; Generating interaction features by linear variation and activation function wherein is represented as a weight matrix, is represented as a bias vector; projecting interaction features of each behavior onto a uniform dimension wherein, is represented as a projection matrix, is a bias vector; Finally, all the behaviors within the window are aggregated to generate a user-level content representation vector: .​ 2. The social media data mining method based on user behavior analysis according to claim 1, characterized in that, The multi-source behavior data in step S1 includes text comments, like records, forwarding trajectories, and browsing dwell time. 3.The social media data mining method based on user behavior analysis of claim 1, wherein, The implementation of data conversion in step S2 is: Context completion modeling of collected user behavior data, using a pre-trained language model to expand the context information of the text fragment; Based on the context attention mechanism, a candidate set of key word groups is extracted, and multi-feature factors such as word frequency, part of speech, dependency relationship, and context semantic density are fused to score the candidate word groups in multiple dimensions, and the key word groups with user intent expression characteristics are selected; Performing syntactic analysis and dependency structure labeling on the candidate key word groups, and identifying the boundaries of the key word groups through graph structure traversal; Generating a matching confidence score between the semantic mapping relationship of the key word group and the structured field, and setting a confidence threshold, if the threshold is exceeded, it means that the mapping structure is successful.

4. The social media data mining method based on user behavior analysis according to claim 1, characterized in that, The specific implementation of step S3 for data cleaning of data after data conversion operation to remove abnormal data is: First, the semantic cosine similarity between the behavior and the user's historical behaviors is calculated by BERT to obtain a semantic consistency measure ; According to a time point of the behavior, whether the time interval falls within a reasonable fluctuation range is calculated to obtain a time series rationalization index , , wherein represents a time stamp of the current behavior, and w is a preset time fluctuation window. Combine semantic consistency measurement and time series rationalization index to get behavior confidence score, if the behavior confidence score is less than the preset threshold, it is determined that the behavior is an abnormal sample behavior and is deleted.

5. The social media data mining method based on user behavior analysis as claimed in claim 1, wherein, The intent recognition model in step S6 is implemented using a BiGRU-Attention network structure.

6. The social media data mining method based on user behavior analysis according to claim 5, characterized in that, The specific implementation of the intent recognition model for outputting behavior probability value is: First, input the time sequence feature representation vector and the content representation vector into a double-channel encoder structure respectively to extract the corresponding deep semantic features; The output features of the two channels are spliced in a high-dimensional space and input into a fully connected fusion layer to generate a unified behavior state representation vector through feature compression and reconstruction; Then, input the behavior state representation vector into an intent discrimination module containing an attention gate mechanism, which adaptively adjusts the response strength of different behavior factors according to the feature weight; Finally, the probability values of each behavior in the future pre-defined time window are outputted by the Softmax activation function.

Images