Abnormal data prediction method for network illegal competition behavior

Abstract

The invention provides an abnormal data prediction method for network illegal competition behaviors, and relates to the technical field of data processing, and the method comprises the steps: obtaining text data of a target complaint event and a historical complaint behavior sequence of a corresponding user; semantic coding is conducted on the text, and text nodes and legal reference intention nodes are extracted; constructing behavior event nodes based on historical complaint events, wherein the behavior event nodes comprise complaint intensity spectrum features obtained through frequency domain transformation; the text nodes, the intention nodes and the behavior event nodes are constructed into a heterogeneous graph, multi-layer feature aggregation and pooling are carried out through a graph neural network, and a whole graph representation vector is obtained; and finally, predicting whether the target complaint event belongs to abnormal data or not based on the vector. According to the method, texts, behavior spectrums and law reference intentions are fused, hidden and periodic collaborative complaint behaviors can be effectively recognized, and recognition accuracy and interpretability are improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method for predicting abnormal data related to unfair competition on the internet. Background Technology

[0002] In consumer rights protection and market supervision scenarios, complaint and reporting data are crucial clues for discovering illegal and irregular activities. However, the number of professional claims for profit is increasing. In processing complaint and reporting data, the text content exhibits both templated and variably worded characteristics. The same intent to claim compensation is often rewritten through synonym substitution, sentence transformation, or implicit expression, making it difficult to reliably identify claims using methods based on fixed rules or keywords. Simultaneously, legal clause citations are flexible and diverse, with explicit and implicit citations intertwined, limiting the ability of traditional text matching techniques to capture semantic intent. Furthermore, professional claims constitute a relatively low proportion of overall complaint data, leading to severe class imbalance. This makes the model prone to bias towards majority class samples during training, hindering the effective learning of minority class feature patterns and impacting the stability and generalization ability of the discrimination performance. Summary of the Invention

[0003] The purpose of this application is to provide an abnormal data prediction method for unfair competition on the Internet, so as to alleviate the above-mentioned technical problems existing in the prior art.

[0004] In a first aspect, the present invention provides a method for predicting abnormal data related to unfair competition on the internet, comprising: Obtain the text data of the target complaint event and the behavioral sequence composed of multiple historical complaint events of the corresponding user; Semantic encoding is performed on text data, semantic units carrying significant attention weights are extracted as text nodes, and legal citation fragments are identified from text data to generate intent nodes representing citation intent. Behavioral event nodes are constructed based on historical complaint events. Each event node includes event attributes and spectral features obtained by converting timestamp sequences. The spectral features are obtained by performing frequency domain transformation on the complaint intensity signal and extracting frequency distribution characteristics. Construct a heterogeneous graph containing text nodes, intent nodes, and behavior event nodes; The heterogeneous graph is subjected to multi-layer feature aggregation through a target graph neural network, and the full graph representation vector is obtained through graph pooling. The multi-layer feature aggregation includes local information fusion between nodes of the same type and global information interaction between nodes of different types. Predict whether a target complaint event is an anomaly related to unfair competition in the network based on the full graph representation vector.

[0005] Secondly, the present invention provides an abnormal data prediction device for unfair competition on the Internet, including an acquisition module for acquiring text data of a target complaint event and a behavioral sequence composed of multiple historical complaint events of the corresponding user. The text node extraction module is used to perform semantic encoding on text data and extract semantic units carrying significant attention weights as text nodes. The intent node generation module is used to identify legal citation fragments from text data to generate intent nodes that represent the citation intent; The behavior node construction module is used to construct behavior event nodes based on historical complaint events. The event node includes event attributes and spectral features obtained by converting the timestamp sequence. The spectral features are obtained by performing frequency domain transformation on the complaint intensity signal and extracting the frequency distribution characteristics. The graph structure building module is used to construct heterogeneous graphs containing text nodes, intent nodes, and behavior event nodes; The graph neural network processing module is used to perform multi-layer feature aggregation on heterogeneous graphs through the target graph neural network and obtain the full graph representation vector through graph pooling. The multi-layer feature aggregation includes local information fusion between nodes of the same type and global information interaction between nodes of different types. The prediction module is used to predict whether a target complaint event is an anomaly related to unfair competition in the network, based on the full graph representation vector.

[0006] Thirdly, the present invention provides an electronic device including a processor and a memory, the memory storing computer-executable instructions that can be executed by the processor, the processor executing the computer-executable instructions to implement the abnormal data prediction method for unfair competition behavior on the network according to any of the foregoing embodiments.

[0007] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the abnormal data prediction method for unfair competition behavior on the network according to any of the foregoing embodiments.

[0008] This application provides an abnormal data prediction method for unfair competition behavior on the Internet. By extracting semantic units with significant attention weights to construct text nodes, it can effectively capture key expressions in complaint texts, thereby overcoming the recognition difficulties caused by text templates or expression variations. By identifying legal citation fragments and generating corresponding intent nodes, it enhances the ability to distinguish between explicit and implicit legal citation intent. By performing frequency domain transformation on complaint behavior sequences and extracting spectral features to construct behavior event nodes, it accurately portrays dynamic behavior patterns such as periodicity and clustering, making up for the limitations of traditional statistical methods in such scenarios. It uses a heterogeneous graph structure to achieve structured association modeling of three types of information: text, intent, and behavior, and combines a hierarchical graph neural network to fuse local and global multi-level information, effectively improving the recognition efficiency of cross-modal collaborative features. In addition, by using a representation learning method based on the whole graph, it enhances the feature expression ability of sparse samples, thereby maintaining high recognition accuracy and model robustness even under conditions of uneven data distribution. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0010] Figure 1 A flowchart illustrating an abnormal data prediction method for unfair competition on the internet, provided as an embodiment of this application; Figure 2 A flowchart illustrating a specific method for predicting abnormal data related to unfair competition on the internet, provided as an embodiment of this application; Figure 3 A structural diagram of an abnormal data prediction device for unfair competition on the Internet provided in an embodiment of this application; Figure 4 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0012] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0013] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0014] Based on this, embodiments of this application provide an anomaly data prediction method for unfair competition behavior on the internet, addressing the problems of insufficient accuracy, poor robustness, and weak model interpretability in existing technologies. See also Figure 1 As shown, the method mainly includes the following steps: S110, Obtain the text data of the target complaint event and the behavioral sequence consisting of multiple historical complaint events of the corresponding user.

[0015] The text data of a target complaint event refers to the text content submitted in the complaint to be identified. This text data may include content such as the consumer's statement of facts, expression of demands, and citation of legal basis. The behavioral sequence, composed of multiple historical complaint events of a corresponding user, is a time-ordered set of several complaint records initiated by the same complainant on the same or different platforms in the past, arranged chronologically. This behavioral sequence is used to characterize the user's behavioral trajectory and pattern characteristics.

[0016] S120, semantically encode the text data, extract semantic units carrying significant attention weights as text nodes, and identify legal citation fragments from the text data to generate intent nodes representing the citation intent.

[0017] When semantically encoding text data, a pre-trained language model can be used to transform the text into a distributed vector representation to obtain the corresponding semantic encoding. The extracted semantic units carrying significant attention weights refer to words or sub-words that are given high attention during model inference. These semantic units are used to characterize their key role in judging occupational claim tendency.

[0018] In this embodiment, the legal citation fragment refers to the portion of the text that explicitly or implicitly cites legal or regulatory provisions. The final generated intent node is an abstract representation of the underlying motivation for citing the legal citation fragment, used to distinguish between three types of intents: legitimate rights protection, templated claims, and circumvention-oriented citations.

[0019] S130, construct behavioral event nodes based on historical complaint events. The event nodes include event attributes and spectral features obtained by converting timestamp sequences. The spectral features are obtained by performing frequency domain transformation on the complaint intensity signal and extracting frequency distribution characteristics.

[0020] Behavioral event nodes are high-dimensional feature representations of each historical complaint event. Event attributes include structured fields such as complaint time, claim amount, and product type. The timestamp sequence is a time sequence of complaint occurrences arranged along a timeline. The complaint intensity signal is a numerical sequence that quantifies the impact of each complaint, typically calculated by weighting the claim amount and complaint frequency.

[0021] S140, construct a heterogeneous graph containing text nodes, intent nodes, and behavior event nodes.

[0022] The aforementioned heterogeneous graph is a graph structure containing multiple types of nodes and edges, enabling more refined modeling of complex relationships across modalities. Text nodes represent deep semantic information, intent nodes represent the psychological motivations behind legal citations, and behavioral event nodes record external action trajectories. Together, these three types of nodes constitute a joint representation space that integrates semantics, behavior, and rule logic.

[0023] S150 performs multi-layer feature aggregation on heterogeneous graphs through a target graph neural network and obtains the full graph representation vector through graph pooling; wherein, multi-layer feature aggregation includes local information fusion between nodes of the same type and global information interaction between nodes of different types.

[0024] The target graph neural network is a pre-trained graph attention network (GAT) used to achieve information transfer and feature updates between nodes. Preferably, it can be trained using a distillation framework consisting of a teacher model and a student model. Multi-layer feature aggregation, or information fusion in stages, first performs local aggregation within nodes of the same type, and then performs global interaction across modal boundaries. Graph pooling is the operation of summing all node features into a single fixed-length vector to represent the overall state of the entire graph.

[0025] In practice, a two-layer graph attention network is used to process heterogeneous graphs. The first layer focuses on local information fusion within a modality, while the second layer realizes global information interaction across modalities. A modal attention gating mechanism is introduced to dynamically adjust the contribution weights of different modal paths, thereby enhancing the model's robustness.

[0026] S160, based on the full graph representation vector, predict whether the target complaint event belongs to abnormal data related to unfair competition on the Internet.

[0027] The abnormal data related to unfair competition in the network referred to in this embodiment refers to behavioral data corresponding to professional claims behaviors such as profit-making, abuse of complaint mechanisms, and disruption of market order.

[0028] In practice, the full-image representation vector is input into a lightweight classifier (such as a single-layer feedforward neural network). The classifier predicts the likelihood that the target complaint event belongs to a professional claimant. The classification results can be directly used in automated early warning systems or to assist law enforcement personnel in conducting key investigations.

[0029] In summary, the abnormal data prediction method for unfair competition in the network provided in this application effectively integrates textual semantic information, behavioral sequence features, and legal citation intent through the constructed cross-modal alignment graph structure, achieving significant technical results in the identification task of specific complaint and reporting behaviors. Compared with traditional methods, this scheme can maintain a high comprehensive performance index (F1 score) even under conditions of extremely imbalanced sample distribution and achieves a significant improvement in recall. Especially when dealing with complex scenarios with periodic and collaborative characteristics, this invention demonstrates excellent recognition capabilities.

[0030] For ease of understanding, the abnormal data prediction method for unfair competition on the Internet provided in the embodiments of this application will be described in detail below.

[0031] The aforementioned acquisition of text data of the target complaint event and a behavioral sequence consisting of multiple historical complaint events of the corresponding user may, in specific implementation, include the following steps 1.1 to 1.3: Step 1.1: Obtain the text data of the target complaint event, clean the text data, remove invalid symbols and format tags, and convert full-width characters to half-width characters.

[0032] The text data of the target complaint includes information such as factual statements, expressions of demands, and legal citations. In one implementation, the raw text data of the target complaint is first obtained from the complaint and reporting platform of the State Administration for Market Regulation. Then, a text cleaning tool (such as Python's BeautifulSoup library) is used to remove embedded HTML tags and other formatting marks to prevent them from interfering with subsequent natural language processing. Next, the `unicodedata.normalize` function is called to convert all full-width letters, numbers, and punctuation marks to their corresponding half-width forms to ensure consistent character encoding. Finally, regular expressions are used to filter out all illegal symbols except for Chinese characters, English letters, Arabic numerals, and common Chinese punctuation, such as emoji codes and invisible control characters, completing basic data purification and obtaining clean, standardized text input for subsequent word segmentation and semantic modeling.

[0033] Step 1.2: Perform word segmentation on the cleaned text data and normalize and replace the keywords based on the thesaurus.

[0034] The keywords in this step specifically refer to core terms with discriminative power in identifying professional claims, such as "refund one, compensate three," "inferior test report," and "suspected false advertising." During word segmentation, mature Chinese word segmentation tools (such as Jieba or LTP) are used to segment the cleaned text, generating a token sequence composed of words. Subsequently, a pre-built domain thesaurus is loaded, and the segmentation results are iterated through, uniformly replacing words belonging to synonym groups with their standard word forms. For example, "tenfold compensation," "claim tenfold," and "demand 10 times refund" are all normalized to "refund one, compensate tenfold"; "Article 148 of the Food Safety Law," "Food Safety Law 148," and "Article 148's basis for claims" are uniformly mapped to "Article 148 of the Food Safety Law." This process effectively alleviates the problem of professional claimants circumventing keyword detection by fine-tuning their wording, improving the consistency and generalization ability of text representation.

[0035] Step 1.3: Arrange the historical complaint events of the user matched with the target complaint time in the order of timestamps to generate an ordered sequence of behavioral events.

[0036] The target complaint time refers to the time when the current complaint event occurred. The matched user is the complainant identified by identity markers (such as mobile phone number, ID card number, device ID, or registered account). Historical complaint events are other complaint records initiated by the user in the past period.

[0037] In practice, the system uses the user identifier in the target complaint event as an index to retrieve all historical complaint records from the past two years from the backend database. Each record contains twelve basic attributes: complaint timestamp, claim amount, product category, unified social credit code of the complainant, complaint channel, processing result, complainant's geographic region code, fingerprint / IP address encryption string of the complaining device, tag indicating the type of violation involved, total number of complaints in the user's history, historical complaint success rate / case closure rate, and time interval between the previous complaint and the previous complaint. After filtering out valid complaint entries, they are sorted in ascending order based on their timestamp field to form a complete time series. For users with no historical records, an empty series is generated, but the system can still make inferences and judgments based on the current text and other modal information.

[0038] Furthermore, the above-mentioned semantic encoding of text data, extraction of semantic units carrying significant attention weights as text nodes, and identification of legal citation fragments from the text data to generate intent nodes representing the citation intent can, in specific implementation, include the following steps 2.1 to 2.5: Step 2.1: Input the preprocessed text data into the pre-trained language model for encoding to obtain the hidden state vector of each semantic unit.

[0039] Pre-trained language models refer to language representation models that have been pre-trained on large-scale general corpora, such as BERT and RoBERTa. They have deep bidirectional context modeling capabilities and can generate high-dimensional vector-based context-related representations for each basic semantic unit (e.g., word or subword) in the input text. These representations are called hidden state vectors and contain the semantic information of the semantic unit in the current context.

[0040] In this step, the original text data, such as legal texts or judicial documents, is first subjected to necessary preprocessing operations, including but not limited to word segmentation, noise character removal, standardization of punctuation marks, and format unification, to form a structured text sequence. Then, the sequence is input into a pre-trained language model, which uses its multi-layer neural network structure to perform forward propagation calculation on the entire input sequence, and finally outputs the context-aware hidden state vector corresponding to each semantic unit.

[0041] Step 2.2: Based on the attention mechanism associated with the classification symbol in the last layer of the pre-trained language model, determine the attention weight corresponding to each semantic unit.

[0042] The classification symbol (usually labeled [CLS]) is a special symbol added to the beginning of the input sequence by the pre-trained language model. Its corresponding hidden state is often used to aggregate the global semantic information of the entire input sequence. The attention mechanism refers to the quantification of the degree of mutual attention between semantic units in the self-attention structure of the model. The attention weights associated with the [CLS] symbol reflect the importance of each semantic unit to the overall sentence meaning.

[0043] Step 2.3: Select target semantic units with attention weights higher than the set threshold, and use the hidden state vectors corresponding to the target semantic units as text nodes.

[0044] A threshold is set to filter semantic units with significant semantic importance; a target semantic unit refers to a semantic unit that exceeds the threshold in the attention weight evaluation, and is used to represent that it carries more prominent information in the context; a text node is a basic element in a graph structure represented by the hidden state vector of the target semantic unit, and is used to represent the vector expression of key semantic content in the document.

[0045] Based on the attention weights of each semantic unit obtained in step 2.2, they are compared with a predefined threshold, and semantic units with weight values ​​greater than the threshold are selected as key components that have a significant impact on the overall semantic expression. For each selected target semantic unit, its original hidden state vector in the output of the pre-trained language model is retained and directly used as an independent text node in the graph structure.

[0046] Step 2.4: Identify legal clause reference fragments contained in the text data using a sequence labeling model.

[0047] Sequence labeling models are task models for text sequences that can assign predefined category labels to each input semantic unit. Common architectures include BiLSTM-CRF and BERT-BiLSTM-CRF, and they are suitable for tasks such as named entity recognition and phrase boundary detection. Legal clause citation fragments refer to the parts of the text that explicitly point to specific legal provisions, such as formal expressions like "according to Article XX of the XX Law". They have fixed patterns but also have diverse variations.

[0048] This step uses a pre-trained sequence labeling model to label the original text data unit by unit, identifying continuous semantic unit intervals belonging to the "legal clause citation" category. By learning language patterns and contextual features from a large number of labeled samples, the model can accurately capture common keyword combinations in legal citations (such as "based on", "in accordance with", "Article X", etc.), regulation name matching rules, and syntactic structure features, thereby accurately locating and segmenting complete legal clause citation fragments.

[0049] Step 2.5: Input the reference fragment into the intent classification model, output the reference intent category, and generate the intent node corresponding to the intent category.

[0050] Intent classification models are text classification models used to determine the purposeful function or behavioral motivation embodied in a text. Common implementation methods include methods based on CNN, RNN, or fine-tuning of pre-trained models. Citation intent category refers to the functional purpose that the cited clause in legal text intends to achieve, such as supporting arguments, limiting the scope of liability, citing precedents, etc. Intent nodes are graph structure nodes built with citation intent category as the core attribute, used to represent the logical function behind a certain citation behavior.

[0051] Each legal clause reference identified in step 2.4 is extracted separately and fed into a pre-trained intent classification model for reasoning. The model determines the category of reference intent based on the semantic content of the reference and its contextual clues, such as supporting a claim, avoiding risks, establishing obligations, etc. For each classification result, an intent node is created with the intent category as its identifier, and a relationship is established between it and the corresponding reference.

[0052] Furthermore, the aforementioned construction of behavioral event nodes based on historical complaint events includes event attributes and spectral features obtained from timestamp sequence transformation. These spectral features are obtained by performing frequency domain transformation on the complaint intensity signal and extracting its frequency distribution characteristics. In specific implementation, this may include steps 3.1 to 3.6: Step 3.1: Extract the timestamp, claim amount, and product category information from each historical complaint.

[0053] Historical complaints are consumer complaints about goods or services that were initiated and recorded in the past. They are usually stored in a structured database. The timestamp corresponds to the specific date and time of the complaint to determine the chronological order of the complaints. The amount claimed can reflect the economic severity of the complaint. The product category is an identifier of the category to which the product belongs, such as "home appliances", "food", "clothing", etc., which is mainly used to characterize the product area involved in the complaint.

[0054] Step 3.2: Sort the historical complaint events according to the timestamp order to obtain an ordered sequence of behavioral events.

[0055] A behavioral event sequence is a linear sequence formed by arranging multiple historical complaint events in chronological order of their occurrence. It is used to represent the continuous behavioral trajectory of an individual or group within a specific time period. After obtaining all relevant complaint events and their timestamps, all events are sorted in ascending order using the timestamp as the key, generating a behavioral event flow that strictly follows the timeline of the organization.

[0056] Step 3.3: Generate a time point sequence sampled in fixed time units based on the behavioral event sequence. The complaint intensity corresponding to each time point is determined by weighting the claim amount and the number of complaints of historical complaint events falling within that time period.

[0057] Fixed time units refer to preset time granularities, such as daily, weekly, or monthly, used to divide continuous time axes; time point sequences are discrete time sets formed by dividing the entire observation period into several equal-length time periods; complaint intensity is a quantitative indicator used to characterize the level of complaint activity within a certain period, which can be calculated by weighting the total amount of claims with the number of complaints.

[0058] The ordered sequence of behavioral events obtained in step 3.2 is mapped onto an equally spaced time grid, that is, the entire time span is divided into a series of continuous and non-overlapping fixed-length time periods (e.g., in "weeks"). For each time period, the number of all complaint events falling within it is counted, and the corresponding claim amounts are accumulated. Then, a preset weighting method (e.g., linear weighting or exponential weighting) is used to fuse the number of complaints and the accumulated claim amounts to generate the complaint intensity value for that time period, thus forming a discrete time series signal, namely the complaint intensity signal, which is used to characterize the trend of complaint behavior over time.

[0059] Step 3.4: Perform frequency domain transformation on the complaint intensity signal corresponding to the time point sequence to obtain the frequency component distribution.

[0060] Frequency domain transformation is the process of converting discrete signals in the time domain to the frequency domain through mathematical transformations (such as Fourier transform). The frequency component distribution shows the distribution of signal energy at different frequencies and can identify whether there are significant periodic patterns (such as peaks that repeat weekly or monthly).

[0061] The complaint intensity signal generated in step 3.3 is processed by Fast Fourier Transform (FFT) to convert it from a time-domain representation to a frequency-domain representation, resulting in a set of complex frequency-domain coefficients. These coefficients are then normalized after taking their modulus to obtain the energy value corresponding to each frequency point, generating a frequency component distribution map. This frequency component distribution is used to characterize the hidden periodic structure in the original complaint signal, such as high-frequency noise, mid-frequency oscillations, or low-frequency long-term trends.

[0062] Step 3.5: Extract the main frequency, spectral concentration, energy proportion of periodic peaks, and periodic stability index from the frequency component distribution to form spectral characteristics.

[0063] The dominant frequency refers to the frequency component with the highest energy in the frequency component distribution, representing the most important periodic rhythm of the complaint behavior; the degree of spectral concentration is used to measure whether the signal energy is concentrated on a few frequencies, reflecting the clarity of the behavior pattern; the energy proportion of periodic peaks refers to the proportion of energy carried by all frequency points with significant periodicity to the total energy, characterizing the strength of the periodicity of the behavior; the periodic stability index is used to assess the persistence and consistency of the main period over a long period of time.

[0064] After obtaining the frequency component distribution, the frequency corresponding to the maximum energy value is identified as the main frequency, which is used to determine the main cycle of the complaint behavior (e.g., once every 7 days). Then, the spectral entropy or similar index is calculated to characterize the degree of spectral concentration. The smaller the value, the more concentrated the energy and the more regular the behavior. Next, the proportion of the sum of the energy of all periodic peaks exceeding a set threshold to the total energy of the entire frequency band is counted as a measure of the periodic intensity. Finally, the degree of shift of the main frequency in different sub-time periods is analyzed through a sliding window, and its variance or coefficient of variation is calculated to generate a periodic stability score. The above four indicators are combined to form a set of multi-dimensional spectral feature vectors to fully express the temporal dynamic characteristics of the complaint behavior.

[0065] Step 3.6: Concatenate the original event attributes and spectral features corresponding to each historical complaint event to generate the corresponding behavioral event node.

[0066] The original event attributes include basic information such as timestamp, claim amount, and product category extracted in step 3.1; the spectral features have been defined above as a feature set composed of main frequency, spectral concentration, energy proportion of periodic peaks, and periodic stability index; the behavioral event node is a high-dimensional vector representation that fuses the above two types of features, which not only retains the basic attributes of individual events, but also incorporates the macroscopic temporal pattern at the sequence level, making it suitable for node modeling tasks in graph neural networks or other complex models.

[0067] For each historical complaint event, after the global spectral feature extraction is completed, the spectral feature vector of its group or individual is concatenated with its original event attributes to generate a unified feature representation. For example, if the original attributes are encoded as low-dimensional dense vectors and the spectral features are four-dimensional numerical vectors, they are finally concatenated into a higher-dimensional joint feature vector.

[0068] The above methods effectively enhance the ability to deeply understand user complaint behavior, not only capturing the explicit characteristics of individual events, but also revealing the periodic and stable patterns hidden in the time series. This helps to identify abnormal behavior patterns, predict risk trends, and provide reliable data support for building intelligent customer behavior analysis systems, credit assessment models, and regulatory early warning mechanisms. It has significant technological advancements and practical application value.

[0069] Furthermore, the above-described construction of a heterogeneous graph containing text nodes, intent nodes, and behavior event nodes may, in specific implementation, include the following steps 4.1 to 4.5: Step 4.1: Establish semantic association edges between text nodes. The connection relationship of semantic association edges is determined based on the contextual dependency relationship between semantic units.

[0070] Text nodes are vector representations of key semantic units extracted from complaint text. They can be sets of semantic vectors composed of several tokens with the highest attention weights (e.g., Top-10) after encoding by a pre-trained language model. Text nodes centrally embody the core content of the text, such as claims, legal citations, etc. Semantic association edges describe the semantic coherence and contextual dependencies between different text segments. Their connection structure represents the grammatical and semantic dependency paths of words or phrases in sentences or paragraphs. The weights of semantic association edges are determined by the output of the attention mechanism within the pre-trained language model, especially the self-attention weight matrix taken from higher levels of the model (e.g., layer 11). This weight can capture long-distance semantic dependencies and effectively reflect the strength of logical connections between keywords.

[0071] In one example, directed edges are created between each of the Top-10 text nodes based on their position order in the original text and their attention scores. Only connections with attention weights higher than a set threshold are retained to ensure the sparsity and semantic focus of the graph structure, thereby generating a subgraph structure with local semantic connectivity.

[0072] Step 4.2: Establish temporal evolution edges between behavioral event nodes in chronological order. The weight of the temporal evolution edges decreases as the time interval between the two events increases.

[0073] Behavioral event nodes correspond to each complaint event in a user's historical complaint record. Each node's feature vector is composed of 12 basic attributes (such as complaint timestamp, claim amount, product category, complaint target's unified social credit code, complaint channel, processing result, complainant's geographic region code, complaint device fingerprint / IP address encryption string, relevant violation type tag, total number of historical complaints, historical complaint success rate / case closure rate, and time interval with the previous complaint) and 10 behavioral spectrum features (including main frequency, spectrum entropy, energy proportion of the top 5 peak frequencies, spectrum centroid, main frequency stability index, frequency domain signal standard deviation, low-frequency component energy ratio, high-frequency noise energy ratio, spectrum flatness, and fundamental frequency power), totaling 22 dimensions, used to characterize the behavioral pattern of a single complaint. Multiple behavioral events arranged in ascending order of timestamps constitute a behavioral sequence.

[0074] Temporal evolution edges are used to model the dynamic evolution of complaint behavior over time, representing trends in user complaint frequency, such as periodic concentration or sudden outbreaks. These edges are directed, pointing from earlier events to later events. The weights are non-linearly decayed using a Gaussian function to account for the time difference between events; shorter time intervals result in higher weights, indicating stronger behavioral continuity, while longer intervals lead to rapidly decreasing weights. Establishing temporal evolution edges effectively identifies the temporal clustering characteristics of typical professional claims behaviors, such as weekly fixed complaints and intensive holiday reporting, and provides temporal prior constraints for information propagation in subsequent graph neural networks.

[0075] Step 4.3: Establish a reference-pointing edge between the text node containing legal references and the intent node. The weight of the reference-pointing edge is determined based on the similarity of their semantic expressions.

[0076] An intent node is a dynamically generated embedded node for legal citation intent, represented by a 64-dimensional vector output by the Legal Citation Intent Discriminator (LID). It is used to characterize the overall intent type of legal clause citation in the current complaint, including three categories: normal rights protection citation, templated claim citation, and evasive implicit citation.

[0077] Reference-pointing edges are used to establish a mapping relationship between specific legal reference fragments in the text and the overall reference intent, achieving semantic alignment from local text to global intent. Reference-pointing edges are only established between text nodes identified as related to legal references (such as token nodes containing expressions like "refund one, compensate three" or "based on the Food Safety Law") and intent nodes. The weight of the reference-pointing edge is calculated using cosine similarity to determine the semantic matching degree between the text node vector and the intent node vector. The higher the similarity, the greater the contribution of the text fragment to the overall legal intent, and the stronger the edge connection. This approach helps strengthen the model's focus on key legal expressions and improves its ability to identify implicit circumvention-type references.

[0078] Step 4.4: Calculate the cross-modal relevance score between each text node and each behavior event node, and establish cross-modal alignment edges based on the cross-modal relevance scores. The weights of the cross-modal alignment edges are determined based on the mutual attention mechanism between text nodes and behavior event nodes.

[0079] Cross-modal alignment edges are key bridges connecting textual and behavioral modalities, aiming to reveal the intrinsic relationship between the content of a complaint text and its corresponding complaint behavior. For example, the frequent appearance of standardized expressions such as "inspection report is unqualified" and "demand ten times compensation" in a certain text may correspond to multiple repeated complaints about high-value, similar products. This collaborative pattern can be modeled using cross-modal edges.

[0080] Cross-modal relevance scores are calculated using a learnable mutual attention mechanism. This involves treating the text node set and the behavioral event node set as the query and key, respectively, and calculating the attention weight matrix between them. These weights reflect the degree of matching between semantic units across different modalities. Finally, node pairs with scores above a threshold are selected to establish directed edges—cross-modal alignment edges—with the edge weights representing the normalized attention scores. This approach achieves fine-grained alignment between text content and historical behavior, enabling the model to recognize composite occupational claim patterns involving templated text and periodic behavior.

[0081] Step 4.5: Structure the text nodes, intent nodes, and behavior event nodes and their corresponding directed edges to generate a heterogeneous graph containing multiple node types and multiple edge types.

[0082] Text nodes, intent nodes, and behavioral event nodes, along with their corresponding directed edges, are structurally integrated. This involves unifying these three types of nodes and their four types of edges (semantic association edges, temporal evolution edges, reference-pointing edges, and cross-modal alignment edges) into a complete heterogeneous graph structure: the Text-BehaviorAlignment Graph (TBAG). This graph exhibits significant heterogeneity in both node and edge types: nodes encompass semantic units at the language level, event records at the behavioral level, and legal intent at the abstract level; edges represent semantic dependencies, temporal evolution, legal intent attribution, and cross-modal associations, respectively. During the integration process, all nodes are incorporated into a unified graph space, and edges are constructed according to their respective rules and independently weighted. The resulting graph structure contains an average of approximately 35 nodes and 82 edges, exhibiting a compact structure and rich information content.

[0083] Furthermore, the above-mentioned multi-layer feature aggregation of heterogeneous graphs using a pre-trained graph neural network, and the obtaining of the full graph representation vector through graph pooling, can include the following steps 5.1 to 5.3 in its specific implementation: Step 5.1: For each type of node, calculate the attention weight between it and its neighboring nodes of the same type. Then, perform a weighted summation of the feature vectors of the neighboring nodes based on the attention weights to obtain the first node feature after local information aggregation.

[0084] Graph Neural Networks (GNNs) employ a two-layer Graph Attention Network (GAT) to achieve phased fusion of multimodal features, with the first layer focusing on the aggregation of local information within a modality. Each node category includes three types: text nodes, behavioral event nodes, and intent nodes, corresponding to abstract representations of language expression units, historical complaint records, and legal citation intent, respectively.

[0085] In this step, only the connection edges between nodes of the same type are considered (such as semantic association edges between text nodes and temporal evolution edges between behavioral event nodes). Cross-type connections are shielded through a masking mechanism to ensure that information aggregation is limited to the modality.

[0086] Attention weights are calculated using a learnable attention mechanism. They are obtained by processing a linear combination of the node's own features and the features of its neighboring nodes using the LeakyReLU activation function, and are used to measure the information contribution of neighboring nodes to the central node. Subsequently, the feature vectors of each node in the neighborhood are weighted and summed using this weight, and a multi-head attention mechanism (configured with 4 attention heads) is introduced to enhance the stability of the representation. Finally, the updated features of the first node are output.

[0087] Step 5.2: Using the first node features as input, calculate the cross-modal attention weights between different types of nodes, and perform the weighted summation operation again to obtain the second node features after global information aggregation.

[0088] In this step, the first node feature output from the previous stage is used as input, the modality type restriction is removed, and all nodes are allowed to exchange information based on the cross-modal edges they connect (such as reference pointing edges and cross-modal alignment edges).

[0089] Specifically, the attention scores between the current node and all its neighboring nodes (regardless of their type) are calculated and then normalized to form cross-modal attention weights. Then, the features of neighboring nodes are weighted and aggregated based on these cross-modal attention weights to complete a global feature update, generating second-node features with more comprehensive discriminative power.

[0090] This method particularly enhances the ability to collaboratively identify high-claim statements mentioned in the text and high-frequency, high-value complaints in historical behavior, effectively capturing the complex behavioral patterns of professional claimants.

[0091] Step 5.3: Perform a weighted average operation on the features of the first node and the features of the second node to generate a full graph representation vector that characterizes the state of the entire heterogeneous graph.

[0092] The full-graph representation vector is the unified feature representation ultimately used for classifying professional claimants, and it must fully preserve the information evolution trajectory of each node during the local and global aggregation process. In this step, instead of directly using the output of a single layer, a hierarchical feature fusion strategy is adopted: First, the first node feature (reflecting the local structure within the modality) and the corresponding second node feature (reflecting the global relationship across modalities) of each node are weighted and averaged to form the fused feature representation of that node. The weights can be set through learnable parameters or fixed proportions to ensure that both the original semantic details are preserved and higher-order relational information is incorporated. Subsequently, attention pooling is performed on the fused features of all nodes, that is, an importance score for each node is generated through a single-layer neural network, and a weighted average is performed accordingly to generate a 128-dimensional fixed-length vector as the global representation of the entire heterogeneous graph.

[0093] The aforementioned target graph neural network is trained using a distillation framework comprised of a teacher model and a student model. The training method for this target graph neural network includes the following steps 6.1 to 6.5: Step 6.1: Construct a teacher model and pre-train it using a category-weighted strategy to obtain auxiliary supervision information.

[0094] A teacher model refers to a leading model with strong discriminative ability on the same task, used to provide soft-label outputs (i.e., class probability distributions) as a knowledge source for student models. It belongs to the guiding model in knowledge distillation. The category weighting strategy is a key measure to address the severe data imbalance problem in the task of identifying professional claimants. Because professional claimant samples account for a very small percentage of the overall complaint data (usually less than 5%, and in extreme cases only 2%), direct training can easily lead to the model being biased towards the majority class.

[0095] Auxiliary supervision information refers to the probability vector with confidence output by the teacher model for each training sample. It not only contains the final classification result, but also contains the uncertainty distribution of the sample near the decision boundary. It is also called soft label and can convey richer learning signals than hard label (real label information).

[0096] Step 6.2: Construct the student model, whose structure is consistent with the target graph neural network used for prediction.

[0097] The structure of the student model is consistent with that of the target graph neural network used in the final deployment. Its structure is the aforementioned two-layer graph attention network (GAT) based on cross-modal aligned graph (TBAG), which includes all components such as intramodal local aggregation, cross-modal global aggregation and attention pooling. It outputs a 128-dimensional full graph representation vector and connects it to a classification head to complete the binary classification task.

[0098] Step 6.3: The student model is jointly optimized using the auxiliary supervision information and the real category labeling information output by the teacher model after pre-training. For each training sample, the first loss between the student model output and the auxiliary supervision information and the second loss between the student model output and the real label are calculated.

[0099] Joint optimization primarily integrates the advantages of knowledge distillation and traditional supervised learning to enhance the generalization ability of student models under sparse sample conditions. The first loss is the knowledge distillation loss (in the form of KL divergence), which measures the difference between the probability distribution output by the student model and the soft labels provided by the teacher model, guiding the student to imitate the teacher's decision-making behavior, especially absorbing its judgment logic for difficult and marginal samples. The second loss is the standard cross-entropy loss, reflecting the error between the student model's prediction and the true class labels, ensuring the model's basic classification performance.

[0100] In this implementation, after each batch of training samples is input, the student model forward propagates to generate a predicted output, and calculates the distillation loss with the soft label output by the teacher model and the cross-entropy loss with the real label. The two together constitute the learning objective for joint optimization.

[0101] Step 6.4: Calculate the weighted sum of the first loss and the second loss for each training sample, and determine it as the target training loss for the training sample.

[0102] The target training loss serves as the basis for updating the student model parameters and is composed of distillation loss and cross-entropy loss in an adjustable ratio. In the specific implementation, the total loss function is set as the sum of three terms: cross-entropy loss, knowledge distillation loss, and hard example mining loss, with the first two terms constituting the weighted sum described here. The weight coefficients are initially set to a balanced state (e.g., each accounting for 0.5), but can also be dynamically adjusted according to the training phase.

[0103] Step 6.5: Adjust the parameters of the student model based on the target training loss of all training samples to obtain the target graph neural network.

[0104] Parameter tuning involves using the backpropagation algorithm to pass the average target training loss of a batch of samples to each layer of the student model, and then using the optimizer to update the network weights to achieve iterative convergence of the model. In this implementation, the AdamW optimizer is used, with an initial learning rate of 2e-5, a batch size of 32, and a total of 30 training epochs. The hardware platform is an NVIDIA Tesla V100 GPU. During training, the F1 score on the validation set is monitored. If there is no improvement for several consecutive epochs, the learning rate is reduced or training is terminated early to prevent overfitting. Finally, the student model with the best performance is saved as the target graphical neural network for formal deployment. Through the above distillation training framework, this scheme effectively solves the model bias and underlearning problems caused by extreme sample imbalance in the professional claimant identification task.

[0105] The technical solution of the present invention will be described in detail below with reference to a specific embodiment, aiming to identify professional claimants in complaint and reporting platform data in practical applications. See also... Figure 2 As shown, it includes the following S1 to S6: S1: Obtain the original complaint text data.

[0106] S2, Data preprocessing and basic feature extraction.

[0107] 2-1) Text Data Preprocessing: The original complaint text is cleaned, specifically including: removing HTML tags using the BeautifulSoup library; converting full-width characters to half-width characters using the `unicodedata.normalize` function; filtering invalid symbols except for Chinese, English, numbers, and common punctuation marks using regular expressions; and performing synonym replacement based on the Harbin Institute of Technology's thesaurus to normalize the text, resulting in the following text sequence: .

[0108] 2-2) Behavioral Data Processing: Collect users' historical complaint records from the past two years, arrange them in ascending order by timestamp, and construct a sequence of behavioral events. Each behavioral event includes 12 basic attributes: complaint timestamp (accurate to the second), claim amount (RMB), product category, unified social credit code of the complainant, complaint channel (e.g., APP, webpage, telephone), processing result (e.g., case filed, case not filed, successful mediation), complainant's geographic region code, fingerprint / IP address encryption string of the complaining device, label of the type of illegal or irregular activity involved, total number of complaints in the user's history, historical complaint success rate / case closure rate, and time interval between the previous complaint and the previous complaint.

[0109] S3: Multidimensional Feature Engineering S3-1, Text Semantic Feature Extraction: Using the "hfl / rbt3" Chinese RoBERTa variant as a pre-trained model, the pre-processed text sequence (truncated to a maximum length of 512 tokens) is input into the model for encoding. The [CLS] vector and the hidden state vectors of the Top-10 tokens with the highest attention weights are extracted to form the initial features of the text node set. Initial semantic vector generation:

[0110] in This represents the encoding function of the pre-trained language model. Let i be the semantic vector of the i-th token. The semantic vector dimension (default 768).

[0111] Sorting users' historical complaint behaviors in ascending order by timestamp, a sequence of behavioral events is constructed:

[0112] Each behavioral event It includes 12 basic attributes: timestamp, claim amount, product category, unified social credit code of the complainant, complaint channel, processing result, complainant's geographic region code, fingerprint / IP address encryption string of the complaining device, label of the type of violation involved, total number of complaints in the user's history, historical complaint success rate / case closure rate, and time interval with the previous complaint.

[0113] S3-2, Behavioral Spectrum Feature (BFS) Generation: Converting timestamps in a sequence of behavioral events into a daily, equally spaced sequence of sampled numerical values. ,in This represents the timestamp of the i-th complaint. The intensity of complaints at the corresponding time point (weighted by the amount of claims and the number of complaints).

[0114] Perform a Fast Fourier Transform using Python's `scipy.fftpack.fft` function:

[0115] in For a frequency domain signal, k is the frequency component index. The following key features are extracted from the spectrum: (1) dominant frequency (1) The frequency component with the highest energy; (2) Spectral entropy (3) Energy percentage of the first 5 peak frequencies; (4) 10 key features such as periodic stability index, including main frequency, spectral entropy, energy percentage of the first 5 peak frequencies, etc. This spectral feature vector is used to capture periodic behavior patterns.

[0116] After standardizing and concatenating the above features, a 22-dimensional behavior spectrum vector is generated:

[0117] This feature can effectively characterize typical professional claim behavior patterns such as weekly / monthly periodic complaints and concentrated complaints during holidays.

[0118] S3-3, Legal Citation Extraction and Intent Determination (LID): First, a pre-trained BiLSTM-CRF model (128-dimensional LSTM hidden layers) is used to extract legal citation fragments from the text. This includes explicit citations (such as "according to Article 148 of the Food Safety Law") and implicit citations (such as "demanding ten times the compensation" corresponding to relevant clauses of the Food Safety Law). The model achieves an F1 score of 0.91 on the legal clause annotation dataset. Subsequently, the extracted citation fragments are input into a legal intent discriminator, which employs a 3-layer Transformer encoder structure and outputs the probability distributions of the three types of intents:

[0119] in and For learnable parameters, This is the hidden layer dimension of the Transformer. Intent categories include: (1) normal rights protection citation (reasonable legal basis citation in real consumer disputes); (2) templated claim citation (professional claim feature citation using fixed sentence structures and standardized expressions); and (3) evasive implicit citation (indirectly citing legal clauses through vague expressions to evade detection). A 64-dimensional legal intent vector is generated by sampling the probability distribution using Gumbel-Softmax. :

[0120] Step 4: Construction of Cross-Modal Alignment Graph (TBAG) 4-1) Node Construction: Construct a heterogeneous graph based on the extracted features. , where the node set It includes three types of nodes: ① Text node set Selecting text semantic vectors The top-10 token vectors with the highest attention weights are used as text nodes. ; ② Behavioral event nodes The 12-dimensional basic attributes from step 2-2 are concatenated with the 10-dimensional spectral features from step 2.2 to form a 22-dimensional feature vector, representing each behavioral event. For each node, the node features are a concatenation of basic behavioral attributes and behavioral spectrum features: ; ③ Legal Intent Node : Including those from the current complaint The dynamic intent nodes that constitute the structure.

[0121] 4-2) Edge set construction: Four types of directed edges are established between nodes to characterize the relationships: ①Semantic edges within the text: weights are derived from the attention matrix of layer 11 of the RoBERTa model; ② Behavioral temporal edges (i.e., the aforementioned temporal evolution edges): The weight is determined by the time interval function. Confirmed, where T is one day (86400 seconds); ③Legal citation semantic edge (i.e., the aforementioned citation pointing edge): The weight is the cosine similarity between the citation fragment and the embedding vector of the intent node; ④ Cross-modal alignment edges: The association weights between text nodes and behavior nodes are calculated through a learnable mutual attention mechanism.

[0122] The final TBAG graph contains 24-30 nodes and 60-80 edges (averaging about 35 nodes and 82 edges), forming a compact and informative cross-modal association representation.

[0123] Step 5: Model Training and Optimization 5-1) Graph Network Feature Fusion: A two-layer GAT network is used to fuse features of TBAG. The first layer of GAT performs local aggregation of nodes within each modality:

[0124] in Let i be the neighborhood of node i. The first layer of attention weights uses a masking mechanism to ensure that only nodes of the same modality participate in the aggregation.

[0125] The first layer maps the input node features of different dimensions (768 dimensions for text, 22 dimensions for behavior, and 64 dimensions for law) to a unified 256-dimensional map.

[0126] The second-layer GAT performs cross-modal global aggregation and introduces modal attention gating:

[0127] in For cross-modal neighborhood, For modal indication vectors, This is for modal bias.

[0128] The second layer outputs 128-dimensional fused node features. The network configuration includes four attention heads, a dropout rate of 0.2, and a LeakyReLU activation function.

[0129] Finally, graph pooling is used to aggregate all node features into a single full graph representation vector:

[0130] This full-graph representation vector integrates cross-modal relational information of text semantics, behavioral patterns, and legal intent.

[0131] 5-2) Adaptive Imbalanced Data Distillation (AIDD) Training: A teacher-student distillation framework is used for model training. The teacher model is trained using cross-entropy loss with class weights (non-compensated:compensated = 1:20). The student model is optimized by learning the soft and hard labels from the teacher model. Total loss function... It includes cross-entropy loss, distillation loss, and hard example mining loss based on sample uncertainty:

[0132] in For cross-entropy loss, Distillation loss ( For teacher models, soft labels Output for student model). Difficult case mining loss. Adaptive weights. Dynamically adjusted based on sample uncertainty:

[0133] in Due to sample uncertainty, This is the temperature parameter (default 2.0). This is the uncertainty threshold (default 0.6).

[0134] Training was performed using the AdamW optimizer with an initial learning rate of 2e-5, a batch size of 32, and a total of 30 training epochs. The training hardware environment was an NVIDIA Tesla V100 GPU. This strategy makes the training weights of the professional claim samples highly dynamic, significantly improving the model's performance on imbalanced data.

[0135] 5-3) Performance Comparison. To verify the superiority of the method of this invention, the following baseline methods were selected for comparison on the same test set: (1) Keyword matching: a rule engine-based method. (2) Statistical features + LR: using only behavioral statistical features and classifying by logistic regression. (3) BERT-Base: using only BERT to classify text.

[0136] The performance comparison results are shown in Table 1:

[0137] As can be seen from the results in Table 1, the professional claimant identification method (TBAG) based on cross-modal alignment graph structure proposed in this invention is significantly better than other baseline methods in all indicators, proving its effectiveness and advancement in identifying professional claimants.

[0138] Step 6: Generation of Reasoning and Explanation Chains 6-1) Identification of Professional Claimants: During the inference phase, the test sample is processed to obtain a full-image representation vector, which is then input into the final classifier, and the output is a prediction of whether the claimant is a professional claimant. This process can be completed in real time on an Intel Core i7 CPU, with a single-sample inference time of less than 200 milliseconds.

[0139] 6-2) Chain-based Explainable Reasoning (CoRA) Generation: To meet the explainability requirements of law enforcement scenarios, starting from the final prediction result, the GNNExplainer method is used to trace back the nodes and edges with the highest contribution in the GAT network. The system generates a structured reasoning chain based on this key information. ,in Key nodes (text token / behavioral event / legal intent). The weight of its contribution to the prediction results is determined. The inference chain contains four dimensions: (1) behavioral cycle inference (based on the peak characteristics of the spectrum, such as "fixed complaints every Friday, cycle characteristics contribute 32%"); (2) legal intent judgment (such as "implicitly circumventing the reference to Article 28 of the Advertising Law, intent characteristics contribute 27%"); (3) text template structure (such as "using the standard sentence structure of 'refund one and compensate ten + test report', text features contribute 23%"); (4) cross-modal association (such as "the text legal reference is highly consistent with the behavioral cycle, association features contribute 18%").

[0140] The generated reasoning chain is output in both natural language and visual graph (implemented using D3.js). For example, the generated explanation is: "User behavior exhibits significant periodic characteristics (dominant frequency f=1 / 7, contribution 32%), the text contains a templated expression of 'refund one, compensate ten' (match degree 0.92, contribution 25%), legal citations are judged as implicit evasion intent (confidence degree 0.87, contribution 28%), and the three factors together point to a 94.3% probability of professional claims." This reasoning chain can be directly used to generate law enforcement verification reports, meeting the interpretability requirements of law enforcement scenarios.

[0141] In summary, by extracting key semantic units and legal citation intent nodes from the text, the ability to identify the core content and legal citation intent of complaint texts is enhanced; by converting behavioral sequences into spectral features, dynamic patterns such as periodicity and clustering are effectively captured; by utilizing heterogeneous graph structures and graph neural networks to achieve structured association and fusion of multimodal information, the efficiency of utilizing cross-modal collaborative features is improved; and by strengthening the representation of sparse samples through full graph representation learning, high recognition accuracy and robustness are maintained even under imbalanced data conditions.

[0142] Based on the above method embodiments, this application also provides an abnormal data prediction device for unfair competition behavior on the internet, see [link to relevant documentation]. Figure 3 As shown, the device includes the following parts: The acquisition module 310 is used to acquire the text data of the target complaint event and the behavior sequence composed of multiple historical complaint events of the corresponding user; The text node extraction module 320 is used to perform semantic encoding on text data and extract semantic units carrying significant attention weights as text nodes. The intent node generation module 330 is used to identify legal citation fragments from text data to generate intent nodes that represent the citation intent; The behavior node construction module 340 is used to construct behavior event nodes based on historical complaint events. The event node includes event attributes and spectral features obtained by converting the timestamp sequence. The spectral features are obtained by performing frequency domain transformation on the complaint intensity signal and extracting the frequency distribution characteristics. Graph structure building module 350 is used to build heterogeneous graphs containing text nodes, intent nodes, and behavior event nodes; The graph neural network processing module 360 ​​is used to perform multi-layer feature aggregation on heterogeneous graphs through the target graph neural network and obtain the full graph representation vector through graph pooling; wherein, multi-layer feature aggregation includes local information fusion between nodes of the same type and global information interaction between nodes of different types. Prediction module 370 is used to predict whether a target complaint event is abnormal data related to unfair competition on the Internet based on the full graph representation vector.

[0143] In one feasible implementation, the target graph neural network is trained using a distillation framework consisting of a teacher model and a student model. The device further includes a model training module for: constructing a teacher model and pre-training it using a category weighting strategy to obtain auxiliary supervision information; constructing a student model whose structure is consistent with the target graph neural network used for prediction; jointly optimizing the student model using the auxiliary supervision information and real category labeling information output by the teacher model after pre-training; for each training sample, calculating a first loss between the student model output and the auxiliary supervision information, and a second loss between the student model output and the real label; calculating the weighted sum of the first loss and the second loss for each training sample to determine the target training loss for the training sample; and adjusting the parameters of the student model according to the target training loss of all training samples to obtain the target graph neural network.

[0144] The abnormal data prediction device for unfair competition on the Internet provided in this application has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the embodiment of the abnormal data prediction device for unfair competition on the Internet can be referred to the corresponding content in the aforementioned embodiment of the abnormal data prediction method for unfair competition on the Internet.

[0145] This application also provides an electronic device, such as... Figure 4 The diagram shows the structure of the electronic device 100, which includes a processor 41 and a memory 40. The memory 40 stores computer-executable instructions that can be executed by the processor 41. The processor 41 executes the computer-executable instructions to implement any of the above-mentioned abnormal data prediction methods for unfair competition in the network.

[0146] exist Figure 4 In the illustrated embodiment, the electronic device further includes a bus 42 and a communication interface 43, wherein the processor 41, the communication interface 43, and the memory 40 are connected via the bus 42.

[0147] The memory 40 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 43 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 42 may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 42 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 4 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0148] Processor 41 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 41 or by software instructions. Processor 41 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in the memory. The processor 41 reads the information in the memory and, in conjunction with its hardware, completes the steps of the abnormal data prediction method for unfair competition behavior on the network as described in the foregoing embodiments.

[0149] This application also provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are called and executed by a processor, they cause the processor to implement the above-described abnormal data prediction method for unfair competition behavior on the network. For specific implementation details, please refer to the foregoing method embodiments, which will not be repeated here.

[0150] The computer program product provided in this application embodiment for an abnormal data prediction method for unfair competition on the Internet includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0151] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0152] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting abnormal data related to unfair competition on the internet, characterized in that, include: Obtain the text data of the target complaint event and the behavioral sequence composed of multiple historical complaint events of the corresponding user; The text data is semantically encoded, semantic units carrying significant attention weights are extracted as text nodes, and legal citation fragments are identified from the text data to generate intent nodes representing the citation intent. Based on the historical complaint events, behavioral event nodes are constructed. The event nodes include event attributes and spectral features obtained by converting timestamp sequences. The spectral features are obtained by performing frequency domain transformation on the complaint intensity signal and extracting frequency distribution characteristics. Construct a heterogeneous graph that includes the text nodes, the intent nodes, and the behavior event nodes; The heterogeneous graph is subjected to multi-layer feature aggregation through a target graph neural network, and a full graph representation vector is obtained through graph pooling; wherein, the multi-layer feature aggregation includes local information fusion between nodes of the same type and global information interaction between nodes of different types. Based on the full graph representation vector, predict whether the target complaint event belongs to abnormal data related to unfair competition in the network.

2. The abnormal data prediction method for unfair competition on the internet according to claim 1, characterized in that, Obtain the text data of the target complaint event and the behavioral sequence composed of multiple historical complaint events of the corresponding user, including: Obtain the text data of the target complaint event, clean the text data, remove invalid symbols and format tags, and convert full-width characters to half-width characters; The cleaned text data is segmented into words, and keywords are normalized and replaced based on a thesaurus. The historical complaint events of the user matched with the target complaint time are arranged in chronological order according to timestamps to generate an ordered sequence of behavioral events.

3. The abnormal data prediction method for unfair competition on the internet according to claim 1, characterized in that, The text data is semantically encoded, semantic units carrying significant attention weights are extracted as text nodes, and legal citation fragments are identified from the text data to generate intent nodes representing the citation intent, including: The preprocessed text data is input into the pre-trained language model for encoding to obtain the hidden state vector of each semantic unit; Based on the attention mechanism associated with classification symbols in the last layer of the pre-trained language model, the attention weight corresponding to each semantic unit is determined; Select target semantic units whose attention weight is higher than a set threshold, and use the hidden state vector corresponding to the target semantic unit as a text node; Identify legal clause citation fragments contained in text data using sequence labeling models; The reference fragment is input into the intent classification model, the reference intent category is output, and an intent node corresponding to the intent category is generated.

4. The abnormal data prediction method for unfair competition on the internet according to claim 1, characterized in that, Behavioral event nodes are constructed based on the historical complaint events. Each event node includes event attributes and spectral features obtained from timestamp sequence transformation. These spectral features are obtained by performing frequency domain transformation on the complaint intensity signal and extracting its frequency distribution characteristics, including: Extract the timestamp, claim amount, and product category information from each historical complaint. The historical complaint events are sorted according to the timestamp order to obtain an ordered sequence of behavioral events; A time point sequence sampled in a fixed time unit is generated based on the behavioral event sequence. The complaint intensity corresponding to each time point is determined by weighting the claim amount and the number of complaints of historical complaint events falling within that time period. The frequency domain transformation of the complaint intensity signal corresponding to the time point sequence is performed to obtain the frequency component distribution; The spectral characteristics are formed by extracting the dominant frequency, the degree of spectral concentration, the energy proportion of periodic peaks, and the periodic stability index from the frequency component distribution. The original event attributes corresponding to each historical complaint event are concatenated with the spectral features to generate the corresponding behavioral event node.

5. The abnormal data prediction method for unfair competition on the internet according to claim 1, characterized in that, Constructing a heterogeneous graph containing the text nodes, intent nodes, and behavior event nodes includes: Semantic association edges are established between the text nodes, and the connection relationship of the semantic association edges is determined based on the contextual dependency relationship between semantic units; A temporal evolution edge is established between the behavioral event nodes in chronological order, and the weight of the temporal evolution edge decreases as the time interval between the two events increases. A reference-pointing edge is established between the text node containing legal references and the intent node, and the weight of the reference-pointing edge is determined based on the similarity of their semantic expressions. Calculate the cross-modal relevance score between each text node and each action event node, and establish the cross-modal alignment edge based on the cross-modal relevance score. The weight of the cross-modal alignment edge is determined based on the mutual attention mechanism between the text node and the action event node. The text nodes, intent nodes, and behavior event nodes, along with their corresponding directed edges, are structurally integrated to generate a heterogeneous graph containing multiple node types and multiple edge types.

6. The abnormal data prediction method for unfair competition on the internet according to claim 1, characterized in that, The heterogeneous graph is subjected to multi-layer feature aggregation using a pre-trained graph neural network, and a full graph representation vector is obtained through graph pooling, including: For each type of node, calculate the attention weight between it and its neighboring nodes of the same type. Then, perform a weighted summation of the feature vectors of the neighboring nodes based on the attention weights to obtain the first node feature after local information aggregation. Using the first node feature as input, calculate the cross-modal attention weights between different types of nodes, and perform a weighted summation operation again to obtain the second node feature after global information aggregation. A weighted average operation is performed on the features of the first node and the features of the second node to generate a full graph representation vector that characterizes the state of the entire heterogeneous graph.

7. The abnormal data prediction method for unfair competition on the internet according to claim 1, characterized in that, The target graph neural network is trained using a distillation framework consisting of a teacher model and a student model; the method further includes: The teacher model is constructed and pre-trained using a category-weighted strategy to obtain auxiliary supervision information; Construct a student model whose structure is consistent with the target graph neural network used for prediction; The student model is jointly optimized using the auxiliary supervision information and the real category labeling information output after the teacher model is pre-trained. For each training sample, the first loss between the student model output and the auxiliary supervision information, and the second loss between the student model output and the real label are calculated respectively. Calculate the weighted sum of the first loss and the second loss for each training sample, and determine it as the target training loss for the training sample; The parameters of the student model are adjusted based on the target training loss of all training samples to obtain the target graphical neural network.

8. An abnormal data prediction device for unfair competition on the internet, characterized in that, The acquisition module is used to acquire the text data of the target complaint event and the behavioral sequence composed of multiple historical complaint events of the corresponding user; The text node extraction module is used to perform semantic encoding on the text data and extract semantic units carrying significant attention weights as text nodes. An intent node generation module is used to identify legal citation fragments from the text data to generate intent nodes that represent the citation intent; A behavior node construction module is used to construct behavior event nodes based on the historical complaint events. The event nodes include event attributes and spectral features obtained by converting timestamp sequences. The spectral features are obtained by performing frequency domain transformation on the complaint intensity signal and extracting frequency distribution characteristics. The graph structure construction module is used to construct a heterogeneous graph that includes the text nodes, the intent nodes, and the behavior event nodes; The graph neural network processing module is used to perform multi-layer feature aggregation on the heterogeneous graph through the target graph neural network and obtain the full graph representation vector through graph pooling; wherein, the multi-layer feature aggregation includes local information fusion between nodes of the same type and global information interaction between nodes of different types. The prediction module is used to predict whether the target complaint event belongs to abnormal data related to unfair competition behavior on the network, based on the full graph representation vector.

9. An electronic device, characterized in that, The device includes a processor and a memory, the memory storing computer-executable instructions that can be executed by the processor, the processor executing the computer-executable instructions to implement the abnormal data prediction method for unfair competition behavior on the network as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the abnormal data prediction method for unfair competition behavior on the network as described in any one of claims 1 to 7.

Images