An ai-driven industry user reputation collection and intelligent public opinion analysis system and method

Abstract

This invention discloses an AI-driven system and method for collecting industry user reviews and analyzing intelligent public opinion, belonging to the field of machine learning. The system includes: a data acquisition module for collecting review data; a normal pattern graph construction module for acquiring historical review data during normal operation and constructing a normal propagation pattern graph; a graph contrastive learning training module for training a graph neural network through graph contrastive learning to obtain the trained graph neural network model and the feature space cluster centers of the normal propagation pattern; a dynamic prototype correction module for dynamically adjusting the feature space cluster centers through online calibration; and an anomaly detection module for calculating the distance between the feature embedding and the current cluster center based on the target review content and the graph neural network model to determine whether there are anomalies in the propagation path of the target review content. This invention can identify abnormal propagation behavior of user reviews across multiple stores of a chain brand.

Description

Technical Field

[0001] This invention relates to the field of machine learning, and in particular to an AI-driven system and method for collecting industry user reputation and analyzing public opinion. Background Technology

[0002] With the widespread use of social media, word-of-mouth has become a key factor in reputation management and marketing decisions for chain brands. User reviews and interactions on platforms such as Xiaohongshu, Douyin, and Dianping directly impact consumer perception and market competitiveness. For brands, timely and accurate monitoring and analysis of word-of-mouth content on social media, and identifying abnormal dissemination behaviors, has become a core requirement of brand management.

[0003] Existing technologies lack an anomaly detection solution for word-of-mouth communication paths that can model normal propagation patterns for multi-store scenarios of chain brands, establish normal / abnormal discrimination boundaries through graph comparison learning, and support differentiated store strategies. Summary of the Invention

[0004] This invention aims to address at least one of the technical problems existing in the prior art. To this end, this invention proposes an AI-driven industry user reputation collection and intelligent public opinion analysis system and method, which can detect abnormal reputation behavior in multi-store scenarios of chain brands where the content is compliant but the dissemination path is abnormal by constructing a normal propagation pattern map and establishing a normal cluster benchmark in the feature space using graph contrastive learning, and by using the distance metric between feature embedding and cluster center.

[0005] This invention provides an AI-driven industry user reputation collection and intelligent public opinion analysis system, comprising: a data collection module for collecting reputation data of a target object from multiple social media platforms, wherein the reputation data includes at least text content and interaction data; a normal pattern graph construction module for acquiring historical reputation data of the target object during normal operation, extracting content nodes, user nodes, and propagation path nodes, and constructing a normal propagation pattern graph using the publishing, forwarding, and interaction relationships between nodes as edges; wherein the content nodes store text semantic feature vectors and publishing time features of the historical reputation data, and the propagation path nodes store feature representations of a propagation link between the initial publication of reputation content and various interaction nodes, and the propagation path nodes are generated by associating their corresponding content nodes and user nodes participating in the propagation path; and a graph comparison learning training module for constructing positive sample pairs and negative sample pairs from the normal propagation pattern graph, wherein the positive sample pair includes two content nodes that are related to each other under the same normal propagation pattern. The negative sample pairs are pairs of normal content nodes and known abnormal content nodes, which are distributed across different platforms. A graph neural network is trained using graph contrastive learning to map content nodes in the normal propagation pattern to normal clustering regions in the feature space, and to map content nodes in the abnormal propagation pattern to locations far from these normal clustering regions, resulting in a trained graph neural network model and feature space cluster centers for the normal propagation pattern. A dynamic prototype correction module is used to respond to brand strategy adjustments or new platform access events, acquire newly added normal propagation samples, dynamically adjust the feature space cluster centers through online calibration, and limit the time range of samples participating in cluster center calculation using a sliding window mechanism. An anomaly detection module is used to input the target word-of-mouth content collected in real time by the data acquisition module into the trained graph neural network model, obtain its feature embedding in the feature space, calculate the distance between the feature embedding and the current cluster center, and determine that the propagation path of the target word-of-mouth content is abnormal when the distance exceeds a preset threshold, and output the anomaly detection result.

[0006] According to some embodiments of the present invention, the contrastive loss function used by the graph contrastive learning training module is:

[0007] ;

[0008] in, and For each anchor point i, the feature embedding of the two samples in the positive sample pair is uniquely determined in each training iteration; 2N is the total number of samples in a training batch, and each original data point is augmented to obtain two positive sample pairs. The cosine similarity function is τ, and the temperature parameter is τ. The construction of the positive sample pairs also includes: taking the feature representations of the same content node in different time windows as positive sample pairs to enhance the model's ability to learn temporal propagation patterns.

[0009] According to some embodiments of the present invention, the online calibration method of the dynamic prototype correction module is as follows: when a new normal propagation sample is obtained, the offset vector between the feature embedding of the sample and the current cluster center is calculated, and the offset vector is weighted and superimposed on the current cluster center with a preset learning rate to realize incremental update of the cluster center; the window size of the sliding window mechanism is adaptively set according to the brand strategy adjustment frequency.

[0010] According to some embodiments of the present invention, the anomaly detection module is further configured to, when it is determined that there is an anomaly in the propagation path of the target word-of-mouth content, further calculate the distance between the feature embedding of the target word-of-mouth content and the cluster centers of multiple known anomaly types, and output the anomaly diagnosis result according to the anomaly type with the closest distance; the anomaly types include at least the anomaly of brushing traffic and water army propagation, the anomaly of secondary propagation of content tampering, and the anomaly of inconsistent propagation across platforms.

[0011] According to some embodiments of the present invention, the preset threshold is a multi-level dynamic threshold, which includes a strategy differentiation threshold set according to store type and a regional threshold set according to region; different store types set different anomaly tolerance levels according to their brand strategy configuration information, and the anomaly tolerance threshold of strategically aggressive stores is higher than that of strategically conservative stores.

[0012] According to some embodiments of the present invention, the system is deployed in a cloud-edge collaborative architecture. The graph contrastive learning training module is deployed on a cloud server to perform offline training and periodic retraining of the graph neural network. The normal pattern graph construction module, dynamic prototype correction module, and anomaly detection module are deployed on edge nodes to perform real-time graph construction, online cluster center calibration, and anomaly detection. Only cluster center parameters and anomaly detection results are transmitted between the edge nodes and the cloud server, and no original reputation data is transmitted.

[0013] According to some embodiments of the present invention, the system further includes a store strategy parameterization module, which is used to obtain brand strategy configuration information of each store and convert the brand strategy configuration information into a store strategy feature vector. The brand strategy configuration information includes at least a store type tag, a content tone preference tag, and a publishing strategy tag. The normal pattern graph construction module is also used to extract store-level normal dissemination samples from the historical reputation data of the store during normal operation for each store and construct a store-level normal dissemination pattern sub-graph. The graph contrastive learning training module is also used to perform store-conditional contrastive learning training in the feature space, using the content nodes in the store-level normal dissemination pattern sub-graph as positive samples for the store and the normal content nodes of other stores and the global abnormal content nodes as negative samples, to obtain the store-specific cluster centers of each store in the feature space.

[0014] According to some embodiments of the present invention, the generation method of the store-specific cluster center is as follows: using the feature space cluster center of the global normal propagation mode as the reference cluster center, the store strategy feature vector is input into a bias generation network, the bias generation network outputs a store-specific bias vector, and the reference cluster center and the store-specific bias vector are added to obtain the store-specific cluster center; the bias generation network is a multilayer perceptron network, and its training objective is to minimize the distance between the normal sample feature embedding of the store and the store-specific cluster center; the dynamic prototype correction module is also used to recalculate the store's strategy feature vector and update the store-specific cluster center through the bias generation network when the store's brand strategy configuration information changes.

[0015] According to some embodiments of the present invention, the anomaly detection module is further configured to: when the target word-of-mouth content of the target store is determined to be normal, further calculate the average distance between the feature embedding of the target word-of-mouth content and the store-specific cluster centers of other stores in the store type to which the target store belongs; if the average distance exceeds a preset same-type migration threshold, then determine that the target word-of-mouth content has an anomaly in cross-store propagation pattern migration; the system also includes a store strategy clustering analysis module, configured to perform clustering analysis on the store-specific cluster centers of all stores; when the distance between the store-specific cluster center of a store and the cluster center of its store type exceeds a store-level deviation threshold, mark the store as a strategy deviation store and generate a store propagation pattern deviation report.

[0016] Another aspect of this invention provides an AI-driven method for collecting industry user reputation data and intelligent public opinion analysis, comprising the following steps: S1, collecting historical reputation data of a target object during normal operation, extracting content nodes, user nodes, and propagation path nodes, and constructing a normal propagation pattern graph; S2, constructing positive sample pairs and negative sample pairs from the normal propagation pattern graph, training a graph neural network through graph contrastive learning, mapping the content nodes of the normal propagation pattern to normal clustering regions in the feature space, obtaining the trained graph neural network model and the feature space cluster centers of the normal propagation pattern; S3, in response to brand strategy adjustments or new platform access events, acquiring newly added normal propagation samples, dynamically adjusting the feature space cluster centers through online calibration, and limiting the time range of samples participating in cluster center calculation using a sliding window mechanism; S4, collecting target reputation content in real time, inputting the target reputation content into the trained graph neural network model, obtaining its feature embedding in the feature space, calculating the distance between the feature embedding and the current cluster center, and determining that the propagation path of the target reputation content is abnormal when the distance exceeds a preset threshold, and outputting the abnormality detection result.

[0017] The embodiments of the present invention achieve at least the following beneficial effects:

[0018] (1) This invention does not rely on text sentiment analysis or topic classification of word-of-mouth content. Instead, it constructs a normal dissemination pattern map and uses graph contrastive learning to establish feature space clusters of normal dissemination patterns. For newly collected word-of-mouth content, the system determines whether there are abnormal dissemination paths by calculating its distance from the normal cluster centers in the feature space. This invention enables the system to identify hidden abnormal behaviors where the text content appears normal, but the dissemination pattern significantly deviates from the brand's normal patterns.

[0019] (2) This invention introduces a dynamic prototype correction module, which can dynamically fine-tune the cluster centers in the feature space through online calibration in response to brand strategy adjustments or new platform access events, and limits the time range of samples participating in cluster center calculation with a sliding window mechanism. When a brand adjusts its strategy, the system does not need to retrain the entire graph neural network model; it only needs to complete the smooth migration of the normal standard through the gradual update of the cluster centers. This solves the technical problem of the sharp increase in false alarm rate of existing static models after brand strategy changes.

[0020] (3) This invention transforms the brand strategy configuration information of each store into a store strategy feature vector through a store strategy parameterization module, and generates a personalized bias vector for each store through a bias generation network, thereby obtaining the store-specific cluster center for each store. The creative content of the flagship store will not be mistakenly judged as abnormal in a one-size-fits-all manner, and the standardized implementation of community stores can also be effectively supervised.

[0021] (4) This invention also has a cross-store propagation pattern migration detection function, which can identify hidden anomalies in which a single store appears normal but is inconsistent with the overall propagation pattern of other stores of the same type; and a store strategy clustering analysis function, which can identify stores whose overall strategy deviates from the macro level. This helps brand headquarters to discover stores that are out of control before problems escalate.

[0022] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0024] Figure 1 This is a schematic block diagram of the system modules according to an embodiment of the present invention;

[0025] Figure 2 This is a schematic block diagram of a system according to another embodiment of the present invention;

[0026] Figure 3 This is a flowchart illustrating the method according to an embodiment of the present invention.

[0027] Figure label:

[0028] Data acquisition module 100, normal pattern graph construction module 200, graph comparison learning and training module 300, dynamic prototype correction module 400, anomaly detection module 500, store strategy parameterization module 600, and store strategy clustering analysis module 700. Detailed Implementation

[0029] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0030] In the description of this invention, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0031] like Figure 1As shown, this embodiment of the invention provides an AI-driven industry user reputation collection and intelligent public opinion analysis system, including a data collection module 100, a normal pattern graph construction module 200, a graph comparison learning and training module 300, a dynamic prototype correction module 400, and an anomaly detection module 500.

[0032] The data collection module 100 collects reputation data of target entities (such as stores under a chain brand) on various platforms in real time through open API interfaces provided by various social media platforms (such as Weibo Open Platform API, Douyin Open Platform API, Xiaohongshu Open Platform API, etc.) or through compliant web crawling methods. The collected reputation data includes at least: text content (such as review text, comments), publication time, publishing account information, interaction data (number of likes, comments, reposts, and favorites), and URLs of images attached to the content. The preferred collection frequency is once every 5-10 minutes to ensure the real-time nature of the data.

[0033] The Normal Pattern Graph Construction Module 200 is used to extract normal dissemination patterns from historical word-of-mouth data. This module first obtains the historical word-of-mouth data of the target object during normal operation (i.e., a period without major marketing activities or public opinion crisis events, preferably the past 30-90 days), and then extracts three types of nodes (content nodes, user nodes, and dissemination path nodes) from it. Using the publishing relationship, forwarding relationship, and interaction relationship between each node as edges, a normal dissemination pattern graph is constructed.

[0034] The graph contrast learning training module 300 is used to perform graph contrast learning training on the above-mentioned normal propagation pattern graph, mapping the content nodes of the normal propagation pattern to the normal clustering regions in the feature space, and obtaining the trained graph neural network model and the feature space cluster centers of the normal propagation pattern.

[0035] The Dynamic Prototype Correction Module 400 is used to perform online dynamic fine-tuning of the cluster centers in the feature space, enabling normal standards to be adaptively updated as brand strategies evolve and new platforms are accessed.

[0036] The anomaly detection module 500 is used to input the real-time collected target word-of-mouth content into the trained graph neural network model, and to determine whether there is an anomaly in the propagation path by calculating the distance between the target word-of-mouth content and the current cluster center in the feature space.

[0037] In this embodiment, the reputation data collected in real time by the data acquisition module 100 is sent to the normal pattern map construction module 200 for constructing and updating the normal propagation pattern map, and to the anomaly detection module 500 for real-time anomaly determination. The feature space cluster centers output by the graph comparison learning training module 300 are simultaneously transmitted to the dynamic prototype correction module 400 and the anomaly detection module 500. The dynamic prototype correction module 400 synchronizes the updated cluster centers to the anomaly detection module 500 in real time.

[0038] Reference Figure 2 In some embodiments, the system of this invention further includes a store strategy parameterization module 600 and a store strategy clustering analysis module 700.

[0039] In some embodiments, the data acquisition module 100 is implemented as follows:

[0040] Data Acquisition: For chain brand scenarios, the data acquisition module 100 first obtains a list of official social media accounts of each store under the brand (including the stores' Xiaohongshu accounts, Douyin accounts, Dianping accounts, etc.), and then periodically crawls all content posted by these accounts and their interaction data through the APIs of each platform. For review platforms such as Dianping, it also collects the overall rating of the store, ratings of various dimensions (such as environment, service, taste, etc.), and text content of user reviews.

[0041] In a preferred embodiment, the data acquisition module 100 also collects publicly available reputation data of other similar competing stores in the same area as a supplementary source of negative samples in subsequent graph comparison learning training.

[0042] Data preprocessing: The collected raw word-of-mouth data undergoes the following preprocessing steps:

[0043] (1) Text cleaning: Remove HTML tags, special characters, emoticons (retain emojis with semantic value), and redundant whitespace, and convert the text to lowercase (Chinese characters do not need to be converted);

[0044] (2) Text semantic feature extraction: A pre-trained Chinese BERT model or a RoBERTa-wwm-ext model finely tuned for social media scenarios is used to transform the cleaned text content into a 768-dimensional semantic feature vector. For texts longer than 512 tokens, a truncation strategy is used to retain the first 512 tokens;

[0045] (3) Behavioral feature extraction: Extract the following behavioral features from the raw data: posting time period (dividing a day of 24 hours into 48 time periods, each time period is 30 minutes), interaction volume 1 hour after posting (number of likes + number of comments + number of reposts), interaction volume 24 hours after posting, interaction volume time decay curve parameters, number of pictures attached to the content, whether the content contains videos, etc.

[0046] (4) Feature fusion: The text semantic feature vector (768-dimensional) and the behavioral feature vector (preferably 32-dimensional, including the encoding of the publication time period, the logarithm of the interaction, the number of images, etc.) are concatenated to obtain the initial feature vector (800-dimensional) of the content node. This feature vector will be used as the input feature of the graph neural network.

[0047] (5) User feature extraction: For each publishing user, extract features such as historical publishing frequency (average daily publishing volume in the past 30 days), account registration duration, number of fans, number of followers, and average interaction rate of historical content to form the initial feature vector of the user node (preferably 16-dimensional).

[0048] In some embodiments, the normal pattern map construction module 200 is implemented as follows:

[0049] Normal Period Determination: The system first needs to determine the normal operating period. In a preferred embodiment, the system automatically determines the normal period by: (1) excluding dates when the brand has large-scale marketing activities (such as new product launches, anniversary celebrations, Double Eleven, etc.), which can be pre-marked by the brand headquarters; (2) excluding dates when public opinion crisis events occur, which can be identified by monitoring the outbreak of negative public opinion related to brand keywords; and (3) excluding dates when the platform itself is abnormal (such as large-scale platform failures, algorithm adjustment periods). After the above exclusions, the reputation data of the remaining periods are regarded as data during the normal operating period.

[0050] Node extraction:

[0051] (1) Content Node: Each piece of word-of-mouth content corresponds to a content node. Each content node stores the following attributes: content ID (platform unique identifier), text semantic feature vector (768 dimensions), publication time feature (48-dimensional one-hot encoding or 2-dimensional sine and cosine encoding), number of images, whether it contains video, interaction volume in 1 hour after publication, interaction volume in 24 hours after publication, platform identifier, and store identifier. In a preferred embodiment, the feature vector dimension of the content node is 800 dimensions (768-dimensional semantic features + 32-dimensional behavioral features).

[0052] (2) User Nodes: Each user who posts or interacts corresponds to a user node. Each user node stores the following attributes: user ID, historical posting frequency (average daily postings over the past 30 days), account registration days, number of followers, number of followings, historical average interaction rate, and user type tags (such as ordinary user, KOC, KOL, enterprise account, etc., which are automatically determined through account authentication information and historical posting patterns). The feature vector of a user node is preferably 16-dimensional.

[0053] (3) Propagation Path Nodes: Each piece of word-of-mouth content corresponds to a propagation path node as the complete propagation chain from its initial publication to subsequent interaction nodes (forwards, comments, replies). The propagation path node is constructed as follows: with the source content as the root node and all direct and indirect forwarding, commenting, and replying relationships as edges, a propagation tree is constructed; then the propagation tree is encoded, and the following propagation path features are extracted: propagation depth (length of the longest chain), propagation breadth (number of nodes in the first layer), propagation time span (time difference from the first post to the last interaction), interaction density distribution in each time period, and fractal dimension of the propagation tree, etc. The feature vector of the propagation path node is preferably 32-dimensional.

[0054] Edge construction: Construct the following types of edges between the three types of nodes mentioned above:

[0055] (1) Publish relationship edge: connects the user node and the content node it publishes, representing "who published what", and the direction of the edge is from the user node to the content node;

[0056] (2) Forwarding relationship edge: connects two content nodes, indicating that one piece of content is forwarded to another piece of content, and the direction of the edge is from the original content to the forwarded content;

[0057] (3) Comment relationship edge: connects two content nodes or connects a user node and a content node to represent a comment relationship;

[0058] (4) Interactive relationship edge: connects user nodes and content nodes, representing interactive behaviors such as liking and collecting;

[0059] (5) Propagation membership edge: connects the propagation path node with its corresponding content node and the user node that participated in the propagation path, indicating "the content belongs to the propagation path" and "the user participated in the propagation path".

[0060] By constructing the nodes and edges described above, a normal propagation pattern graph is obtained. , where V is the set of nodes (including content nodes, user nodes, and propagation path nodes), and E is the set of edges (including publishing relationships, forwarding relationships, comment relationships, interaction relationships, and propagation membership relationships).

[0061] In some embodiments, the graph comparison learning training module 300 is implemented as follows:

[0062] This embodiment uses a Graph Attention Network (GAT) as the graph neural network encoder. GAT assigns different weights to neighboring nodes through an attention mechanism, adaptively highlighting the core interactions in the graph. The GAT encoder preferably has two layers, with each layer having a hidden layer dimension of 256 and an output dimension (i.e., the final feature embedding dimension) of 128. Each GAT layer is followed by a ReLU activation function and a Dropout layer (preferably with a dropout rate of 0.2).

[0063] In an alternative embodiment, a graph convolutional network (GCN) can also be used as the encoder. GCN aggregates node features through spectral convolution of the adjacency matrix, which is more computationally efficient and suitable for scenarios with large graph scales.

[0064] Positive sample pair construction: Positive sample pairs are constructed from the normal propagation pattern map. Positive sample pairs include the following two categories:

[0065] (1) Positive sample pair under the same propagation mode: Two content nodes that are related to each other under the same normal propagation mode. Specifically, if two content nodes are directly connected through forwarding or commenting relationships, and they belong to the same propagation path node (i.e. connected to the same propagation path node through the propagation membership edge), then they constitute a positive sample pair.

[0066] (2) Cross-platform positive sample pairs: copies of the same content on different platforms. For example, a store posts a note on Xiaohongshu and a short video version of the same content on Douyin. Although these two content nodes are on different platforms, they belong to different carriers of the same marketing content and constitute a positive sample pair. The identification of cross-platform positive sample pairs is achieved by matching the similarity of content fingerprints (such as SimHash of text and perceptual hash of images). When the similarity of the content fingerprints of two content nodes exceeds a preset threshold (preferably 0.85), they are determined to be copies of the same content on different platforms.

[0067] In a preferred embodiment, it further includes: (3) temporal positive sample pairs: the feature representations of the same content node in different time windows are used as positive sample pairs. Specifically, for the same content node, its feature vector is extracted 1 hour, 6 hours and 24 hours after its release (the feature vectors in different time windows will be different due to the accumulation of interactive data), and the feature representations of the same content node in different time windows are used as positive sample pairs to enhance the model's ability to learn temporal propagation patterns.

[0068] Negative sample pair construction: Negative sample pairs consist of pairing normal content nodes with known abnormal content nodes. Known abnormal content nodes include: (1) content that has historically been marked as illegal or false by the platform; (2) illegal content manually marked by the brand headquarters; and (3) content with significantly abnormal dissemination patterns in competitor stores (as external negative samples). In a preferred embodiment, the number of negative sample pairs is 3-5 times the number of positive sample pairs to ensure sufficient comparative learning.

[0069] The learning and training process is compared with the diagram:

[0070] Mapping the normal transmission pattern Input the GAT encoder to obtain the feature embedding of each node. ,in .

[0071] For each anchor node Its positive samples are (and (Forming positive sample pairs), the negative sample set consists of all other nodes k in the batch ( (and k and i do not form a positive sample pair).

[0072] The following contrastive loss function (InfoNCE loss) is used for training:

[0073] ;

[0074] in, and 2N represents the feature embedding of a pair of positive samples, and 2N represents the total number of samples in a training batch (each original data corresponds to two augmented views, so the total number of samples in the batch is twice the number of original samples). For indicator functions, when The value is 1 when k=i and 0 when k=i, used to exclude the anchor point itself when summing the denominators (to avoid...). (leading to training collapse) Let τ be the cosine similarity function, and τ be the temperature parameter (preferably 0.07-0.1). This loss function is calculated independently for each anchor point i in each training iteration, and the loss value is the InfoNCE loss for a single sample.

[0075] In one specific embodiment, the batch size 2N is preferably 256 or 512 (i.e., each batch contains 128 or 256 original samples and their corresponding augmented views), and the number of positive sample pairs in each batch is approximately 1 / 4 to 1 / 3 of the batch size (i.e., approximately 64-128 pairs). Training uses the Adam optimizer with an initial learning rate of 3×10−4 / 3×10−4, and the number of training epochs is preferably 50-100 epochs. The patience value of the early stopping strategy is set to 10 epochs.

[0076] Cluster center calculation: After training, calculate the cluster centers in the feature space of the normal propagation pattern.

[0077] ;

[0078] in, This is the set of all content nodes in the normal propagation pattern graph. The feature embedding of node v is obtained by the trained GAT encoder.

[0079] The cluster center This represents the central position of the brand's normal communication model in the feature space.

[0080] During training, through comparative learning, content nodes exhibiting normal propagation patterns are mapped to normal clustering regions (gathered around cluster center C) in the feature space, while content nodes exhibiting abnormal propagation patterns are pushed away from cluster center C. This paradigm of normal baseline modeling combined with distance measurement enables the model to determine whether there are abnormal propagation paths by calculating the distance between new content and the cluster center.

[0081] In some embodiments, the dynamic prototype correction module 400 is implemented as follows:

[0082] In actual brand operations, normal communication patterns are not static. Brand strategies may be adjusted (such as shifting from high-end professionalism to approachable and energetic), new stores may open (stores in different regions may have different communication styles), and new social media platforms may be integrated (different platforms have different communication patterns). If the core element C remains fixed, over time, content that was originally normal may be misjudged as abnormal, while truly abnormal content may be missed because the core element becomes outdated.

[0083] Online calibration mechanism: The dynamic prototype correction module 400 performs online fine-tuning of the cluster center in the following ways:

[0084] When a new, normally propagated sample is acquired (i.e., new content that has been manually confirmed or automatically determined to be normal by the system), the feature embedding of that sample is calculated. With the current cluster center offset vector Then, the offset vectors are weighted and superimposed onto the current cluster center using a preset learning rate η:

[0085] ;

[0086] The learning rate η ranges from 0.01 to 0.1, preferably 0.05. A smaller learning rate ensures that the cluster center updates are smooth and gradual, avoiding drastic fluctuations in cluster centers due to individual noisy samples.

[0087] In a preferred embodiment, a momentum update method is used instead of a simple linear update:

[0088] ;

[0089] Wherein, α is the momentum parameter (with a value range of 0.9-0.99, preferably 0.95). This is a newly added batch of samples from normal transmission.

[0090] The momentum update method can better smooth out random fluctuations in the data stream, making the evolution of cluster centers more stable.

[0091] Sliding window mechanism: The dynamic prototype correction module 400 also employs a sliding window mechanism to limit the time range of samples participating in cluster center calculation. Specifically, the system maintains a time window of length W (preferably 30 days), and only normally propagated samples falling within this time window are used for cluster center calculation. Older samples exceeding the time window are gradually eliminated.

[0092] In one specific embodiment, the sliding window is updated as follows: Every morning at midnight, the system automatically checks the start and end dates of the current time window, removes old samples that have exceeded the window from the cluster center calculation set, and adds the newly confirmed normal samples from the previous day to the set, then recalculates the cluster centers. The window size W can be adaptively set according to the frequency of brand strategy adjustments: brands with frequent strategy adjustments (such as the FMCG industry) can set a shorter window (such as 14 days), while brands with relatively stable strategies (such as high-end luxury goods) can set a longer window (such as 60 days).

[0093] Brand Strategy Change Response: When a major adjustment to the brand strategy occurs (such as brand refresh, change of spokesperson, adjustment of target customer group, etc.), the brand headquarters can manually trigger the cluster center reset operation through the system backend. At this time, the system will recalculate the cluster centers based on the new normal communication samples generated after the brand strategy adjustment, realizing the rapid migration of cluster centers without retraining the entire graph neural network model.

[0094] In some embodiments, the anomaly detection module 500 is implemented as follows:

[0095] Real-time anomaly detection: When the data acquisition module 100 collects a new piece of word-of-mouth content, the anomaly detection module 500 performs the following steps:

[0096] (1) Perform the same preprocessing and feature extraction on the content as on the training data to obtain its initial feature vector;

[0097] (2) Input the initial feature vector into the trained GAT encoder to obtain the feature embedding of the content. ;

[0098] (3) Calculation Cosine distance from the current cluster center C:

[0099] ;

[0100] The value of cosine distance ranges from 0 to 2, where 0 represents completely identical, 1 represents orthogonal (unrelated), and 2 represents completely opposite.

[0101] (4) Compare the calculated distance d with the preset threshold θ: if d>θ, it is determined that the propagation path of the target word-of-mouth content is abnormal; otherwise, it is determined to be normal.

[0102] Anomaly type diagnosis: When it is determined that the propagation path of the target word-of-mouth content is abnormal, the anomaly detection module 500 further performs anomaly type diagnosis. Specifically, the system maintains multiple known anomaly type cluster centers (obtained by clustering historical anomaly samples), including: (1) Abnormal propagation by brushing traffic and water army: manifested as a large number of similar content being released in a short period of time, and the interactive accounts are mostly newly registered or low-activity accounts; (2) Abnormal secondary propagation of content tampering: manifested as normal content being modified and re-released, and the modified part usually involves the replacement of brand keywords or the addition of sensitive information; (3) Abnormal cross-platform propagation: manifested as the same content having significantly different propagation patterns on different platforms (such as being normal on Xiaohongshu but abnormally active on Douyin).

[0103] The method for anomaly type diagnosis is: calculating the target content feature embedding. Clustering centers for each anomaly type The distance (t=1,2,3) is used to select the abnormal type that is closest to the diagnosis result.

[0104] In some embodiments, the store strategy parameterization module 600 and the store-specific cluster core are generated in the following ways:

[0105] Within chain brands, different stores often play different market roles. Flagship stores focus on brand image building, community stores on convenient services, and mall stores on experiential marketing. The content tone, release schedule, and interaction methods of different stores vary significantly. Mapping all store reputation data to the same normal communication model can lead to flagship stores' high-quality creative content being misjudged as abnormal, while community stores' regular content is judged as acceptable.

[0106] Store strategy parameterization: The store strategy parameterization module 600 acquires the brand strategy configuration information of each store and transforms it into a store strategy feature vector. .

[0107] Brand strategy configuration information should include at least:

[0108] (1) Store type labels: flagship store, community store, shopping mall store, experience store, etc., using one-hot encoding (4-dimensional);

[0109] (2) Content tone preference tags: formal / professional, friendly / warm, lively / fun, high-end / luxury, etc., using multi-tag encoding (4 dimensions, each dimension is a continuous value of 0-1, representing the degree of preference for the tone).

[0110] (3) Publishing strategy tags: high frequency publishing type (≥3 items per day), medium frequency publishing type (1-2 items per day), low frequency publishing type (<1 item per day), using one-hot encoding (3-dimensional);

[0111] (4) Target audience characteristics: main audience age group, gender preference, etc. (5-dimensional continuous value).

[0112] The above features are concatenated to obtain the store strategy feature vector. .

[0113] Store-specific cluster core generation: The store-specific cluster core is generated by using the feature space cluster core of the global normal propagation mode. Using the baseline cluster center, the store strategy feature vector is... Input a bias generator network The network outputs a personalized bias vector for each store. Then, the baseline cluster center is added to the store-specific bias vector to obtain the store-specific cluster center:

[0114] ;

[0115] Bias generation It is a two-layer Multilayer Perceptron (MLP) with an input dimension of 16 (store strategy feature vector dimension), a hidden layer dimension of 64, and an output dimension of 128 (consistent with the cluster center dimension), using ReLU as the activation function. The training objective of the biased generative network is to minimize the distance between the normal sample feature embeddings of the store and the store-specific cluster centers.

[0116] ;

[0117] in, This is the normal propagation sample set for store i.

[0118] Store strategy change response: When the brand strategy configuration information of a store changes (such as upgrading from a community store to a flagship store), the store strategy parameterization module 600 recalculates the strategy feature vector of the store. Then, a bias-generated network is generated. Recalculate the store-specific bias vector This will update the store's exclusive cluster. The entire process does not require retraining the graph neural network model; it can be completed with only one forward propagation of the MLP, resulting in minimal computational overhead.

[0119] In some embodiments, the method for setting the dynamic threshold for store hierarchy is as follows:

[0120] Differentiated threshold settings: Different stores are configured with different anomaly detection thresholds. The hierarchical dynamic thresholds for stores are set based on a combination of the store's historical anomaly detection rate, false alarm rate, and store strategy feature vector.

[0121] In one specific embodiment, the anomaly detection threshold for store i Calculated in the following way:

[0122] ;

[0123] in, As the baseline threshold, Rate the aggressiveness of the store strategy (values ​​range from -1 to 1, derived from the store strategy feature vector). (obtained through a linear mapping), where λ is the adjustment coefficient (preferably 0.1-0.2).

[0124] For stores with aggressive strategies (such as flagship stores or new store opening periods). A positive value indicates a higher tolerance threshold for anomalies, allowing for greater creative deviations; this is suitable for stores with conservative strategies (such as standard community stores). A negative value indicates a lower tolerance threshold for anomalies, requiring stricter adherence to brand guidelines.

[0125] In some embodiments, the cross-store propagation pattern migration detection and store strategy clustering analysis module 700 is implemented as follows:

[0126] Cross-store propagation pattern migration detection: When the target store's word-of-mouth content is judged to be normal, the anomaly detection module 500 further calculates the average distance between the feature embedding of the target word-of-mouth content and the store-specific cluster centers of other stores in the same store type as the target store:

[0127] ;

[0128] in, For other stores of the same type as store i, This is a dedicated cluster for store j. If Exceeding the preset threshold for migration of the same type If this occurs, it is determined that the target's word-of-mouth content exhibits an abnormal cross-store dissemination pattern, and a cross-store consistency warning is issued. This situation may indicate that the store's dissemination pattern is deviating from the overall pattern of similar stores, requiring attention from headquarters.

[0129] Store Strategy Cluster Analysis: The store strategy cluster analysis module generates store-specific clusters for all 700 stores. Cluster analysis (using the K-means algorithm, where K is the number of store types) was performed to identify abnormal stores whose propagation patterns significantly deviated from their respective store type clusters.

[0130] Specifically, for each store type t, calculate the mean of the cluster centers of all stores of that type. (i.e., the cluster center of this type), and then calculate the cluster center for each store i. and its type cluster center Distance:

[0131] ;

[0132] when Exceeding the store-level deviation threshold When this happens, the store is marked as a strategy deviation store, and a store propagation pattern deviation report is generated. The method for determining this type is: take all stores of this type. The mean of the values ​​plus twice the standard deviation.

[0133] In some embodiments, the deployment architecture of the system according to this invention is as follows:

[0134] The system is deployed as a cloud-edge collaborative architecture: (1) Cloud server: Deploy graph comparison learning training module 300 to perform offline training and periodic retraining of graph neural networks (preferably once a month). The cloud server is equipped with GPU to accelerate model training. The cloud server also stores historical graph data of the entire brand for regular retraining and version iteration of the model. (2) Edge node: Deploy normal mode graph construction module 200, dynamic prototype correction module 400 and anomaly detection module 500 to perform real-time graph construction, online cluster center calibration and anomaly detection. The edge node can be a regional server (one deployed in each region, serving 10-50 stores in the region) or a store-level local server (one lightweight device deployed in each store). The edge node is equipped with a CPU (2-4 cores) and 4-8GB of memory to meet the computing requirements of real-time detection. (3) Data transmission: Only cluster center parameters and anomaly detection results are transmitted between the edge node and the cloud server. The original reputation data is not transmitted to protect user privacy and brand data security.

[0135] Specifically, edge nodes periodically (e.g., daily) upload their updated cluster core parameters to the cloud, and the cloud then distributes the globally updated model parameters and cluster cores to each edge node. Anomaly detection results (including anomaly content ID, anomaly type, anomaly score, etc.) are reported to the cloud in real time for the brand headquarters to view and perform statistical analysis. In one specific embodiment, the real-time anomaly detection latency of the edge nodes is controlled within 500 milliseconds, which can meet the immediate feedback needs of store staff after publishing content.

[0136] Reference Figure 3 The method proposed in this invention is an AI-driven method for collecting industry user reviews and analyzing intelligent public opinion, including the following steps:

[0137] S1. Collect historical reputation data of the target object during normal operation, extract content nodes, user nodes and propagation path nodes, and construct a normal propagation pattern map;

[0138] S2. Construct positive and negative sample pairs from the normal propagation pattern graph, train the graph neural network through graph contrast learning, map the content nodes of the normal propagation pattern to the normal clustering region in the feature space, and obtain the trained graph neural network model and the feature space cluster center of the normal propagation pattern.

[0139] S3. In response to brand strategy adjustments or new platform access events, acquire newly added normal dissemination samples, dynamically adjust the cluster centers of the feature space through online calibration, and limit the time range of samples participating in cluster center calculation using a sliding window mechanism.

[0140] S4. Collect target word-of-mouth content in real time, input the target word-of-mouth content into the trained graph neural network model, obtain its feature embedding in the feature space, calculate the distance between the feature embedding and the current cluster center, and when the distance exceeds the preset threshold, determine that the propagation path of the target word-of-mouth content is abnormal and output the abnormality detection result.

[0141] The following describes the complete workflow of an embodiment of the present invention using a specific application scenario.

[0142] Scenario: A chain milk tea brand has 200 stores and uses this system to collect and analyze social media reputation and public opinion of each store.

[0143] Initialization Phase: Data Acquisition Module 100 collects reputation data from Xiaohongshu, Douyin, and Dianping for each store over the past 90 days, totaling approximately 54,000 pieces of content (200 stores × 90 days × 3 pieces / day). Normal Pattern Graph Construction Module 200 extracts content nodes, user nodes, and propagation path nodes to construct a normal propagation pattern graph. This graph contains approximately 54,000 content nodes, approximately 150,000 user nodes, and approximately 54,000 propagation path nodes, with a total of approximately 800,000 edges.

[0144] The graph comparison learning training module 300 was trained on the graph on a cloud GPU server. After 80 iterations, it converged, yielding the GAT encoder and global cluster centers. After training, the average distance between normal samples and cluster centers was 0.38, the standard deviation was 0.12, and the threshold was... Set to 0.62.

[0145] The store strategy parameterization module 600 generates store strategy feature vectors based on the type labels of each store (20 flagship stores, 150 community stores, and 30 shopping mall stores), trains a bias generation network, and obtains the store-specific cluster center for each store. .

[0146] During the routine operation phase (after the system went live): An employee of a community store posted a Xiaohongshu note promoting a new product at 3 pm (the store's normal posting time).

[0147] The data acquisition module 100 acquires this content in real time, and the anomaly detection module 500 extracts its features and inputs them into the GAT encoder to obtain the feature embedding. Calculate the cosine distance d=0.35 between it and the exclusive cluster center of the community store.

[0148] because (The threshold for community stores is low), and the system determines that the content propagation path is normal, with no abnormal warnings.

[0149] Another flagship store posted a creative short video at 2 a.m. (not during the store's normal posting hours; flagship stores usually post on weekday afternoons).

[0150] The anomaly detection module 500 calculated d=0.72, because... (The flagship store has a higher threshold) The system judges it as normal (the creative content of the flagship store is allowed to deviate to a greater extent).

[0151] However, if this content occurs in a community store, If it does, it will be judged as abnormal and an alert will be triggered.

[0152] Suppose a community store publishes 50 similar positive reviews between midnight and 4 a.m. (the copywriting templates are highly consistent, and the accounts that publish them are mostly newly registered users). The anomaly detection module 500 detects that the average distance between the feature embedding of this batch of content and the cluster center of the community store is 0.85, which is far beyond the threshold of 0.58. The system judges it as an abnormal spread of fake traffic and automatically generates an anomaly report and pushes it to the brand headquarters and the regional manager.

[0153] After reviewing the anomaly report, the brand headquarters confirmed that the store had engaged in review manipulation and ordered rectification. Simultaneously, the system added this batch of abnormal content to the negative sample library for use in the next round of model training.

[0154] Strategy Adjustment Phase: Three months later, the chain brand underwent a refresh and upgrade, shifting from a more affordable to a higher-end positioning, resulting in a significant adjustment to its brand strategy. The brand headquarters triggered a cluster center reset in the system backend. Based on the newly generated normal dissemination samples after the strategy adjustment, the system recalculated the cluster centers using the Dynamic Prototype Correction Module 400. After two weeks of gradual updates, the cluster centers smoothly migrated to their new normal positions. During this period, content published by stores according to the new strategy was gradually incorporated into the normal sample set, and the cluster centers gradually shifted as new samples accumulated, while content from the old strategy was gradually phased out via a sliding window. Throughout the migration process, the system did not require retraining the GAT encoder; only updating the cluster center parameters was needed to complete the adaptation.

[0155] Although specific embodiments are described herein, those skilled in the art will recognize that many other modifications or alternative embodiments are also within the scope of this disclosure. For example, any of the functions and / or processing capabilities described in connection with a particular device or component can be performed by any other device or component. Furthermore, while various exemplary embodiments and architectures have been described according to embodiments of this disclosure, those skilled in the art will recognize that many other modifications to the exemplary embodiments and architectures described herein are also within the scope of this disclosure.

[0156] The foregoing description, with reference to block diagrams and flowcharts of systems, methods, systems, and / or computer program products according to exemplary embodiments, has described certain aspects of this disclosure. It should be understood that one or more blocks in the block diagrams and flowcharts, as well as combinations of blocks in the block diagrams and flowcharts, can be implemented by executing computer-executable program instructions, respectively. Similarly, according to some embodiments, some blocks in the block diagrams and flowcharts may not need to be executed in the order shown, or may not all need to be executed. Furthermore, additional components and / or operations beyond those shown in the blocks in the block diagrams and flowcharts may exist in some embodiments.

[0157] Therefore, blocks in block diagrams and flowcharts support combinations of means for performing a specified function, combinations of elements or steps for performing a specified function, and program instruction means for performing a specified function. It should also be understood that each block in a block diagram and flowchart, and combinations of blocks in block diagrams and flowcharts, can be implemented by a dedicated hardware computer system or a combination of dedicated hardware and computer instructions that performs a specific function, element, or step.

[0158] The program modules, applications, etc., described herein may include one or more software components, including, for example, software objects, methods, data structures, etc. Each such software component may include computer-executable instructions that, in response to execution, cause at least a portion of the functionality described herein (e.g., one or more operations of the exemplary methods described herein) to be performed.

[0159] Software components can be coded using any of a variety of programming languages. An exemplary programming language could be a low-level programming language, such as assembly language associated with a specific hardware architecture and / or operating system platform. Software components including assembly language instructions may need to be converted into executable machine code by an assembler before being executed by the hardware architecture and / or platform. Another exemplary programming language could be a higher-level programming language that is portable across multiple architectures. Software components including higher-level programming languages ​​may need to be converted into an intermediate representation by an interpreter or compiler before execution. Other examples of programming languages ​​include, but are not limited to, macro languages, shell or command languages, job control languages, scripting languages, database query or search languages, or report writing languages. In one or more exemplary embodiments, a software component containing instructions from one of the above-described programming language examples can be executed directly by the operating system or other software components without first being converted into another form.

[0160] Software components can be stored as files or other data storage structures. Software components of similar type or related function can be stored together in a specific directory, folder, or library. Software components can be static (e.g., pre-defined or fixed) or dynamic (e.g., created or modified at runtime).

[0161] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. An AI-driven industry user reputation collection and intelligent public opinion analysis system, characterized in that, include: The data acquisition module is used to collect reputation data of target objects from multiple social media platforms. The reputation data includes at least text content and interaction data. The normal mode graph construction module is used to acquire historical reputation data of the target object during normal operation, extract content nodes, user nodes, and propagation path nodes, and construct a normal propagation mode graph with the publishing relationship, forwarding relationship, and interaction relationship between each node as edges. The content nodes store the textual semantic feature vector and publishing time period features of the historical reputation data, and the propagation path nodes store the feature representation of a propagation link between the initial publication of reputation content and each interactive node. Each propagation path node is generated by associating its corresponding content node with the user nodes participating in the propagation path. The graph contrastive learning training module is used to construct positive and negative sample pairs from the normal propagation pattern graph. The positive sample pairs include two related content nodes under the same normal propagation pattern or propagation copies of the same content on different platforms. The negative sample pairs include pairings of normal content nodes and known abnormal content nodes. The graph neural network is trained through graph contrastive learning, mapping the content nodes of the normal propagation pattern to normal clustering regions in the feature space, and mapping the content nodes of the abnormal propagation pattern to positions far away from the normal clustering regions, thereby obtaining the trained graph neural network model and the cluster centers of the feature space of the normal propagation pattern. The dynamic prototype correction module is used to respond to brand strategy adjustments or new platform access events, acquire new normal propagation samples, dynamically adjust the cluster centers of the feature space through online calibration, and limit the time range of samples participating in the cluster center calculation with a sliding window mechanism. An anomaly detection module is used to input the target word-of-mouth content collected in real time by the data acquisition module into the trained graph neural network model, obtain its feature embedding in the feature space, calculate the distance between the feature embedding and the current cluster center, and determine that the propagation path of the target word-of-mouth content is abnormal when the distance exceeds a preset threshold, and output the anomaly detection result.

2. The AI-driven industry user reputation collection and intelligent public opinion analysis system according to claim 1, characterized in that, The contrastive loss function used in the graph contrastive learning training module is: ； in, and For each anchor point i, the feature embedding of the two samples in the positive sample pair is uniquely determined in each training iteration; 2N is the total number of samples in a training batch, and each original data point is augmented to obtain two positive sample pairs. The cosine similarity function is τ, and the temperature parameter is τ. The construction of the positive sample pairs also includes: taking the feature representations of the same content node in different time windows as positive sample pairs to enhance the model's ability to learn temporal propagation patterns.

3. The AI-driven industry user reputation collection and intelligent public opinion analysis system according to claim 1, characterized in that, The online calibration method of the dynamic prototype correction module is as follows: when a new normal propagation sample is acquired, the offset vector between the feature embedding of the sample and the current cluster center is calculated, and the offset vector is weighted and superimposed on the current cluster center with a preset learning rate to achieve incremental update of the cluster center; the window size of the sliding window mechanism is adaptively set according to the brand strategy adjustment frequency.

4. The AI-driven industry user reputation collection and intelligent public opinion analysis system according to claim 1, characterized in that, The anomaly detection module is also used to further calculate the distance between the feature embedding of the target word-of-mouth content and the cluster centers of multiple known anomaly types when it is determined that there is an anomaly in the propagation path of the target word-of-mouth content, and output the anomaly diagnosis result according to the anomaly type with the closest distance; the anomaly types include at least the anomaly of brushing traffic and water army propagation, the anomaly of secondary propagation of content tampering, and the anomaly of inconsistent propagation across platforms.

5. The AI-driven industry user reputation collection and intelligent public opinion analysis system according to claim 1, characterized in that, The preset threshold is a multi-level dynamic threshold, which includes a strategy-differentiated threshold set according to store type and a regional threshold set according to region. Different store types have different anomaly tolerance levels set according to their brand strategy configuration information. The anomaly tolerance threshold for stores with aggressive strategies is higher than that for stores with conservative strategies.

6. The AI-driven industry user reputation collection and intelligent public opinion analysis system according to claim 1, characterized in that, The system is deployed in a cloud-edge collaborative architecture. The graph contrast learning training module is deployed on the cloud server to perform offline training and periodic retraining of the graph neural network. The normal pattern graph construction module, dynamic prototype correction module, and anomaly detection module are deployed on the edge nodes to perform real-time graph construction, online cluster center calibration, and anomaly detection. The edge nodes and the cloud server only transmit cluster center parameters and anomaly detection results, and do not transmit the original reputation data.

7. The AI-driven industry user reputation collection and intelligent public opinion analysis system according to claim 1, characterized in that, The system also includes a store strategy parameterization module, which is used to obtain the brand strategy configuration information of each store and convert the brand strategy configuration information into a store strategy feature vector. The brand strategy configuration information includes at least store type tags, content tone preference tags, and publishing strategy tags. The normal pattern graph construction module is also used to extract store-level normal propagation samples from the store’s historical word-of-mouth data during normal operation for each store, and construct a store-level normal propagation pattern sub-graph. The graph contrastive learning training module is also used to perform store-specific contrastive learning training in the feature space, using the content nodes in the store-level normal propagation pattern subgraph as positive samples for the store, and the normal content nodes of other stores and the global abnormal content nodes as negative samples, to obtain the store-specific cluster centers of each store in the feature space.

8. The AI-driven industry user reputation collection and intelligent public opinion analysis system according to claim 7, characterized in that, The method for generating the store-specific cluster center is to use the feature space cluster center of the global normal propagation mode as the reference cluster center, input the store strategy feature vector into a bias generation network, the bias generation network outputs the store-specific bias vector, and add the reference cluster center and the store-specific bias vector to obtain the store-specific cluster center. The bias generation network is a multilayer perceptron network, and its training objective is to minimize the distance between the normal sample feature embedding of the store and the store-specific cluster center. The dynamic prototype correction module is also used to recalculate the store's strategy feature vector and update the store's store-specific cluster center through the bias generation network when the store's brand strategy configuration information changes.

9. The AI-driven industry user reputation collection and intelligent public opinion analysis system according to claim 7, characterized in that, The anomaly detection module is also used to calculate the average distance between the feature embedding of the target word-of-mouth content and the store-specific cluster center of other stores in the store type to which the target store belongs when the target word-of-mouth content is determined to be normal. If the average distance exceeds the preset same-type migration threshold, it is determined that the target word-of-mouth content has an abnormal cross-store propagation mode migration. The system also includes a store strategy clustering analysis module, which is used to perform clustering analysis on the store-specific cluster centers of all stores. When the distance between the store-specific cluster center of a store and the cluster center of its store type exceeds the store-level deviation threshold, the store is marked as a strategy deviation store and a store propagation pattern deviation report is generated.

10. An AI-driven method for collecting industry user reputation and analyzing intelligent public opinion, applied to the system described in any one of claims 1 to 9, characterized in that, Includes the following steps: S1. Collect historical reputation data of the target object during normal operation, extract content nodes, user nodes and propagation path nodes, and construct a normal propagation pattern map; S2. Construct positive sample pairs and negative sample pairs from the normal propagation pattern map, train the graph neural network through graph contrast learning, map the content nodes of the normal propagation pattern to the normal clustering region in the feature space, and obtain the trained graph neural network model and the feature space cluster center of the normal propagation pattern. S3. In response to brand strategy adjustments or new platform access events, acquire newly added normal propagation samples, dynamically adjust the cluster centers of the feature space through online calibration, and limit the time range of samples participating in cluster center calculation using a sliding window mechanism. S4. Collect target word-of-mouth content in real time, input the target word-of-mouth content into the trained graph neural network model, obtain its feature embedding in the feature space, calculate the distance between the feature embedding and the current cluster center, and when the distance exceeds a preset threshold, determine that the propagation path of the target word-of-mouth content is abnormal and output the abnormality detection result.

Images