Behavior analysis method and system for multi-modal data fusion and dynamic behavior prediction
Through multimodal data fusion and dynamic behavior prediction methods, the user behavior is analyzed using Transformer and LSTM models, and the prediction lag problem caused by the single data dimension in the existing technology is solved, and real-time user behavior analysis and personalized marketing strategies are realized with high accuracy.
Patent Information
- Application Number
- CN202510554133.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-29
- Publication Date
- 2025-08-15
AI Technical Summary
The existing customer behavior analysis system relies on single-dimensional historical transactions or browsing data, and lacks comprehensive analysis of users' real-time behavior, emotional state and environmental scenarios, resulting in lagging prediction results and insufficient accuracy, making it difficult to support dynamic marketing decisions.
The multimodal data fusion method is adopted to collect historical user behavior, product metadata and user basic information, and data preprocessing and predictive analysis are carried out through Transformer and LSTM models, and long-term interests are modeled in combination with self-attention mechanisms to achieve multi-dimensional data fusion and second-level behavior prediction.
It improves the depth and real-time nature of user behavior analysis, supports second-level behavior prediction, optimizes business decisions, improves personalized recommendations and inventory scheduling efficiency, reduces operating costs, and enhances user stickiness.
Smart Images
Figure CN120494880A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of big data analysis, and more specifically to a behavior analysis method and system for multimodal data fusion and dynamic behavior prediction. Background Art
[0002] Existing customer behavior analysis systems mostly rely on single-dimensional historical transaction or browsing data, and lack comprehensive analysis of users' real-time behavior, emotional state, and environmental scenarios. This leads to delayed and inaccurate prediction results, making it difficult to support dynamic marketing decisions.
[0003] The single data dimension and poor real-time prediction in the user demand analysis process are technical problems that need to be solved. Summary of the Invention
[0004] The technical task of the present invention is to address the above shortcomings and provide a behavior analysis method and system for multimodal data fusion and dynamic behavior prediction to solve the problems of single data dimension and poor real-time prediction in the process of user demand analysis.
[0005] In a first aspect, the present invention provides a behavior analysis method for multimodal data fusion and dynamic behavior prediction, comprising the following steps:
[0006] Data collection: Collect historical user behavior data, product metadata, and basic user information as raw data. User behavior data includes browsing, clicking, and add-to-cart records; product metadata includes product ID, category, and price; and basic user information includes user ID and profile.
[0007] Data preprocessing: Encode the raw data, encode the product ID and user row type, normalize the execution time corresponding to the user behavior, and use labels to mark whether the user purchased the product to obtain preprocessed data. The preprocessed data is then sorted by time to form a time series.
[0008] Predictive analysis: Using time series as input, the trained Transformer model predicts the probability of users purchasing various products as the output result. Using time series as input, the trained LSTM model predicts the probability of users purchasing various products as the output result. The output results of the Transformer model and the LSTM model are weighted and summed to obtain the final prediction result.
[0009] Preferably, when encoding the original data, the product ID is mapped to a number.
[0010] Preferably, when the pre-processed data is sorted according to time, the pre-processed data is truncated or padded to a fixed length.
[0011] Preferably, the method further comprises the following steps:
[0012] Based on the final prediction results, user risks are identified, products are pushed to users based on the final prediction results, and user behavior reports are constructed based on the final prediction results.
[0013] In a second aspect, the present invention provides a behavior analysis system for multimodal data fusion and dynamic behavior prediction, which is used to perform user behavior analysis using a behavior method for multimodal data fusion and dynamic behavior prediction as described in any one of the first aspects, wherein the system includes a data acquisition module, a data preprocessing module, and a prediction analysis module;
[0014] The data collection module is used to perform the following operations: collect historical user behavior data, product metadata, and user basic information as raw data. User behavior data includes browsing, clicking, and add-to-purchase records; product metadata includes product ID, category, and price; and user basic information includes user ID and profile.
[0015] The data preprocessing module is used to perform the following operations: encode the raw data, encode the product ID and user row type, normalize the execution time corresponding to the user behavior, mark whether the user purchased the product with a label, obtain preprocessed data, and sort the preprocessed data based on time to form a time series sequence;
[0016] The predictive analysis module is used to perform the following operations: using a time series as input and using a trained Transformer model to predict the probability of a user purchasing various products as output, and using a time series as input and using a trained LSTM model to predict the probability of a user purchasing various products as output. The output of the Transformer model and the output of the LSTM model are weighted and summed to obtain the final prediction result.
[0017] Preferably, when encoding the original data, the data preprocessing module is used to map the product ID into a number.
[0018] Preferably, when the pre-processed data is sorted according to time, the data pre-processing module is used to truncate or fill the pre-processed data to a fixed length.
[0019] Preferably, the system further includes a summary report module, which is used to perform the following: identify risks for users based on the final prediction results, push products to users based on the final prediction results, and construct a user behavior report based on the final prediction results.
[0020] The behavior analysis method and system for multimodal data fusion and dynamic behavior prediction of the present invention have the following advantages:
[0021] 1. Improve analytical depth: Multimodal data integration comprehensively depicts user behavior motivations, breaking through the limitations of a single dimension;
[0022] 2. Enhanced real-time performance: The time series model supports behavior prediction within seconds, enabling proactive services that "respond before being reached";
[0023] 3. Optimize business decisions: Dynamic forecast results directly drive personalized recommendations, inventory scheduling, and promotion strategies, reducing operating costs;
[0024] 4. Improve user stickiness: Accurate demand matching reduces user decision fatigue and increases repurchase rate. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or descriptions of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0026] The present invention will be further described below with reference to the accompanying drawings.
[0027] Figure 1 This is a flowchart of a behavior analysis method for multimodal data fusion and dynamic behavior prediction in Example 1. DETAILED DESCRIPTION
[0028] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments given are not intended to limit the present invention. Unless there is a conflict, the embodiments of the present invention and the technical features in the embodiments may be combined with each other.
[0029] The embodiments of the present invention provide a behavior analysis method and system for multimodal data fusion and dynamic behavior prediction, which are used to solve the technical problems of single data dimension and poor real-time prediction in the process of user demand analysis.
[0030] Example 1:
[0031] The present invention provides a behavior analysis method for multimodal data fusion and dynamic behavior prediction, which includes three steps: data acquisition, data preprocessing and prediction analysis.
[0032] Step S100 Data Collection: Collect historical user behavior data, product metadata, and user basic information as raw data, where user behavior data includes browsing, clicking, and purchase records, product metadata includes product ID, category, and price, and user basic information includes user ID and portrait.
[0033] In this embodiment, when encoding the original data, the product ID is mapped to a number, and when sorting the pre-processed data by time, the pre-processed data is truncated or padded to a fixed length.
[0034] Step S200: Data preprocessing: Encode the original data, encode the product ID and user row type through encoding, normalize the execution time corresponding to the user behavior, and mark whether the user purchases the product through a label to obtain preprocessed data, and sort the preprocessed data based on time to form a time series.
[0035] Step S300 Prediction Analysis: Using the time series as input, the trained Transformer model predicts the probability of users purchasing various commodities as the output result, and using the time series as input, the trained LSTM model predicts the probability of users purchasing various commodities as the output result. The output results of the Transformer model and the output results of the LSTM model are weighted and summed to obtain the final prediction result.
[0036] In this embodiment, the LSTM model performs predictive analysis to analyze short-term user behavior sequences (e.g., recent clicks, browsing, and add-to-cart) and capture temporal dependencies. Its output is a predicted probability of a user purchasing each product.
[0037] For the Transformer model, its function is to model long-term interests (such as preferences across time periods) through the self-attention mechanism, and its output is to predict the probability of users purchasing each product.
[0038] Suppose a user's behavior sequence is: [browsing phones, browsing headphones, browsing computers,..., purchasing phone accessories]. Self-attention analysis: The model may find that "browsing phones" and "purchasing phone accessories" are highly correlated (high attention weight). Even though these two behaviors are far apart, the self-attention mechanism can still directly model their relationship without relying on step-by-step transfer (such as LSTM). Result: The model infers that the user's long-term interest is "mobile phone accessories," rather than focusing solely on recent behaviors (such as "browsing computers").
[0039] The prediction result fusion strategy is weighted averaging, assigning weights based on validation set performance (e.g., 40% for LSTM and 60% for Transformer). Dynamic adjustments are made during model training as follows: weights are optimized in real time based on the business scenario. The final output is the fused probability result, generating the final purchase prediction.
[0040] As an improvement, the method of this embodiment further includes the following steps:
[0041] S400 Summary Report: Identify user risks based on the final prediction results, push products to users based on the final prediction results, and build a user behavior report based on the final prediction results.
[0042] For example, ranking users’ purchase probability (such as top-10 products) and identifying high-risk users (such as users with a high probability of not purchasing) can be done.
[0043] Application scenarios of this method include:
[0044] (1) Real-time recommendation: Push high-probability products to users.
[0045] (2) Marketing strategy: Design promotional activities based on the predicted results.
[0046] (3) User analysis: Generate behavioral insight reports (such as preference change trends).
[0047] For the two models implemented in this paper, the optimization direction during evaluation and iteration involves the following two directions:
[0048] Adjust the model structure (such as the number of LSTM layers and Transformer heads).
[0049] Update training data to adapt to changes in user behavior.
[0050] Example 2:
[0051] The present invention provides a behavior analysis system for multimodal data fusion and dynamic behavior prediction, comprising a data acquisition module, a data preprocessing module and a prediction analysis module.
[0052] The data collection module is used to perform the following: collect historical user behavior data, product metadata and user basic information as raw data, among which user behavior data includes browsing, clicking and purchase records, product metadata includes product ID, category and price, and user basic information includes user ID and portrait.
[0053] When encoding the original data, the data preprocessing module is used to map the product ID to a number. When sorting the preprocessed data by time, the data preprocessing module is used to truncate or fill the preprocessed data to a fixed length.
[0054] The data preprocessing module is used to perform the following: encode the original data, encode the product ID and user row type through encoding, normalize the execution time corresponding to the user behavior, mark whether the user purchases the product through labels, obtain preprocessed data, and sort the preprocessed data based on time to form a time series sequence.
[0055] The predictive analysis module is used to perform the following operations: using a time series as input and using a trained Transformer model to predict the probability of a user purchasing various products as output, and using a time series as input and using a trained LSTM model to predict the probability of a user purchasing various products as output. The output of the Transformer model and the output of the LSTM model are weighted and summed to obtain the final prediction result.
[0056] The system of this embodiment also includes a summary report module, which is used to perform the following: identify risks for users based on the final prediction results, push products to users based on the final prediction results, and build a user behavior report based on the final prediction results.
[0057] The system of this embodiment can execute the method disclosed in Example 1.
[0058] The above is a detailed introduction to the behavior analysis method and system for multimodal data fusion and dynamic behavior prediction provided by the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for general technical personnel in this field, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.
Claims
1. A behavior analysis method for multimodal data fusion and dynamic behavior prediction, characterized in that: The steps include: Data collection: Collect historical user behavior data, product metadata, and basic user information as raw data. User behavior data includes browsing, clicking, and add-to-cart records; product metadata includes product ID, category, and price; and basic user information includes user ID and profile. Data preprocessing: Encode the raw data, encode the product ID and user row type, normalize the execution time corresponding to the user behavior, and use labels to mark whether the user purchased the product to obtain preprocessed data. The preprocessed data is then sorted by time to form a time series. Predictive analysis: Using time series as input, the trained Transformer model predicts the probability of users purchasing various products as the output result. Using time series as input, the trained LSTM model predicts the probability of users purchasing various products as the output result. The output results of the Transformer model and the LSTM model are weighted and summed to obtain the final prediction result.
2. The behavior analysis method for multimodal data fusion and dynamic behavior prediction according to claim 1 is characterized in that: When encoding the raw data, the product ID is mapped to a number.
3. The behavior analysis method for multimodal data fusion and dynamic behavior prediction according to claim 1 is characterized in that: When sorting preprocessed data by time, truncate or pad the preprocessed data to a fixed length.
4. The behavior analysis method for multimodal data fusion and dynamic behavior prediction according to claim 1 is characterized in that: The method further comprises the steps of: Based on the final prediction results, user risks are identified, products are pushed to users based on the final prediction results, and user behavior reports are constructed based on the final prediction results.
5. A behavior analysis system for multimodal data fusion and dynamic behavior prediction, characterized by: Used for performing user behavior analysis by a behavior method of multimodal data fusion and dynamic behavior prediction according to any one of claims 1 to 4, the system comprising a data acquisition module, a data preprocessing module and a prediction analysis module; The data collection module is used to perform the following operations: collect historical user behavior data, product metadata, and user basic information as raw data. User behavior data includes browsing, clicking, and add-to-purchase records; product metadata includes product ID, category, and price; and user basic information includes user ID and profile. The data preprocessing module is used to perform the following operations: encode the raw data, encode the product ID and user row type, normalize the execution time corresponding to the user behavior, mark whether the user purchased the product with a label, obtain preprocessed data, and sort the preprocessed data based on time to form a time series sequence; The predictive analysis module is used to perform the following operations: using a time series as input and using a trained Transformer model to predict the probability of a user purchasing various products as output, and using a time series as input and using a trained LSTM model to predict the probability of a user purchasing various products as output. The output of the Transformer model and the output of the LSTM model are weighted and summed to obtain the final prediction result.
6. The behavior analysis system for multimodal data fusion and dynamic behavior prediction according to claim 5 is characterized in that: When encoding the original data, the data preprocessing module is used to map the product ID to a number.
7. The behavior analysis system for multimodal data fusion and dynamic behavior prediction according to claim 5 is characterized in that: When the preprocessed data is sorted according to time, the data preprocessing module is used to truncate or pad the preprocessed data to a fixed length.
8. The behavior analysis system for multimodal data fusion and dynamic behavior prediction according to claim 5 is characterized in that: The system also includes a summary report module, which is used to perform the following: identify risks for users based on the final prediction results, push products to users based on the final prediction results, and build a user behavior report based on the final prediction results.