Private domain community user implicit demand mining method based on multi-dimensional semantic fusion

Abstract

The application discloses a private domain community user implicit demand mining method based on multi-dimensional semantic fusion. The method comprises the following steps: acquiring multi-dimensional community data and user space-time behavior information, and using a lightweight multi-modal recognition model for preliminary demand recognition in the edge analysis layer; using a multi-source fusion model to comprehensively evaluate the preliminary results and original data; assigning tracking identifiers to the behavior logs and combining the space-time information to determine the continuous behavior sequence of the user, and obtaining a multi-dimensional health feature vector based on the evaluation; using a demand evolution prediction model to predict the life cycle state, demand decay trend and behavior health index of the user; determining the demand response level and triggering the corresponding operation strategy; adjusting the judgment threshold of the demand response level and updating the detection parameters of the edge analysis layer. Through the implementation of the method of the application, multi-modal information deep fusion, accurate user state prediction, operation strategy adaptive optimization and demand cause space-time tracing of community user implicit demand mining can be realized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, and more specifically to a method for mining the implicit needs of users in private communities based on multi-dimensional semantic fusion. Background Technology

[0002] With the development of digital marketing, private domain traffic operation has gradually become an important way for enterprises to establish deep connections with users. Especially in community environments, such as WeChat groups, enterprise WeChat groups, and Discord channels, enterprises can directly reach target users and achieve efficient user conversion. However, current community operation technologies have significant shortcomings in meeting the needs of modern enterprises.

[0003] First, existing community operation technologies primarily rely on keyword matching and basic text sentiment analysis to identify user needs, neglecting comprehensive analysis of image content such as product screenshots and usage scenario images, as well as behavioral logs such as clicks, dwell time, and forwarding behavior. This single-dimensional approach to need identification limits the ability to capture users' implicit needs, reducing user experience and satisfaction. Second, these technologies typically make decisions based on static rules or data from a single point in time, lacking effective predictive mechanisms for changes in user interests and the dynamic evolution of social relationships. Consequently, they cannot adjust operational strategies in a timely manner to adapt to users' latest needs.

[0004] Furthermore, the existing solution lacks an effective closed-loop feedback mechanism between the demand identification model, the operation strategy engine, and the data acquisition module, resulting in fixed model parameters that cannot be optimized or adjusted based on actual performance. In addition, when new implicit demands emerge within the community, current technology struggles to trace the source of these demands and their propagation path within the community, limiting the ability of operations personnel to take targeted measures.

[0005] Therefore, it is necessary to design a new method to achieve deep fusion of multimodal information, accurate prediction of user status, adaptive optimization of operational strategies, and spatiotemporal tracing of demand incentives to mine the implicit needs of community users, so as to solve the technical problems of insufficient multimodal fusion, lack of prediction of user status evolution, separation of operational strategies and model parameters, and inability to trace demand incentives in existing technologies. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for mining the implicit needs of users in private communities based on multi-dimensional semantic fusion.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for mining implicit needs of private community users based on multi-dimensional semantic fusion, applied to the cloud platform layer of a demand mining system. The demand mining system further includes a data acquisition layer, an edge analysis layer, and a cloud platform layer, comprising:

[0008] Acquire multi-dimensional community data and user spatiotemporal behavior information;

[0009] In the edge analysis layer, a preliminary implicit demand identification is performed on the multi-dimensional community data based on a preset lightweight multimodal recognition model to obtain preliminary demand identification results.

[0010] At the cloud platform layer, the preliminary demand identification results and the multi-dimensional community data are fused based on a preset multi-source fusion model to obtain a comprehensive demand evaluation result;

[0011] Assign tracking identifiers to user behavior paths in the behavior log data, and determine continuous behavior sequences for the same user behavior path by combining user spatiotemporal behavior information;

[0012] A multidimensional state evolution assessment is performed based on the continuous behavioral sequence to obtain a multidimensional health feature vector;

[0013] Based on the preset demand evolution prediction model, and using the continuous behavior sequence and the multidimensional health feature vector as input, the user life cycle status, demand decay trend and user behavior health index are predicted.

[0014] Based on the comprehensive demand evaluation results, the demand decay trend, the user behavior health index, and the user lifecycle status, the demand response level is determined, and the target operation strategy corresponding to the demand response level is triggered.

[0015] The system acquires feedback behavior data of the user terminal in response to the reach signal, dynamically adjusts the threshold for determining the demand response level based on the feedback behavior data, and sends the adjusted threshold as a demand detection parameter to the edge analysis layer.

[0016] This invention also provides a system for mining the implicit needs of users in private communities based on multi-dimensional semantic fusion, including:

[0017] The acquisition unit is used to acquire multi-dimensional community data and user spatiotemporal behavior information;

[0018] The preliminary identification unit is used to perform preliminary identification of implicit needs on the multi-dimensional community data at the edge analysis layer based on a preset lightweight multimodal identification model, and obtain preliminary needs identification results.

[0019] The fusion unit is used to fuse the preliminary demand identification results and the multi-dimensional community data based on a preset multi-source fusion model at the cloud platform layer to obtain a comprehensive demand evaluation result.

[0020] The sequence determination unit is used to assign tracking identifiers to user behavior paths in the behavior log data and determine continuous behavior sequences of the same user behavior path by combining user spatiotemporal behavior information.

[0021] An evaluation unit is used to perform multidimensional state evolution evaluation based on the continuous behavior sequence to obtain a multidimensional health feature vector;

[0022] The prediction unit is used to predict the user's life cycle status, demand decay trend and user behavior health index based on a preset demand evolution prediction model, taking the continuous behavior sequence and the multidimensional health feature vector as input.

[0023] The triggering unit is used to determine the demand response level based on the comprehensive demand evaluation result, the demand decay trend, the user behavior health index, and the user lifecycle status, and to trigger the target operation strategy corresponding to the demand response level.

[0024] The sending unit is used to acquire the feedback behavior data of the user terminal to the touch signal, dynamically adjust the judgment threshold of the demand response level according to the feedback behavior data, and send the adjusted judgment threshold as a demand detection parameter to the edge analysis layer.

[0025] The advantages of this invention compared to existing technologies are as follows: This invention acquires multi-dimensional community data, including text, images, and behavioral logs, at the data acquisition layer, and uses a lightweight multimodal recognition model at the edge analysis layer to perform preliminary demand identification to quickly capture users' potential needs. Then, at the cloud platform layer, a multi-source fusion model is used to deeply integrate the preliminary identification results with the original data to generate a comprehensive demand evaluation result. Next, by assigning tracking identifiers to user behavior paths and combining spatiotemporal information to evaluate multi-dimensional health feature vectors, a preset demand evolution prediction model is used to predict user lifecycle status, demand decay trends, and behavioral health indices, thereby achieving accurate user status prediction. Based on these prediction results, the demand response level is determined and corresponding operational strategies are triggered. At the same time, the threshold for judging the demand response level is dynamically adjusted based on user feedback to the reach signal. While optimizing the operational strategy, updated parameters are also sent to the edge analysis layer. This realizes a complete process from deep integration of multimodal information, accurate prediction of user status evolution, adaptive optimization of operational strategies, to spatiotemporal tracing of demand causes. This effectively solves the problems of insufficient multimodal fusion, lack of user status evolution prediction, separation of operational strategies and model parameters, and inability to trace demand causes in existing technologies.

[0026] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 A flowchart illustrating the method for mining the implicit needs of private community users based on multi-dimensional semantic fusion, as provided in an embodiment of the present invention.

[0029] Figure 2 A schematic block diagram of a system for mining the implicit needs of users in a private community based on multi-dimensional semantic fusion, provided in an embodiment of the present invention;

[0030] Figure 3 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0033] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0034] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0035] Please see Figure 1 , Figure 1This is a flowchart illustrating the method for mining implicit needs of private community users based on multi-dimensional semantic fusion, as provided in this embodiment of the invention. This method is applied to a server. It achieves deep fusion of multimodal information through a three-layer architecture: a data acquisition layer, an edge analysis layer, and a cloud platform layer. At the edge analysis layer, a lightweight multimodal recognition model is used to perform preliminary processing of community data, obtaining the results of identifying users' implicit needs. Subsequently, at the cloud platform layer, a multi-source fusion model is used to comprehensively evaluate the preliminary results combined with the original data. A continuous behavior sequence is constructed to assess the user's multi-dimensional health feature vector, thereby predicting the user's lifecycle status and demand change trends. Based on these analysis results, the demand response level is determined, and corresponding operational strategies are triggered. Furthermore, a feedback mechanism dynamically adjusts the judgment threshold and optimizes model parameters to ensure adaptive optimization of operational strategies. Finally, a demand causal analysis model is used to generate a demand tracing report, updating the implicit demand content and achieving spatiotemporal tracing of demand causes. This solves the problems of insufficient multimodal fusion, lack of user state evolution prediction, disconnect between operational strategies and model parameters, and inability to trace demand causes in existing technologies.

[0036] Figure 1 This is a flowchart illustrating the method for mining implicit needs of private community users based on multi-dimensional semantic fusion, as provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps S110 to S180.

[0037] S110: Obtain multi-dimensional community data and user spatiotemporal behavior information.

[0038] In this embodiment, the multi-dimensional community data includes user-generated text content, image content, and behavior log data, which are collected by the data acquisition layer.

[0039] User-generated text content: This includes all textual communication information users exchange within the community, such as posts, comments, chat logs, etc. This text data not only reflects users' interests and preferences but also reveals their emotional attitudes and social relationships.

[0040] Image content: Images uploaded or shared by community members are also an important source of data. By analyzing image content (such as usage scenarios and object recognition), we can further understand users' lifestyles, aesthetic preferences, and social activities.

[0041] Behavioral log data: This records users' behavioral patterns on social media platforms, such as clicks, views, likes, and shares. This type of data is crucial for analyzing user behavior patterns and changes in interests.

[0042] User spatiotemporal behavior information includes the location information of user nodes in the community channel, topic participation timestamps, and interaction link information.

[0043] The data acquisition layer is responsible for collecting the three types of data mentioned above from different data sources, ensuring data diversity and integrity. This step may involve interfacing with multiple data interfaces to cover as many user interaction scenarios as possible. Since directly collected data may contain noise or be incomplete, necessary cleaning and preprocessing are required. This includes removing duplicate data, filling in missing values, and standardizing data formats to ensure subsequent steps can analyze and utilize the data more accurately. Throughout the data acquisition process, relevant laws and regulations must be strictly followed to ensure full protection of user privacy. This means that anonymization, encryption, and other technical measures must be adopted when collecting and processing user data, and explicit user consent must be obtained.

[0044] By acquiring multi-dimensional community data and user spatiotemporal behavior information, this embodiment can construct a comprehensive and in-depth user profile, covering not only basic user attributes (such as age and gender) but also capturing their dynamic interests, social network structure, and behavioral habits. This is of great significance for accurately identifying users' implicit needs and providing personalized services. Furthermore, this comprehensive data analysis also provides rich material for subsequent steps, such as preliminary identification at the edge analysis layer and comprehensive evaluation at the cloud platform layer.

[0045] The data acquisition layer collects user-generated text content, image content, and behavior log data in real time through community open interfaces or SDK tracking. The text content is transmitted to the edge analysis layer and the cloud platform layer after being anonymized; the image content is transmitted after compression and encoding; and the behavior log data is transmitted in structured JSON format, including user ID, behavior type, behavior object, timestamp, and channel location information.

[0046] S120. In the edge analysis layer, based on a preset lightweight multimodal recognition model, the implicit needs of the multidimensional community data are initially identified to obtain preliminary needs identification results.

[0047] In this embodiment, the preliminary requirement identification results include requirement type, semantic matching degree, and identification confidence level.

[0048] Based on the comprehensive analysis of the aforementioned multimodal data, the edge analysis layer performs preliminary identification of implicit needs. The key to this stage lies in determining the basic attributes of user needs, including but not limited to need type (e.g., product recommendations, service feedback, social interactions), semantic matching degree (measuring the relevance between a user's expressed needs and their underlying needs), and identification confidence (representing the system's confidence in the identified need type). By calculating these indicators, the edge analysis layer can generate a preliminary need profile for each user, providing a foundation for further in-depth analysis.

[0049] The initial demand identification results not only lay the foundation for more accurate demand analysis later on, but also directly serve the real-time response mechanism. Specifically, demand type helps the system quickly locate the user's areas of interest, thereby guiding the delivery of personalized content or services; semantic matching degree is used to evaluate whether the currently provided solutions meet the user's actual needs, providing a basis for adjustment and optimization; and identification confidence reflects the reliability of the current identification results, helping decision-makers determine whether additional manual review or other forms of verification are needed. Furthermore, because the processing of the edge analysis layer occurs close to the data source, this distributed architecture effectively reduces latency and improves the system's real-time responsiveness. Therefore, the initial demand identification of the edge analysis layer plays an indispensable role in both improving user experience and enhancing enterprise competitiveness. It enables enterprises to more keenly capture market dynamics, respond promptly to changes in customer needs, and thus achieve precise marketing and service upgrades.

[0050] The preset lightweight multimodal recognition model includes:

[0051] A text encoding branch based on DistilBERT is used to encode the text content to obtain edge text features;

[0052] The image coding branch based on MobileNet-V3 is used to encode the image content to obtain edge image features;

[0053] A feature splicing layer is used to splice the edge text features and the edge image features;

[0054] The preliminary identification layer includes a first fully connected layer and a Softmax classifier, which are used to output the demand type and the identification confidence based on the concatenated features;

[0055] A similarity calculation layer is used to calculate the cosine similarity between the edge text features and the edge image features, and output the semantic matching degree.

[0056] In this embodiment, a preset lightweight multimodal recognition model is deployed in the edge analysis layer 20. This model includes a text encoding branch based on DistilBERT and an image encoding branch based on MobileNet-V3.

[0057] Specifically, the text encoding branch receives user-generated text content, extracts features through DistilBERT's 6-layer Transformer encoder, and outputs an edge text feature vector with a dimension of 768. Compared to standard BERT, DistilBERT reduces the number of parameters, improves inference speed, and is suitable for deployment on edge nodes.

[0058] The image encoding branch receives image content uploaded by the user and extracts features through the inverse residual bottleneck structure of MobileNet-V3 Large, outputting edge image feature vectors with a dimension of 1280. MobileNet-V3 is optimized through hardware-aware Neural Architecture Search (NAS) and the NetAdapt algorithm, which significantly reduces the computational cost while maintaining high accuracy.

[0059] The feature splicing layer splices the edge text feature vector (768-dimensional) and the edge image feature vector (1280-dimensional) into a 2048-dimensional joint feature vector.

[0060] The preliminary identification layer includes a first fully connected layer (2048-dimensional input, 256-dimensional output, ReLU activation function) and a Softmax classifier (output dimension is the preset number of demand types N), which is used to output the demand type and identification confidence based on the joint feature vector.

[0061] The similarity calculation layer calculates the cosine similarity between the edge text feature vector and the edge image feature vector, and outputs the semantic matching degree.

[0062] S130. At the cloud platform layer, the preliminary demand identification results and the multi-dimensional community data are fused based on a preset multi-source fusion model to obtain a comprehensive demand evaluation result.

[0063] In this embodiment, based on a preset multi-source fusion model, the preliminary demand identification result is used as the gating input and the multi-dimensional community data is used as the encoding input to perform multi-source weighted fusion and obtain a comprehensive demand evaluation result.

[0064] The comprehensive demand evaluation results include demand type, urgency score, and multimodal fusion confidence level;

[0065] The multi-source fusion model includes:

[0066] A text encoding sub-network based on RoBERTa is used to encode the text content to obtain text semantic features;

[0067] A Swing Transformer-based image coding sub-network is used to encode the image content to obtain image visual features;

[0068] A behavior encoding subnetwork based on a temporal convolutional network (TCN) is used to encode the behavior log data to obtain temporal behavior features.

[0069] A cross-modal attention fusion layer is used to perform multi-head cross-attention calculation with the text semantic features as the query vector and the image visual features and the behavior temporal features as key-value pairs, and output multi-modal interaction features; the image visual features and the behavior temporal features are projected or concatenated to a unified dimension and then used together as key-value pairs for multi-head cross-attention calculation.

[0070] The gated fusion layer is used to calculate dynamic fusion weights based on the identification confidence in the preliminary requirement identification results and the historical data quality labels of each data source, and to perform weighted fusion of the multimodal interaction features and the semantic matching degree in the preliminary requirement identification results using the dynamic fusion weights, and output the comprehensive requirement evaluation result.

[0071] In step S110, historical data quality labels of each of the data acquisition sources are obtained, and the historical data quality labels are used to represent the data quality error of each of the data acquisition sources.

[0072] The text encoding sub-network adopts a RoBERTa-base architecture (12-layer Transformer, 768 hidden dimensions) to perform deep semantic encoding on the text content, obtaining text semantic features T∈R. n×768 , where n is the sequence length.

[0073] The image coding subnetwork adopts the Swing Transformer Tiny architecture (window size 7×7, embedding dimension 96, depth [2,2,6,2]) to perform hierarchical visual coding on the image content, obtaining image visual features V∈R. m×768 , where m is the length of the image patch sequence. The Swin output is mapped to 768 dimensions through a linear projection layer to align with the text features.

[0074] The behavior encoding subnetwork employs a Temporal Convolutional Network (TCN), comprising four residual blocks. Each residual block contains two dilated causal convolutional layers (dilation factors of 1, 2, 4, and 8, respectively), with a kernel size of 3. This subnetwork performs temporal encoding on the behavior log data to obtain the behavior temporal feature B∈R. k×768 , where k is the time step.

[0075] The cross-modal cross-attention fusion layer uses the text semantic feature T as the query vector and the image visual feature V and the behavioral temporal feature B as key-value pairs for multi-head cross-attention calculation. Specifically, for the h-th attention head: ;in, d is a learnable projection matrix. k =64 dimensions per head. The representation is feature concatenation. The multi-head attention output is linearly projected to obtain the multimodal interaction feature F. cross ∈R n×768 .

[0076] The gated fusion layer calculates dynamic fusion weights based on the recognition confidence level *c* in the preliminary requirement identification results and the historical data quality labels of each data source. Let the historical accuracy rates of the text, image, and behavioral data sources be respectively... Then the confidence level of the data source is: Dynamic fusion weights Calculated using a Sigmoid gated network. Ultimately, the urgency score S in the comprehensive demand evaluation result is: ,in, For average pooling operation, Match represents the semantic matching degree.

[0077] S140. Assign tracking identifiers to user behavior paths in the behavior log data, and determine continuous behavior sequences of the same user behavior path by combining user spatiotemporal behavior information.

[0078] Specifically, a multidimensional state evolution assessment is performed based on the continuous behavior sequence to obtain a multidimensional health feature vector. The multidimensional state evolution assessment includes: calculating the contrast value between the user's active signal and the community background baseline to obtain an activity visibility score; detecting the skeleton continuity and edge roughness of the user's social network relationship to obtain a relationship integrity score; mapping the user's interest tag data to the semantic heat space and calculating the heat deviation to obtain an interest decay score; and using time-series curve fitting and residual analysis to calculate the degree of deviation of the user's behavior trajectory to obtain a behavior consistency score.

[0079] A unique tracking identifier (Track ID) is assigned to each user behavior path in the behavior log data. Combining user spatiotemporal behavior information (channel location information, topic participation timestamps), a nearest neighbor matching algorithm based on time windows is used for cross-time window tracking and location. Specifically, if the same user's behavior events within a time window are located in the same channel or related channels (such as sub-channels or referenced channels), and the time interval is less than a preset threshold (e.g., 30 minutes), then the user behavior paths are determined to be unified and merged into a continuous behavior sequence.

[0080] Specifically, the user's spatiotemporal behavior information includes the user node's location information in the community channel, i.e., channel location information, topic participation timestamps, and interaction link information.

[0081] S150. Based on the continuous behavior sequence, perform multidimensional state evolution evaluation to obtain a multidimensional health feature vector.

[0082] In this embodiment, the multidimensional health feature vector refers to

[0083] In one embodiment, step S150 described above may include steps S151 to S155.

[0084] S151. Extract user activity signals from the continuous behavior sequence, calculate the Weber contrast or Michelson contrast between the user activity signals and the community background baseline, and obtain the activity visibility score.

[0085] S152. Construct a social network graph based on the social relationship data in the continuous behavior sequence, extract the relationship skeleton, and calculate the skeleton continuity index and edge roughness index to obtain the relationship integrity score;

[0086] S153. Map the interest tag data in the continuous behavior sequence to a preset semantic popularity space, calculate the deviation between the current popularity value and the historical benchmark popularity value, and perform aging trend analysis in combination with the time decay function to obtain the interest decay score.

[0087] S154. Perform cubic spline curve fitting on the behavior trajectory points in the continuous behavior sequence, calculate the fitting residual and lateral deviation, and obtain the behavior consistency score.

[0088] S155. The activity visibility score, the relationship integrity score, the interest decay score, and the behavior consistency score are concatenated into the multidimensional health feature vector.

[0089] User activity signals (such as message sending frequency and online duration) are extracted from the continuous behavior sequence. A Michelson comparison is calculated between these signals and the community background baseline (such as the average activity level of the same channel during the same period). A social network graph G=(V,E) is constructed based on social relationship data (such as @ interactions, reply relationships, and mutual likes) in the continuous behavior sequence. The relationship skeleton is extracted using the Zhang-Suen refinement algorithm, and skeleton continuity indicators (skeleton pixel ratio) and edge roughness indicators (average gradient magnitude after Canny edge detection) are calculated. The relationship integrity score is defined as the weighted average of the inverses of the skeleton continuity indicator and the edge roughness indicator. User interest tag data (such as clicked product categories and search keywords) is mapped to a preset semantic heat space (a word vector space obtained through pre-trained Word2Vec or BERT embedding). The cosine similarity deviation between the current time window interest vector and the historical baseline heat vector is calculated. Combined with a time decay function, an interest decay score is obtained. Cubic spline curves are fitted to the behavioral trajectory points (such as daily active time points and consumption amount sequences) in the continuous behavior sequence to obtain the fitted curve S(t). Calculate the sum of squared residuals between each actual trajectory point and the fitted value, as well as the lateral deviation (time axis offset). The behavioral consistency score is defined as the reciprocal of the normalized residuals. The four scores are concatenated to form the multidimensional health feature vector.

[0090] Specifically,

[0091] Activity visibility score: Extract user activity signals from the continuous behavior sequence and calculate their contrast value with the community background baseline. Specifically, calculate the difference between the user activity signal and the background baseline, divide it by the sum of the two, and add a small constant to prevent division by zero to obtain the activity visibility score.

[0092] Relationship integrity score: A social network graph is constructed based on the social relationship data in the continuous behavior sequence. The relationship skeleton is extracted using a refinement algorithm, and the skeleton pixel ratio is calculated as the skeleton continuity index. The average gradient magnitude is calculated using an edge detection algorithm as the edge roughness index. The relationship integrity score is obtained by a weighted average of the skeleton continuity index and the inverse of the edge roughness index.

[0093] Interest Decline Score: This function maps user interest tag data to a predefined semantic popularity space. It calculates the cosine similarity deviation between the current time window's interest vector and the historical baseline popularity vector, and then combines this with a time decay function to analyze aging trends, thus obtaining the interest decline score.

[0094] Behavioral consistency score: A cubic spline curve is fitted to the behavioral trajectory points in the continuous behavioral sequence to obtain the fitted curve. The sum of squared residuals between each actual trajectory point and the fitted value, as well as the lateral deviation, are calculated. The behavioral consistency score is obtained from the reciprocal of the normalized residuals.

[0095] Finally, the activity visibility score, relationship integrity score, interest decay score, and behavior consistency score are concatenated to form the multidimensional health feature vector.

[0096] S160. Based on the preset demand evolution prediction model, using the continuous behavior sequence and the multidimensional health feature vector as input, predict the user life cycle status, demand decay trend and user behavior health index.

[0097] Specifically, based on a preset demand evolution prediction model, the user lifecycle state, demand decay trend, and user behavior health index are predicted using the continuous behavior sequence and the multidimensional health feature vector as input; wherein, the user lifecycle state is output in the form of a probability distribution, and the user behavior health index is obtained by mapping the multidimensional health feature vector through the fully connected layer in the demand evolution prediction model.

[0098] The demand evolution prediction model includes:

[0099] An input embedding layer is used to map the continuous sequence of behaviors into a sequence of embedding vectors;

[0100] The feature concatenation layer is used to concatenate the multidimensional health feature vector with the embedded vector sequence;

[0101] The Bi-LSTM layer of the bidirectional long short-term memory network is used to capture bidirectional temporal dependencies of the concatenated vector;

[0102] A multi-head self-attention layer is used to calculate the importance weights of each time step and perform weighted aggregation.

[0103] The first output branch, including the first fully connected layer and the Softmax activation function, is used to output the probability distribution of the user's lifecycle state;

[0104] The second output branch, including a second fully connected layer, is used to output continuous predicted values ​​of the demand decay trend;

[0105] The third output branch, including the third fully connected layer and the Sigmoid activation function, is used to output the user behavior health index.

[0106] Specifically, the input embedding layer maps the continuous behavioral sequence (such as the daily behavioral event encoding of the last 30 days) into an embedding vector sequence E∈R. L×de Where L=30 is the time step. For the embedded dimension.

[0107] The feature splicing layer combines the multidimensional health feature vector H∈R 4 After expanding to the dimension aligned with the embedded sequence, it is concatenated with E to obtain X∈R.L×(de+4) .

[0108] The Bi-LSTM layer comprises two bidirectional long short-term memory (LSTM) layers, each with 128 hidden units and an output dimension of 256. The forward LSTM captures historical dependencies, and the backward LSTM captures future trends; the concatenated layers output H. lstm ∈R L×256 .

[0109] The multi-head self-attention layer contains 8 attention heads, each with a dimension of 32, and calculates the importance weights for each time step: The weighted aggregation outputs a context vector C∈R. 256 .

[0110] The first output branch includes a first fully connected layer (256→64, ReLU) and a Softmax activation function, outputting the probability distribution of the user's lifecycle state. .

[0111] The second output branch includes a second fully connected layer (256→32→1), which outputs continuous predictions of the demand decay trend. A larger value indicates that demand is declining more rapidly.

[0112] The third output branch includes a third fully connected layer (256→16→1) and a Sigmoid activation function, which outputs the user behavior health index after nonlinear mapping of the multidimensional health feature vector. .

[0113] S170. Based on the comprehensive demand evaluation results, the demand decay trend, the user behavior health index, and the user lifecycle status, determine the demand response level and trigger the target operation strategy corresponding to the demand response level. The target operation strategy is used to instruct the edge analysis layer to perform a response action and send a reach signal to the user terminal.

[0114] In this embodiment, a demand response level is determined based on the urgency score, the demand decay trend, the user behavior health index, and the churn tendency probability in the user lifecycle state probability distribution, and a target operation strategy corresponding to the demand response level is triggered; the target operation strategy is used to instruct the edge analysis layer to perform a response action and send a reach signal to the user terminal.

[0115] Specifically, the churn tendency probability is extracted from the probability distribution of the user's lifecycle state;

[0116] A correction coefficient is calculated, which is determined based on the ratio of the churn tendency probability to the user behavior health index. The higher the ratio, the larger the correction coefficient, in order to reflect the nonlinear amplification effect of churn risk on the urgency.

[0117] The dynamic urgency index is obtained by multiplying the urgency score by the demand decay trend.

[0118] Multiply the dynamic urgency index by the correction coefficient to obtain the level determination value;

[0119] The level determination value is compared with a preset first threshold and a second threshold: when the level determination value is less than or equal to the first threshold, the demand response level is determined to be the first level; when the level determination value is greater than the first threshold and less than or equal to the second threshold, it is determined to be the second level; when the level determination value is greater than the second threshold, it is determined to be the third level. An automated telephone call outreach message is sent to the user terminal, and automated marketing outreach to that user is suspended.

[0120] S180. Obtain the feedback behavior data of the user terminal to the touch signal, dynamically adjust the judgment threshold of the demand response level according to the feedback behavior data, and send the adjusted judgment threshold as a demand detection parameter to the edge analysis layer.

[0121] In this embodiment, the edge analysis layer maps the determination threshold to the semantic matching threshold according to a preset ratio, and synchronously adjusts the temporal window length according to the adjustment range of the determination threshold.

[0122] The system acquires feedback behavior data of the user terminal to the touch signal, calculates the average user response time and ignore rate based on the feedback behavior data, dynamically adjusts the judgment threshold of the demand response level based on the average user response time and the ignore rate, and sends the adjusted judgment threshold as a demand detection parameter to the edge analysis layer to update the detection parameters of the lightweight multimodal recognition model in the edge analysis layer.

[0123] In this embodiment, when the ignore rate is higher than a preset ignore threshold and the average user response time exceeds a preset time threshold, the first threshold and the second threshold are reduced.

[0124] When the ignore rate is lower than the preset ignore threshold and the average user response time is lower than the preset time threshold, the first threshold and the second threshold are increased.

[0125] The adjusted first threshold and second threshold are sent to the edge analysis layer as the demand detection parameters. The edge analysis layer updates the semantic matching threshold and the temporal window length according to the demand detection parameters, and uses the updated semantic matching threshold for the similarity calculation layer, and uses the updated temporal window length for the time window division of the cross-time window tracking and positioning.

[0126] In response to different levels of demand, differentiated target operational strategies are triggered:

[0127] Level 1: Sending application push notifications (such as in-app pushes) to user terminals, with the content primarily focused on information delivery;

[0128] Level 2: Send instant messaging messages to the user terminal (such as WeChat service notifications, SMS), while controlling the edge analysis layer to reduce the frequency of automatic outreach to the user (such as from 3 messages per day to 1 message per day) to avoid excessive disturbance.

[0129] Level 3: Send an automated phone call to the user's terminal and suspend automated marketing outreach to the user, then transfer the call to a dedicated human customer service representative.

[0130] The system acquires user terminal feedback behavior data to the touch signal, including whether the user clicked, responded, ignored, and the response time. It calculates the user's average response time and ignore rate. When the ignore rate is higher than a preset ignore threshold and the user's average response time exceeds a preset time threshold, it lowers the first and second thresholds to improve response sensitivity; when the ignore rate is lower than the preset ignore threshold and the user's average response time is lower than the preset time threshold, it raises the first and second thresholds to reduce response sensitivity. The adjusted first and second thresholds are sent to the edge analysis layer as demand detection parameters. The edge analysis layer updates the semantic matching threshold and the temporal window length based on the demand detection parameters, and uses the updated semantic matching threshold for score filtering in the similarity calculation layer, and uses the updated temporal window length for time window division in cross-time window tracking and positioning.

[0131] In another embodiment, the step S180 described above is followed by:

[0132] S190. Generate a collaborative optimization strategy based on the comprehensive demand evaluation results, the user lifecycle status, and the demand decay trend. Optimize the parameters of the multi-source fusion model and the demand evolution prediction model, and send the collaborative optimization strategy to the data acquisition layer and the edge analysis layer.

[0133] In this embodiment, when the confidence level of the multi-source fusion model is lower than a preset confidence threshold, the collaborative optimization strategy generates instructions to increase the learning rates of the text encoding subnetwork and image encoding subnetwork of the multi-source fusion model; when the prediction error of the demand decay trend exceeds a preset error threshold, the collaborative optimization strategy generates instructions to adjust the dropout rate, learning rate, or gradient pruning threshold of the bidirectional long short-term memory network layer of the demand evolution prediction model, or generates instructions to enable the model ensemble mechanism to increase the number of parallel subnetworks; when the prediction deviation of the user lifecycle state exceeds a preset deviation threshold, the collaborative optimization strategy generates instructions to adjust the learning rate of the input embedding layer or enable the auxiliary embedding branch.

[0134] The collaborative optimization strategy is sent to the data acquisition layer and the edge analysis layer. The data acquisition layer adjusts the data acquisition parameters according to the collaborative optimization strategy, and the edge analysis layer adjusts the detection parameters according to the collaborative optimization strategy and the demand detection parameters.

[0135] Based on the comprehensive demand evaluation results, the user lifecycle status, and the demand decay trend, a collaborative optimization strategy is generated:

[0136] When the confidence level of the multimodal fusion (i.e., the confidence component in the comprehensive demand evaluation result) is lower than the preset confidence threshold (e.g., 0.85), an instruction is generated to increase the learning rate of the RoBERTa text encoding subnetwork and the Swin Transformer image encoding subnetwork (e.g., from 1e-5 to 3e-5).

[0137] When the prediction error (root mean square error RMSE) of the demand decay trend exceeds a preset error threshold (e.g., 0.15), instructions are generated to adjust the hidden layer dimension of the Bi-LSTM layer (e.g., from 128 to 256) and the number of attention heads in the multi-head self-attention layer (e.g., from 8 to 16). When the prediction error of the demand decay trend exceeds the preset error threshold, the collaborative optimization strategy generates instructions to adjust the dropout rate, learning rate, or gradient pruning threshold of the bidirectional long short-term memory network layer; or generates instructions to enable the model ensemble mechanism to increase the number of parallel sub-networks.

[0138] When the prediction deviation (cross-entropy loss) of the user lifecycle state exceeds a preset deviation threshold (e.g., 0.5), an instruction is generated to reconstruct the embedding dimension of the input embedding layer (e.g., adjust it from 128 to 256).

[0139] The collaborative optimization strategy is sent to the edge analysis layer. The data acquisition layer adjusts the data acquisition parameters according to the collaborative optimization strategy, such as increasing the text sampling frequency, reducing the image resolution threshold to alleviate transmission pressure, or refining the granularity of behavior log collection. The edge analysis layer adjusts the detection parameters according to the collaborative optimization strategy and the demand detection parameters.

[0140] When the confidence level of the multimodal fusion is lower than a preset confidence threshold, the collaborative optimization strategy generates instructions to increase the learning rates of the text coding subnetwork and the image coding subnetwork.

[0141] When the prediction error of the demand decay trend exceeds a preset error threshold, the collaborative optimization strategy generates instructions to adjust the hidden layer dimension of the bidirectional long short-term memory network layer and the number of attention heads of the multi-head self-attention layer.

[0142] When the prediction deviation of the user lifecycle state exceeds a preset deviation threshold, the collaborative optimization strategy generates instructions to adjust the learning rate of the input embedding layer or enable the auxiliary embedding branch.

[0143] The collaborative optimization strategy is sent to the edge analysis layer. The data acquisition layer adjusts the data acquisition parameters according to the collaborative optimization strategy, and the edge analysis layer adjusts the detection parameters according to the collaborative optimization strategy and the demand detection parameters.

[0144] S200. Obtain strategy execution feedback data and the user's spatiotemporal behavior information. Based on the user's spatiotemporal behavior information, map the strategy execution feedback data to a spatiotemporal coordinate system with the community channel as the spatial dimension and time as the temporal dimension to obtain a user behavior spatiotemporal correlation map.

[0145] In this embodiment, the user behavior spatiotemporal correlation graph refers to a heterogeneous graph constructed by mapping policy execution feedback data to a coordinate system with community channels as the spatial dimension and time as the temporal dimension, based on user spatiotemporal behavior information, forming spatiotemporal distributed nodes, and establishing directed correlation edges according to the interaction links.

[0146] Specifically, spatial coordinates are determined based on the location information of the user node in the community channel, and time coordinates are determined based on the topic participation timestamp;

[0147] The demand identification results, user behavior health index, and historical demand handling records in the strategy execution feedback data are used as node attribute features and mapped to a spatiotemporal coordinate system to form spatiotemporally distributed nodes.

[0148] Based on the interaction link information, directed association edges are established between spatiotemporally distributed nodes, and the weight of the directed association edges is determined by the interaction frequency and the interaction delay time.

[0149] The spatiotemporal correlation graph of user behavior is constructed based on spatiotemporal distributed nodes and directed related edges.

[0150] Based on the user's spatiotemporal behavior information, the strategy execution feedback data is mapped to a spatiotemporal coordinate system. Specifically: Spatial dimension: using the community channel as the spatial axis, the spatial coordinates are determined according to the user node's position information within the channel; Temporal dimension: the temporal coordinates are determined by the topic participation timestamp. The demand identification results, user behavior health index, and historical demand handling records are used as node attribute features and mapped to the spatiotemporal coordinate system to form spatiotemporally distributed nodes. Directed association edges are established between the spatiotemporally distributed nodes based on interaction link information, with edge weights determined by interaction frequency and interaction delay time. Specifically, the higher the interaction frequency and the shorter the interaction delay time, the greater the edge weight. The user behavior spatiotemporal association graph is constructed based on the spatiotemporal distributed nodes and directed association edges.

[0151] S210. Input the spatiotemporal correlation graph of user behavior into a preset demand causal analysis model for analysis to obtain a demand tracing report.

[0152] In this embodiment, the demand tracing report refers to the analysis results obtained by inputting the spatiotemporal correlation graph of user behavior into the demand causal analysis model and then performing graph attention network and temporal convolution analysis. This analysis is used to characterize the triggering factors and propagation paths of implicit demands of community users.

[0153] The demand causal analysis model includes:

[0154] The graph construction layer is used to construct a heterogeneous graph based on the user's spatiotemporal behavior information. The nodes include user nodes, content nodes, and channel nodes, and the edges include interaction edges, temporal edges, and co-occurrence edges.

[0155] The graph attention network GAT layer has 4 attention heads and is used to aggregate neighbor node information through a multi-head graph attention mechanism.

[0156] Temporal convolutional layers are used to perform one-dimensional convolution on node features along the time axis to capture the temporal pattern of demand propagation.

[0157] The causal inference layer is used to backpropagate based on the attention weights of the GAT layer, identify the demand-inducing nodes and propagation paths, and generate the demand tracing report.

[0158] The spatiotemporal correlation graph of user behavior is input into a preset demand causal analysis model. The graph construction layer constructs a heterogeneous graph based on user spatiotemporal behavior information. The node types include user nodes, content nodes, and channel nodes, and the edge types include interaction edges, temporal edges, and co-occurrence edges.

[0159] The graph attention network layer employs a multi-head attention mechanism. For each node and its neighboring nodes, under each attention head, the node features are concatenated with the neighboring node features, followed by linear transformation and nonlinear activation, and then normalized exponential processing (obtaining attention coefficients through either Softmax function normalization or exponential transformation and normalization) to obtain attention coefficients. Based on these attention coefficients, the neighboring node features are weighted and summed, and after nonlinear activation, aggregated features are obtained. The outputs of multiple attention heads are concatenated to form the final node representation.

[0160] The temporal convolutional layer performs one-dimensional convolution on the node features along the time axis. The convolutional kernel expands the receptive field layer by layer to capture the temporal pattern of demand propagation and output a temporally enhanced node representation.

[0161] The causal inference layer is based on the attention weight backpropagation of the graph attention network layer. It traces the content node and channel node with the highest attention weight as the demand inducement node, and determines the propagation path of demand from the inducement node to the target user by tracing back the time sequence edge and interaction edge, and generates the demand tracing report.

[0162] S220. Update the preset implicit requirement content according to the requirement tracing report, and send the updated implicit requirement content to the edge analysis layer to optimize the next round of implicit requirement preliminary identification.

[0163] Specifically, high-frequency demand-inducing nodes and key propagation paths are extracted from the demand sourcing report;

[0164] Update the set of text descriptions in the preset implicit demand content according to the high-frequency demand trigger nodes;

[0165] The updated implicit requirement content is sent to the edge analysis layer. The edge analysis layer expands the description templates in the requirement type retrieval library based on the updated implicit requirement content, or adds keyword matching rules that can identify requirement types, so as to optimize the next round of preliminary identification of implicit requirements.

[0166] Extract high-frequency demand trigger nodes (such as the top 5 most frequent content themes or channel topics) and key dissemination paths (such as the path from the "Sharing of Good Baby Products" channel to the general user group via KOL user A) from the demand tracing report.

[0167] The text description set in the preset implicit demand content is updated based on the high-frequency demand trigger nodes. For example, if the traceability report shows that "baby food tools" is a high-frequency trigger, then related words such as "baby food", "grinding bowl", and "bib" are added to the text description set.

[0168] The updated implicit requirements are sent to the edge analysis layer.

[0169] The edge analysis layer expands the description templates in the requirement type retrieval library based on the updated implicit requirement content, or adds keyword matching rules that can identify requirement types, in order to optimize the next round of preliminary identification of implicit requirements.

[0170] Alternatively, the edge analysis layer, based on the updated implicit demand content, can fine-tune the classification output layer of the preliminary identification layer by incrementally learning samples while maintaining the original network weights, in order to support the identification of new demand types.

[0171] In this embodiment, a collaborative optimization strategy is generated based on the comprehensive demand evaluation results, the user lifecycle status, and the demand decay trend. The parameters of the multi-source fusion model and the demand evolution prediction model are optimized according to the collaborative optimization strategy, and the collaborative optimization strategy is sent to the edge analysis layer. The data acquisition layer is used to adjust data acquisition parameters according to the collaborative optimization strategy, and the edge analysis layer is used to adjust the detection parameters in the preliminary identification of implicit demands according to the collaborative optimization strategy and the demand detection parameters.

[0172] Acquire strategy execution feedback data and the user's spatiotemporal behavior information; the strategy execution feedback data includes the demand identification results after the data acquisition layer adjusts the data acquisition parameters and the historical demand handling records of the community users; the user's spatiotemporal behavior information includes the location information of the user node in the community channel, the topic participation timestamp, and the interaction link information.

[0173] Based on the user's spatiotemporal behavior information, the strategy execution feedback data is mapped to a spatiotemporal coordinate system with the community channel as the spatial dimension and time as the temporal dimension to obtain a user behavior spatiotemporal correlation graph; wherein, the node features of the user behavior spatiotemporal correlation graph include the demand identification results, the user behavior health index and the historical demand handling records, and the edge features are determined by the interaction link information;

[0174] The spatiotemporal correlation graph of user behavior is input into a preset demand causal analysis model for analysis to obtain a demand tracing report; the demand tracing report is used to represent the causes and propagation paths of the implicit demands of the community users;

[0175] The preset implicit requirement content is updated according to the requirement tracing report, and the updated implicit requirement content is sent to the edge analysis layer to optimize the next round of implicit requirement preliminary identification.

[0176] The data acquisition system deployed in this method has its data collection layer deployed on the community client or community server side. This layer is used to collect user-generated text content (such as chat messages and post text), image content (such as user-uploaded product images and emoticons), and behavioral log data (such as clicks, dwell time, forwarding, adding to cart, liking, and other behavioral events) in real time. The data collection parameters include text sampling frequency (such as sampling the message queue once per second), image resolution threshold (such as a minimum of 512×512 pixels), and behavioral log collection granularity (such as event-level millisecond-level timestamps).

[0177] The edge analysis layer is deployed on edge computing nodes (such as regional servers or 5G MEC nodes) to perform low-latency preliminary identification of implicit needs in the multi-dimensional community data, obtaining preliminary need identification results. The preliminary need identification results include need type (such as price-sensitive, quality-seeking, social fission, etc.), semantic matching degree (the degree of semantic consistency between text and image), and identification confidence.

[0178] The cloud platform layer is deployed on the central cloud server and is used to receive the preliminary demand identification results and the multi-dimensional community data, and to perform deep multi-source fusion, demand evolution prediction, demand response level determination, collaborative optimization strategy generation, spatiotemporal graph construction, and demand causal analysis.

[0179] The method in this embodiment forms a two-level progressive recognition architecture by combining a lightweight multimodal recognition model at the edge layer with a multi-source fusion model in the cloud. This ensures low-latency response at the edge while improving the accuracy of demand recognition through deep semantic fusion of RoBERTa, SwinTransformer, and TCN. By coupling multidimensional state evolution assessment with demand evolution prediction model, user activity, social relationships, interest decay, and behavioral consistency are quantified into health feature vectors, enabling accurate prediction of user lifecycle status and demand decay trends. The demand response level is determined by coupling urgency score, demand decay trend, user behavior health index, and churn probability, and the judgment threshold is dynamically adjusted based on user feedback, achieving adaptive closed-loop optimization of operational strategies. By constructing a spatiotemporal correlation graph of user behavior and performing causal analysis based on GAT and temporal convolution, the method achieves accurate tracing of implicit demand triggers and propagation paths, and supports reverse writing and updating of demand content, forming a complete closed loop from recognition, prediction, operation to tracing.

[0180] For example, taking a private community of a maternal and infant brand as a complete scenario, the community has three channels: "Complementary Food Exchange", "Formula Q&A" and "Daily Parenting", which gather more than 5,000 mother users. The system uses a three-layer architecture to conduct full-link demand mining on user "Mother A".

[0181] One day, the data collection layer captured complete data posted by Mother A in the "Baby Food Exchange" channel through the community open interface: the text content was "My baby is 8 months old, and he always turns his head away when eating rice cereal. Should I switch to something else?", along with two uploaded pictures: one was a real photo of the baby turning his head away and refusing to eat, and the other was a screenshot of a shopping cart containing a baby food grinder bowl and a silicone bib; the accompanying behavior log showed that this user had clicked on baby food product links 12 times in the past seven days, spending a total of 45 minutes on each link, and had forwarded a "Baby Food Introduction Guide" to a group of close friends, but had not placed any orders in the past three days. After being anonymized, compressed, and structured, this data was simultaneously transmitted to the edge analysis layer and the cloud platform layer.

[0182] At the edge analysis layer, a lightweight multimodal recognition model deployed on a regional server quickly begins computation. The DistilBERT-based text encoding branch performs semantic extraction on the posted text, identifying the core focus as "baby food tools and tableware" rather than simply milk powder or parenting knowledge. The MobileNet-V3-based image encoding branch performs visual encoding on two images, identifying the first image as a scene of an infant refusing to eat, and the second image as a screenshot of a baby food grinder and bib. Subsequently, the feature concatenation layer concatenates the 768-dimensional text features and 1280-dimensional image features into a joint feature vector. This vector is then processed by the fully connected layer of the preliminary recognition layer and a Softmax classifier, outputting the demand type as "baby food tool demand" with a confidence score of 0.82. The similarity calculation layer further calculates the cosine similarity between the text and the image, yielding a semantic matching score of 0.79. At this point, the edge layer completes the low-latency preliminary recognition and uploads the results to the cloud.

[0183] After receiving the preliminary results and raw data, the multi-source fusion model at the cloud platform layer initiates deep semantic fusion. A RoBERTa-based text encoding sub-network performs deep semantic encoding on the text, a Swin Transformer-based image encoding sub-network performs hierarchical visual encoding on the images, and a temporal convolutional network (TCN)-based behavior encoding sub-network performs temporal encoding on the behavioral sequences of "Mom A" over the past seven days, including clicks, dwell times, forwarding, and no-orders. The cross-modal cross-attention fusion layer uses text semantic features as query vectors and image visual features and behavioral temporal features as key-value pairs for multi-head cross-attention calculation. The model discovers that text semantics such as "eight months," "turning head," and "changing to another" are highly correlated with the image of the grinding bowl and the "click without order" behavior, thus outputting multimodal interaction features. The gated fusion layer combines the recognition confidence score of 0.82 from the edge layer with the historical quality labels from the three data sources to calculate the dynamic fusion weight. It then performs weighted fusion on the multimodal interaction features and semantic matching score of 0.79, and finally outputs the comprehensive demand evaluation result: the demand type is the need for complementary food tools, the urgency score is 0.75, and the multimodal fusion confidence score is 0.88.

[0184] Meanwhile, the system assigned a unique tracking identifier to Mother A's behavior logs over the past thirty days. Combining channel location information and topic participation timestamps, it performed cross-time window tracking, determining that her browsing, clicking, and posting behaviors within the "Complementary Food Exchange" channel and its sub-channels belonged to a continuous sequence of behaviors along the same user behavior path. Based on this continuous behavior sequence, the system conducted a multi-dimensional state evolution assessment: by calculating the contrast between the user's activity signal and the community background baseline, the activity visibility score was found to be 0.45, indicating a significant weakening of her activity signal; by constructing a social network graph and extracting the relationship skeleton, it was found that the reply link between Mother A and her frequently interacting friends was broken, with increased edge roughness, and the relationship integrity score dropped to 0.38; after mapping her interest tags to the semantic heat space, the deviation between the current heat vector and the historical baseline, combined with time decay analysis, yielded an interest decay score of 0.62; after fitting her behavior trajectory points with a cubic spline curve, a significant lateral deviation was found between the actual trajectory and the fitted curve, resulting in a behavior consistency score of 0.41. These four scores were concatenated to form a multi-dimensional health feature vector with specific values ​​of 0.45, 0.38, 0.62, and 0.41.

[0185] The multidimensional health feature vector, along with the continuous behavior sequence, is input into the demand evolution prediction model. The model first maps the behavior sequence to an embedding vector through an input embedding layer. After being fused with the health vector through a feature concatenation layer, a bidirectional long short-term memory (Bi-LSTM) network is used to capture temporal dependencies, and a multi-head self-attention layer assigns the highest weight to the "silent period" of the past three days. The first output branch, activated by Softmax, outputs a probability distribution of the user's lifecycle state, with a churn tendency probability as high as 35%. The second output branch gives a demand decay trend prediction value of 0.68, indicating that demand is rapidly decaying. The third output branch maps the health feature vector to a user behavior health index through a fully connected layer and sigmoid activation, resulting in a value of 0.42, lower than the health threshold. Based on these outputs, the system extracts the churn tendency probability and calculates a correction coefficient. Then, the urgency score of 0.75 is multiplied by the decay trend of 0.68 to obtain a dynamic urgency index, which is further multiplied by the correction coefficient to obtain a level determination value. This value exceeds the preset second threshold of 0.7, therefore the demand response level is determined to be level three. The system immediately paused the automated marketing outreach to Mother A and instead triggered an automated phone call, with a dedicated customer service representative intervening and preparing a limited-time offer for a grinding bowl set and bib combination specifically for the eight-month-old baby's food preparation needs.

[0186] Within fifteen minutes of the phone call, Mother A placed an order. The system recorded an extremely short response time and a zero ignore rate for her feedback behavior data. Based on this, it determined that the current threshold setting was too conservative. Therefore, it increased the first and second thresholds and sent the adjusted thresholds as demand detection parameters to the edge analysis layer. Upon receiving this, the edge analysis layer updated the semantic matching threshold and the temporal window length, making the screening criteria for the next round of preliminary identification more accurate. At this point, the basic operational closed loop of the first embodiment was fully formed, realizing an adaptive cycle from edge identification, cloud fusion, evolution prediction, hierarchical operation to feedback parameter tuning.

[0187] Building upon this, the second embodiment enters the model self-optimization and demand tracing stage. The system detects that the prediction deviation of the user's lifecycle state exceeds a preset threshold—the model predicted that Mother A had a 35% probability of churn, but in reality, she was successfully retained via phone—therefore, a collaborative optimization strategy is generated, instructing the reconstruction of the embedding dimension of the demand evolution prediction model's input embedding layer, expanding it from 128 dimensions to 256 dimensions to enhance the ability to perceive subtle signals from users during the silent period. Simultaneously, the data acquisition layer, based on the collaborative optimization strategy, increases the text sampling frequency of the "Baby Food Communication" channel where Mother A is located from once per second to twice per second, and refines the granularity of behavior log collection from minute-level to second-level. Subsequently, the system acquires strategy execution feedback data and user spatiotemporal behavior information, determines spatial coordinates based on channel location information, determines time coordinates based on topic participation timestamps, maps demand identification results, user behavior health index, and historical handling records into spatiotemporal distribution nodes, and establishes directed association edges based on interaction links. The system discovered that Mom A's purchase occurred after replying to "Mom B's" post, and that after Mom A placed her order, 23 other users in the same channel clicked on the same grinding bowl link. Based on this, a spatiotemporal correlation graph of user behavior was constructed.

[0188] After inputting the graph into the demand causal analysis model, the graph construction layer establishes a heterogeneous graph containing user nodes, content nodes, and channel nodes; the graph attention network GAT layer aggregates neighbor information through a four-head attention mechanism, discovering that the attention weights of the "Mom B" node and the "Grinding Bowl Review Post" content node are significantly higher than other nodes; the temporal convolutional layer captures the demand propagation pattern along the time axis, reconstructing the complete chain of "Mom B posted on the evening of April 8th, Mom A browsed and asked for help at noon on April 9th, Mom A placed an order that afternoon, and 23 people followed and clicked in the evening"; the causal inference layer uses attention weights for backpropagation, ultimately identifying Mom B's post about her experience using the grinding bowl as the demand trigger node, with the propagation path exhibiting a typical structure of key opinion recommendation, precise user conversion, and community follow-up effect. The resulting demand tracing report clearly points out that the root cause of this surge in demand for complementary food tools lies in Mom B's content recommendation.

[0189] The system extracts high-frequency demand triggers and key propagation paths from the source tracing report, updates related terms such as "grinding bowl," "baby food tool set," and "eight-month-old tableware" to a pre-defined set of implicit demand descriptions, and then sends the updated content to the edge analysis layer. Based on this, the edge analysis layer adjusts the label space of the Softmax classifier in the initial identification layer, adding a subcategory of "baby food tool demand." The next day, when another mother posted a picture and text about her nine-month-old baby constantly spilling noodles, the edge layer could directly identify it as "baby food tool demand - tableware category," increasing the confidence level of the identification, thus completing a full closed loop from identification, prediction, operation, source tracing to model self-evolution.

[0190] The aforementioned method for mining latent needs of private community users based on multi-dimensional semantic fusion acquires multi-dimensional community data, including text, images, and behavioral logs, at the data acquisition layer. At the edge analysis layer, a lightweight multimodal recognition model is used for preliminary need identification to quickly capture users' potential needs. Then, at the cloud platform layer, a multi-source fusion model is used to deeply integrate the preliminary identification results with the original data to generate a comprehensive need evaluation result. Next, by assigning tracking identifiers to user behavior paths and combining spatiotemporal information to evaluate multi-dimensional health feature vectors, a pre-defined need evolution prediction model is used to predict user lifecycle status, need decay trends, and... The behavioral health index enables accurate prediction of user status. Based on these predictions, the demand response level is determined and corresponding operational strategies are triggered. At the same time, the threshold for judging the demand response level is dynamically adjusted based on user feedback to the reach signals. While optimizing the operational strategies, updated parameters are also sent to the edge analysis layer. This realizes a complete process from deep integration of multimodal information, accurate prediction of user status evolution, adaptive optimization of operational strategies, to spatiotemporal tracing of demand causes. It effectively solves the problems of insufficient multimodal fusion, lack of user status evolution prediction, separation of operational strategies and model parameters, and inability to trace demand causes in existing technologies.

[0191] Figure 3 This is a schematic block diagram of a private community user latent needs mining system 300 based on multi-dimensional semantic fusion, provided in an embodiment of the present invention. Figure 3 As shown, corresponding to the above-described method for mining implicit needs of private community users based on multi-dimensional semantic fusion, this invention also provides a system 300 for mining implicit needs of private community users based on multi-dimensional semantic fusion. This system 300 includes a unit for executing the above-described method for mining implicit needs of private community users based on multi-dimensional semantic fusion, and the system can be configured in a server. Specifically, please refer to... Figure 3The system 300 for mining the implicit needs of users in private communities based on multi-dimensional semantic fusion includes an acquisition unit 301, a preliminary identification unit 302, a fusion unit 303, a sequence determination unit 304, an evaluation unit 305, a prediction unit 306, a triggering unit 307, and a sending unit 308.

[0192] The system comprises the following components: an acquisition unit 301, used to acquire multi-dimensional community data and user spatiotemporal behavior information; a preliminary identification unit 302, used at the edge analysis layer to perform preliminary identification of implicit needs on the multi-dimensional community data based on a preset lightweight multimodal identification model, to obtain preliminary need identification results; a fusion unit 303, used at the cloud platform layer to fuse the preliminary need identification results and the multi-dimensional community data based on a preset multi-source fusion model, to obtain a comprehensive need evaluation result; a sequence determination unit 304, used to assign tracking identifiers to user behavior paths in the behavior log data, and determine continuous behavior sequences of the same user behavior path by combining user spatiotemporal behavior information; and an evaluation unit 305, used to perform multi-dimensional state evolution evaluation based on the continuous behavior sequences, to obtain... A multidimensional health feature vector; a prediction unit 306, used to predict user lifecycle status, demand decay trend, and user behavior health index based on a preset demand evolution prediction model, using the continuous behavior sequence and the multidimensional health feature vector as input; a triggering unit 307, used to determine the demand response level based on the comprehensive demand evaluation result, the demand decay trend, the user behavior health index, and the user lifecycle status, and trigger the target operation strategy corresponding to the demand response level; and a sending unit 308, used to acquire the user terminal's feedback behavior data to the touch signal, dynamically adjust the judgment threshold of the demand response level based on the feedback behavior data, and send the adjusted judgment threshold as a demand detection parameter to the edge analysis layer.

[0193] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned private domain community user implicit demand mining system 300 based on multi-dimensional semantic fusion and each unit can be referred to the corresponding description in the aforementioned method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0194] The aforementioned system 300 for mining the implicit needs of users in private communities based on multi-dimensional semantic fusion can be implemented as a computer program, which can be used in, for example... Figure 3 It runs on the computer device shown.

[0195] Please see Figure 3 , Figure 3This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0196] See Figure 3 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0197] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a method for mining implicit needs of private community users based on multi-dimensional semantic fusion.

[0198] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0199] The internal memory 504 provides an environment for the execution of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a method for mining the implicit needs of private community users based on multi-dimensional semantic fusion.

[0200] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0201] The processor 502 is used to run the computer program 5032 stored in the memory to implement all the steps of the method for mining the implicit needs of private community users based on multi-dimensional semantic fusion.

[0202] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0203] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0204] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform all the steps of the method for mining implicit needs of private community users based on multi-dimensional semantic fusion.

[0205] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0206] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0207] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0208] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the system of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0209] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part 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, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0210] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for mining the implicit needs of users in private domain communities based on multi-dimensional semantic fusion, characterized in that, The cloud platform layer applied to the demand mining system, which also includes a data acquisition layer, an edge analysis layer, and a cloud platform layer, includes: Acquire multi-dimensional community data and user spatiotemporal behavior information; In the edge analysis layer, a preliminary implicit demand identification is performed on the multi-dimensional community data based on a preset lightweight multimodal recognition model to obtain preliminary demand identification results. At the cloud platform layer, the preliminary demand identification results and the multi-dimensional community data are fused based on a preset multi-source fusion model to obtain a comprehensive demand evaluation result; Assign tracking identifiers to user behavior paths in the behavior log data, and determine continuous behavior sequences for the same user behavior path by combining user spatiotemporal behavior information; A multidimensional state evolution assessment is performed based on the continuous behavioral sequence to obtain a multidimensional health feature vector; Based on the preset demand evolution prediction model, the user lifecycle status, demand decay trend and user behavior health index are predicted by taking the continuous behavior sequence and the multidimensional health feature vector as input. Based on the comprehensive demand evaluation results, the demand decay trend, the user behavior health index, and the user lifecycle status, the demand response level is determined, and the target operation strategy corresponding to the demand response level is triggered. The system acquires feedback behavior data of the user terminal in response to the reach signal, dynamically adjusts the threshold for determining the demand response level based on the feedback behavior data, and sends the adjusted threshold as a demand detection parameter to the edge analysis layer.

2. The method for mining implicit needs of private community users based on multi-dimensional semantic fusion according to claim 1, characterized in that, After acquiring the user terminal's feedback behavior data to the reach signal, dynamically adjusting the threshold for determining the demand response level based on the feedback behavior data, and sending the adjusted threshold as a demand detection parameter to the edge analysis layer, the method further includes: Based on the comprehensive demand evaluation results, the user lifecycle status, and the demand decay trend, a collaborative optimization strategy is generated to optimize the parameters of the multi-source fusion model and the demand evolution prediction model, and the collaborative optimization strategy is sent to the edge analysis layer. Obtain strategy execution feedback data and user spatiotemporal behavior information. Based on the user spatiotemporal behavior information, map the strategy execution feedback data to a spatiotemporal coordinate system with community channel as the spatial dimension and time as the temporal dimension to obtain a user behavior spatiotemporal correlation map. The spatiotemporal correlation graph of user behavior is input into a preset demand causal analysis model for analysis to obtain a demand tracing report. The preset implicit requirement content is updated according to the requirement tracing report, and the updated implicit requirement content is sent to the edge analysis layer to optimize the next round of implicit requirement preliminary identification.

3. The method for mining implicit needs of private community users based on multi-dimensional semantic fusion according to claim 1, characterized in that, The preset lightweight multimodal recognition model includes: A text encoding branch based on DistilBERT is used to encode the text content to obtain edge text features; The image coding branch based on MobileNet-V3 is used to encode the image content to obtain edge image features; A feature splicing layer is used to splice the edge text features and the edge image features; The preliminary identification layer includes a first fully connected layer and a Softmax classifier, which are used to output the demand type and the identification confidence based on the concatenated features; A similarity calculation layer is used to calculate the cosine similarity between the edge text features and the edge image features, and output the semantic matching degree.

4. The method for mining implicit needs of private community users based on multi-dimensional semantic fusion according to claim 1, characterized in that, The multidimensional state evolution evaluation based on the continuous behavioral sequence yields a multidimensional health feature vector, including: User activity signals are extracted from the continuous behavior sequence, and the Weber contrast or Michelson contrast between the user activity signals and the community background baseline is calculated to obtain the activity visibility score. A social network graph is constructed based on the social relationship data in the continuous behavior sequence. The relationship skeleton is extracted and the skeleton continuity index and edge roughness index are calculated to obtain the relationship integrity score. The interest tag data in the continuous behavior sequence is mapped to a preset semantic popularity space, the deviation between the current popularity value and the historical benchmark popularity value is calculated, and the aging trend analysis is performed in combination with the time decay function to obtain the interest decay score. A cubic spline curve is fitted to the behavior trajectory points in the continuous behavior sequence, and the fitting residual and lateral deviation are calculated to obtain the behavior consistency score. The activity visibility score, the relationship integrity score, the interest decay score, and the behavior consistency score are concatenated to form the multidimensional health feature vector.

5. The method for mining implicit needs of private community users based on multi-dimensional semantic fusion according to claim 3, characterized in that, The step of acquiring feedback behavior data of the user terminal to the reach signal, dynamically adjusting the judgment threshold of the demand response level based on the feedback behavior data, and sending the adjusted judgment threshold as a demand detection parameter to the edge analysis layer includes: When the ignore rate is higher than a preset ignore threshold and the average user response time exceeds a preset time threshold, the first threshold and the second threshold are reduced. When the ignore rate is lower than the preset ignore threshold and the average user response time is lower than the preset time threshold, the first threshold and the second threshold are increased. The adjusted first threshold and second threshold are sent to the edge analysis layer as the demand detection parameters. The edge analysis layer updates the semantic matching threshold and the temporal window length according to the demand detection parameters, and uses the updated semantic matching threshold for the similarity calculation layer, and uses the updated temporal window length for the time window division of the cross-time window tracking and positioning.

6. The method for mining implicit needs of private community users based on multi-dimensional semantic fusion according to claim 2, characterized in that, The parameter optimization of the multi-source fusion model and the demand evolution prediction model includes: When the confidence level of the multi-source fusion model is lower than a preset confidence threshold, the collaborative optimization strategy generates instructions to increase the learning rates of the text coding subnetwork and the image coding subnetwork of the multi-source fusion model. When the prediction error of the demand decay trend exceeds a preset error threshold, the collaborative optimization strategy generates instructions to adjust the dropout rate, learning rate, or gradient pruning threshold of the bidirectional long short-term memory network layer of the demand evolution prediction model; or generates instructions to enable the model integration mechanism to increase the number of parallel sub-networks. When the prediction deviation of the user lifecycle state exceeds a preset deviation threshold, the collaborative optimization strategy generates instructions to adjust the learning rate of the input embedding layer or enable the auxiliary embedding branch.

7. The method for mining implicit needs of private community users based on multi-dimensional semantic fusion according to claim 2, characterized in that, The step of mapping the strategy execution feedback data onto a spatiotemporal coordinate system with the community channel as the spatial dimension and time as the temporal dimension based on the user's spatiotemporal behavior information to obtain a user behavior spatiotemporal correlation graph includes: The spatial coordinates are determined based on the location information of the user node in the community channel, and the time coordinates are determined based on the topic participation timestamp. The demand identification results, user behavior health index, and historical demand handling records in the strategy execution feedback data are used as node attribute features and mapped to a spatiotemporal coordinate system to form spatiotemporally distributed nodes. Based on the interaction link information, directed association edges are established between spatiotemporally distributed nodes, and the weight of the directed association edges is determined by the interaction frequency and the interaction delay time. The spatiotemporal correlation graph of user behavior is constructed based on spatiotemporal distributed nodes and directed related edges.

8. The method for mining implicit needs of private community users based on multi-dimensional semantic fusion according to claim 2, characterized in that, The demand causal analysis model includes: The graph construction layer is used to construct a heterogeneous graph based on the user's spatiotemporal behavior information. The nodes include user nodes, content nodes, and channel nodes, and the edges include interaction edges, temporal edges, and co-occurrence edges. The graph attention network GAT layer has 4 attention heads and is used to aggregate neighbor node information through a multi-head graph attention mechanism. Temporal convolutional layers are used to perform one-dimensional convolution on node features along the time axis to capture the temporal pattern of demand propagation. The causal inference layer is used to backpropagate based on the attention weights of the GAT layer, identify the demand-inducing nodes and propagation paths, and generate the demand tracing report.

9. The method for mining implicit needs of private community users based on multi-dimensional semantic fusion according to claim 3, characterized in that, The step of updating the preset implicit requirement content based on the requirement tracing report includes: Extract high-frequency demand-inducing nodes and key propagation paths from the demand sourcing report; Update the set of text descriptions in the preset implicit demand content according to the high-frequency demand trigger nodes; The updated implicit requirement content is sent to the edge analysis layer. The edge analysis layer expands the description templates in the requirement type retrieval library based on the updated implicit requirement content, or adds keyword matching rules that can identify requirement types, so as to optimize the next round of preliminary identification of implicit requirements.

10. A system for mining the implicit needs of users in private domain communities based on multi-dimensional semantic fusion, characterized in that: include: The acquisition unit is used to acquire multi-dimensional community data and user spatiotemporal behavior information; The preliminary identification unit is used to perform preliminary identification of implicit needs on the multi-dimensional community data at the edge analysis layer based on a preset lightweight multimodal identification model, and obtain preliminary needs identification results. The fusion unit is used to fuse the preliminary demand identification results and the multi-dimensional community data based on a preset multi-source fusion model at the cloud platform layer to obtain a comprehensive demand evaluation result. The sequence determination unit is used to assign tracking identifiers to user behavior paths in the behavior log data and determine continuous behavior sequences of the same user behavior path by combining user spatiotemporal behavior information. An evaluation unit is used to perform multidimensional state evolution evaluation based on the continuous behavior sequence to obtain a multidimensional health feature vector; The prediction unit is used to predict the user's life cycle status, demand decay trend and user behavior health index based on a preset demand evolution prediction model, taking the continuous behavior sequence and the multidimensional health feature vector as input. The triggering unit is used to determine the demand response level based on the comprehensive demand evaluation result, the demand decay trend, the user behavior health index, and the user lifecycle status, and to trigger the target operation strategy corresponding to the demand response level. The sending unit is used to acquire the feedback behavior data of the user terminal to the touch signal, dynamically adjust the judgment threshold of the demand response level according to the feedback behavior data, and send the adjusted judgment threshold as a demand detection parameter to the edge analysis layer.

Images