E-commerce customer behavior prediction method based on big data analysis

Abstract

The invention discloses an e-commerce customer behavior prediction method based on big data analysis. The method comprises the following steps: S1, acquiring original behavior data of e-commerce customers from a plurality of heterogeneous data sources in real time; s2, fusing and preprocessing the original behavior data, and constructing a unified customer behavior portrait data set which takes a customer ID as a main key and comprises time sequence features and static features; s3, based on a business target driven by new quality productivity, defining a customer behavior tag to be predicted; and S4, constructing a hybrid prediction model integrating multi-modal deep learning and a graph neural network. The method has the beneficial effects that multi-source heterogeneous data are fused, and the hybrid model integrating an attention mechanism, the graph neural network and multi-task learning is innovatively constructed; according to the method, the time sequence evolution rule and network association characteristics of customer behaviors can be captured at the same time, and accurate and collaborative prediction of multi-dimensional behaviors such as short-term conversion, long-term value, loss risks and potential demands is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electronic commerce artificial intelligence, in particular to an electronic commerce customer behavior prediction method based on big data analysis. BACKGROUND

[0002] With the rapid development of the Internet and mobile communication technology, electronic commerce has become an important part of the global economy. The increasingly competitive market environment makes electronic commerce enterprises no longer rely solely on the competition of product prices and types, but turn to customer-centric fine operation. Understanding customer intentions, predicting their future behavior, and providing highly personalized experiences have become the key to improving customer satisfaction, enhancing customer loyalty, and ultimately achieving business value growth.

[0003] Traditional methods based on customer behavior analysis mainly rely on statistical models (such as logistic regression, collaborative filtering) or early machine learning models (such as decision trees, support vector machines). These methods have certain effects when dealing with structured historical transaction data or simple user portraits. However, in the face of today's complex customer behavior, their limitations are increasingly evident: first, they usually process a single type or source of data, making it difficult to effectively integrate and utilize multi-source heterogeneous data from web pages, APPs, social media, customer service conversations, and other sources, and lack a comprehensive understanding of customers; second, traditional methods are difficult to effectively capture and model complex, non-linear temporal dependencies (such as interest evolution paths in browsing sequences) and networked relationships (such as customer-goods complex networks formed through common purchases or social relationships) in customer behavior. This limits the accuracy, timeliness, and explainability of the prediction, and cannot meet the needs of fine operation.

[0004] In recent years, the new generation of information technology represented by deep learning and big data technology has developed rapidly, injecting strong impetus into the formation of new productivity. The essence of "new productivity" lies in the use of technological innovation to drive industrial transformation, and its core characteristics are high-level data factor empowerment, deep integration of digital technologies such as artificial intelligence, and a leap in total factor productivity. In the field of electronic commerce, this means that advanced methods need to be built that can deeply integrate global data, have advanced intelligent analysis capabilities, and directly drive automated decision-making and value creation.

[0005] In the prior art, although there are researches trying to apply convolutional neural networks to process image-type commodity information, or using recurrent neural networks to analyze time-series click streams, the model structure is often single, the prediction target is isolated, and a complete, unified and adaptive intelligent system cannot be formed from multi-modal data fusion, multi-relation network modeling to multi-task collaborative prediction. Especially in terms of the self-evolution ability of the model and the interpretability of the prediction results, the existing solutions generally have deficiencies, and cannot meet the requirements of the rapid changes of e-commerce business scenarios and the transparency of decision-making. Therefore, we propose an e-commerce customer behavior prediction method based on big data analysis. SUMMARY

[0006] The purpose of the present application is to provide an e-commerce customer behavior prediction method based on big data analysis to solve the problems raised in the background art.

[0007] To achieve the above-mentioned purpose, the e-commerce customer behavior prediction method based on big data analysis comprises the following steps:

[0008] S1, real-time collection of original behavior data of e-commerce customers from multiple heterogeneous data sources, the original behavior data including but not limited to page browsing sequence, commodity interaction log, transaction record, social media mention and customer service conversation text;

[0009] S2, fusion and preprocessing of the original behavior data, construction of a unified customer behavior portrait data set with customer ID as the primary key and containing time-series features and static features;

[0010] S3, based on the business target driven by new quality productivity, definition of the customer behavior label to be predicted, the behavior label including short-term conversion intention, long-term value level, loss risk probability and potential demand category;

[0011] S4, construction of a hybrid prediction model integrating multi-modal deep learning and graph neural network, the model using attention mechanism to process time-series behavior sequence and using graph structure to analyze the complex correlation between customers, commodities and customers;

[0012] S5, training and optimization of the hybrid prediction model using the processed customer behavior portrait data set, and continuous incremental updating of the model using online learning technology;

[0013] S6, input of real-time or batch processing customer data into the trained hybrid prediction model, and output of multi-dimensional behavior prediction results with confidence for each customer;

[0014] S7, dynamic injection of the prediction results into the personalized recommendation engine, precision marketing system and customer relationship management platform of the e-commerce platform, driving automated and intelligent business decisions.

[0015] Preferably, in step S1, data is collected in real time from multiple heterogeneous data sources, specifically including: collecting client behavior flow data through a tracking SDK, collecting server-side transaction and business data through a log collection tool, and synchronizing external data from third-party social media and customer relationship management systems through API interfaces to form a global data lake.

[0016] Preferably, step S2, which involves constructing a unified customer behavior profile dataset, further includes:

[0017] S21. Use natural language processing technology to perform sentiment analysis and key information extraction on text data;

[0018] S22. Use the sliding window technique to construct time-series features with time decay weights for behavioral sequence data;

[0019] S23. Place the customer in a heterogeneous information network and extract its graph embedding features as a node.

[0020] Preferably, in step S4, the hybrid prediction model includes a temporal feature extraction module, a graph feature extraction module, and a multi-task learning head;

[0021] The temporal feature extraction module employs a long short-term memory network with an attention mechanism or a Transformer encoder to capture long-term and key patterns in customer behavior sequences.

[0022] The graph feature extraction module uses a graph neural network to learn aggregated domain information of customers based on customer-product interaction bipartite graphs or customer social similarity graphs.

[0023] The multi-task learning head shares the aforementioned features and outputs predictions of multiple behavior labels in parallel, thereby leveraging the correlation between tasks to improve the model's generalization ability.

[0024] Preferably, in step S5, the model is continuously and incrementally updated using online learning technology. Specifically, a model performance monitoring and data drift detection mechanism is established. When a decline in prediction performance or a significant change in data distribution is detected, the model parameters are automatically fine-tuned using the latest batch of data to achieve autonomous evolution of the model.

[0025] Preferably, step S7, which drives automated and intelligent business decisions, specifically includes:

[0026] Based on short-term conversion intention predictions, adjust product rankings and ad content in the customer's current session in real time; based on long-term value levels and potential demand categories predictions, develop differentiated customer lifecycle management strategies and cross-category marketing plans; based on churn risk probability predictions, automatically trigger high-value customer retention intervention processes.

[0027] As a preferred option, it also includes S8, an interpretive analysis module for establishing prediction results, which uses SHAP or LIME technology to perform attribution analysis on individual prediction outputs of the hybrid prediction model and generate an understandable customer behavior motivation report.

[0028] As a preferred approach, when calculating the Long Short-Term Memory (LSTM) network unit, for the customer behavior sequence X = (x1, x2, ..., x...),... T The Long Short-Term Memory (LSTM) network units at each time step t are calculated as follows:

[0029] Forgotten Gate: f t =σ(W f ·[h t-1 ,x t ]+b f );

[0030] Input gate: i t =σ(W i ·[h t-1 ,x t ]+b i );

[0031] Candidate memory units:

[0032] Memory unit update:

[0033] Output gate: O t =σ(W O ·[h t-1 ,x t ]+b O );

[0034] Hidden state output: h t =O t ⊙tanh(C t );

[0035] Where σ is the sigmoid function; ⊙ is element-wise multiplication; W and b are trainable parameters; h t The hidden state of the Long Short-Term Memory network at time step t carries the sequence information up to the current moment; C t The long short-term memory network cell state at time step t is responsible for transmitting long-term information in the sequence;

[0036] To identify key behaviors in the sequence, the hidden state sequence H = (h1, h2, ..., h) output by the Long Short-Term Memory network is used. T Apply attention to:

[0037] Attention Score:

[0038] Attention weights:

[0039] Weighted context vector:

[0040] Here, vector c is the temporal feature representation that incorporates attention weights, exp(·) is the natural exponential function, and tanh(·) is the hyperbolic tangent activation function.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] 1. This invention integrates multi-source heterogeneous data and innovatively constructs a hybrid model that integrates attention mechanism, graph neural network and multi-task learning. It can simultaneously capture the temporal evolution pattern and networked correlation characteristics of customer behavior, and achieve accurate and collaborative prediction of multi-dimensional behaviors such as short-term conversion, long-term value, churn risk and potential demand.

[0043] 2. This invention seamlessly integrates real-time prediction results into business systems such as recommendation, marketing, and CRM, which can automatically drive a series of business decisions such as adjusting product rankings, formulating personalized marketing strategies, and intervening in the retention of high-value customers. This not only greatly improves the speed and accuracy of operational response, but also achieves the automation and intelligent innovation of productivity by replacing repetitive manual judgment with machine intelligence.

[0044] 3. This invention introduces online learning and data drift detection mechanisms, enabling the model to continuously self-optimize based on new data and maintain long-term predictive performance effectiveness, reflecting the typical characteristic of "self-evolution" in new-quality productivity. Simultaneously, by integrating an interpretability analysis module, "black box" predictions are transformed into transparent, attributable business insights, significantly improving the credibility and operability of intelligent decision-making. Attached Figure Description

[0045] Figure 1 This is a flowchart of the method modules of the present invention. Detailed Implementation

[0046] 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 embodiments of the present invention, and not all embodiments. 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.

[0047] Please see Figure 1 This invention provides a technical solution: a method for predicting e-commerce customer behavior based on big data analysis, comprising the following steps:

[0048] S1. Collect raw behavioral data of e-commerce customers in real time from multiple heterogeneous data sources. The raw behavioral data includes, but is not limited to, page browsing sequences, product interaction logs, transaction records, social media mentions, and customer service conversation text.

[0049] Data is collected in real time from multiple heterogeneous data sources, specifically including: collecting client behavior flow data through the event tracking SDK, collecting server-side transaction and business data through log collection tools, and synchronizing external data from third-party social media and customer relationship management systems through API interfaces to form a global data lake.

[0050] It should be noted that client-side behavioral data is collected in the website and mobile applications using self-developed or third-party event tracking SDKs. This data includes real-time collection of user event streams such as page views, clicks, search terms, dwell times, and swipe patterns, which are then sent to the data center in JSON format via a message queue. Server-side business data is collected using log collection tools, including structured and unstructured logs such as order transaction logs, payment records, inventory changes, and promotional activity participation records. External ecosystem data is collected through open API interfaces provided by various platforms, synchronizing customer service tickets and communication records from social media and customer relationship management systems (CRM) on a scheduled or triggered basis.

[0051] S2. The raw behavioral data is fused and preprocessed to construct a unified customer behavior profile dataset with customer ID as the primary key and containing time-series and static features.

[0052] S21. Use natural language processing techniques to perform sentiment analysis and extract key information from text data; S22. Use sliding window techniques to construct temporal features with time decay weights for behavioral sequence data.

[0053] It should be noted that all collected data flows into a global data lake built on Hadoop or object storage in real time or near real time, is stored in subject-domain partitions, and is linked together from different sources through ETL operations such as data cleaning, deduplication, and timestamp alignment. For example, the clickstream in a session can be linked to the final generated order.

[0054] For customer service conversation text, product reviews, and social media mentions, pre-trained BERT or ERNIE models are used for sentiment analysis and key information extraction. Sliding windows are used for statistical analysis of behavioral sequence data, and a time decay function is introduced to give higher weight to recent behaviors, thus constructing feature vectors with temporal weights. A heterogeneous information network is constructed, with node types including customer, product, brand, and category; edge types include "purchase," "browse," "belong to," and "mention." An offline graph computation framework is used to pre-compute basic graph embedding features for each customer node, representing its structural position in the global network.

[0055] S3. Based on business objectives driven by new quality productivity, define customer behavior tags to be predicted. These tags include short-term conversion intention, long-term value level, churn risk probability, and potential demand categories.

[0056] It should be noted that short-term conversion intentions aim to capture customers' immediate purchase signals in the current session or intake, and improve the transaction conversion rate and the return on investment of marketing resources during real-time interaction. From the perspective of new quality productivity, short-term conversion intentions directly affect the optimized allocation of real-time production factors, such as dynamically adjusting web page content, advertising placement and customer service priorities based on predictions, so as to realize the real-time command of business flow by data flow.

[0057] In the business objectives of Long-Term Value Rank, the refined management of customer assets and the differentiated allocation of resources go beyond a single transaction. It assesses and predicts the total value of customers from the perspective of the entire life cycle, providing a basis for differentiated services, loyalty programs and high-end resource allocation. From the perspective of new quality productivity, Long-Term Value Rank reflects the quantitative and forward-looking management of customer capital as a new type of production factor.

[0058] Identifying customers with a tendency to churn in advance, especially high-value customers, enables businesses to take cost-effective interventions to reduce customer churn rates and protect existing customer assets.

[0059] S4. Construct a hybrid prediction model integrating multimodal deep learning and graph neural networks. This model uses an attention mechanism to process temporal behavior sequences and uses graph structures to analyze the complex relationships between customers, products, and customers. The hybrid prediction model includes a temporal feature extraction module, a graph feature extraction module, and a multi-task learning head.

[0060] The temporal feature extraction module employs a long short-term memory network with an attention mechanism or a Transformer encoder to capture long-term and key patterns in customer behavior sequences; the graph feature extraction module uses a graph neural network to learn aggregated domain information of customers based on customer-product interaction bipartite graphs or customer social similarity graphs; the multi-task learning head shares the above features and outputs predictions of multiple behavioral labels in parallel to improve the model's generalization ability by leveraging the correlation between tasks.

[0061] It should be noted that when calculating Long Short-Term Memory (LSTM) network units, for the customer behavior sequence X = (x1, x2, ..., x...),... T The Long Short-Term Memory (LSTM) network units at each time step t are calculated as follows:

[0062] Forgotten Gate: f t =σ(W f ·[h t-1 ,x t ]+bf );

[0063] Input gate: i t =σ(W i ·[h t-1 ,x t ]+b i );

[0064] Candidate memory units:

[0065] Memory unit update:

[0066] Output gate: O t =σ(W O ·[h t-1 ,x t ]+b O );

[0067] Hidden state output: h t =O t ⊙tanh(C t );

[0068] Where σ is the sigmoid function; ⊙ is element-wise multiplication; W and b are trainable parameters; h t The hidden state of the Long Short-Term Memory network at time step t carries the sequence information up to the current moment; C t The long short-term memory network cell state at time step t is responsible for transmitting long-term information in the sequence;

[0069] To identify key behaviors in the sequence, the hidden state sequence H = (h1, h2, ..., h) output by the Long Short-Term Memory network is used. T Apply attention to:

[0070] Attention Score:

[0071] Attention weights:

[0072] Weighted context vector:

[0073] Wherein, vector c is the temporal feature representation that incorporates attention weights, exp(·) is the natural exponential function, and tanh(·) is the hyperbolic tangent activation function;

[0074] In the graph feature extraction module, let the customer-product interaction graph be G = (V, E), where V contains customer node u and product node i. For the feature update of customer node u at the k-th layer:

[0075] Neighbor aggregation: The AGGREGATE function can be average pooling, max pooling, or LSTM aggregation, etc.

[0076] Feature update: After K layers of iteration That is, the graph embedding feature of node u, which encodes the interaction structure information in its multi-hop neighborhood;

[0077] In multi-task learning and prediction output, during feature fusion, the temporal feature c is combined with the graph embedding feature. And the fusion of the customer's static features:

[0078]

[0079] In a multi-task prediction head, for M prediction tasks, the feature z is shared, but each task has an independent output layer:

[0080] Predicted output for task m: Among them, g m (·) is an activation function adapted to the task;

[0081] In the loss function, the total loss is a weighted sum of the losses from each task, used for joint optimization:

[0082]

[0083] Where, λ m As task weight, It could be cross-entropy loss or mean squared error.

[0084] S5. Use the processed customer behavior profile dataset to train and optimize the hybrid prediction model, and utilize online learning technology to further refine the model.

[0085] Continuous incremental updates;

[0086] The model is continuously and incrementally updated using online learning technology. Specifically, a model performance monitoring and data drift detection mechanism is established. When a decline in prediction performance or a significant change in data distribution is detected, the model parameters are automatically fine-tuned using the latest batch of data, thereby achieving autonomous evolution of the model.

[0087] It should be noted that a continuous learning strategy based on mini-batch stochastic gradient descent is adopted, when the new data batch D... new Upon arrival, the update formula for the model parameter θ is:

[0088]

[0089] Where η is the learning rate. To prevent catastrophic forgetting, an Elastic Weight Consolidation (EWC) regularization term is introduced:

[0090]

[0091] Among them, F i The parameter θ is from the old task data. i The diagonal elements of the Fisher information matrix, where Υ is the regularization strength.

[0092] S6. Input real-time or batch-processed customer data into the trained hybrid prediction model and output multi-dimensional behavioral prediction results with confidence for each customer.

[0093] It should be noted that when a customer generates a new behavioral event on the platform (such as clicking on a product or searching), the event triggers a prediction request. The system immediately retrieves the customer's latest contextual data and updated static features from the cache or real-time database, and combines them with pre-loaded graph structure information to construct a formatted prediction sample. The preprocessed sample data is input into a hybrid prediction model deployed as a high-performance API service. The model synchronously outputs structured prediction results across four dimensions, each with a confidence score.

[0094] Short-term conversion intention: Output value The confidence level represents the conversion probability and can be provided by the model calibration layer or characterized by the predicted probability itself.

[0095] Long-term value ranking: Output a probability distribution vector P2 = [P 高 P 中高 P 中 P 低 The category with the highest probability is taken as the prediction level, and its maximum probability value max(P2) is taken as the confidence level of the prediction.

[0096] Churn Risk Probability: Output Value Indicates the probability of churn;

[0097] Potential demand categories: Output a probability vector P4 = [P 品类1 P 品类2 [, ...], returns the Top-K categories with the highest probabilities and their corresponding probability values, where the probability value is the confidence level.

[0098] S7. Dynamically inject the prediction results into the personalized recommendation engine, precision marketing system and customer relationship management platform of the e-commerce platform to drive automated and intelligent business decisions;

[0099] Driving automated and intelligent business decisions includes:

[0100] Based on short-term conversion intention predictions, adjust product rankings and ad content in the customer's current session in real time; based on long-term value levels and potential demand categories predictions, develop differentiated customer lifecycle management strategies and cross-category marketing plans; based on churn risk probability predictions, automatically trigger high-value customer retention intervention processes.

[0101] S8. Establish an interpretive analysis module for the prediction results, using SHAP or LIME technology to perform attribution analysis on individual prediction outputs of the hybrid prediction model, and generate an understandable customer behavior motivation report.

[0102] It should be noted that for a single prediction SHAP value φ j The explanation is provided by calculating the marginal contribution of feature j, and its core idea originates from the Shapley value in cooperative game theory:

[0103]

[0104] Where p is the total number of features, and S is the feature subset. This is the expected output of the model when only a subset of features S is used. This quantifies the direction and magnitude of the contribution of each feature (such as "browsed high-priced items last night" or "had similar purchases with high-value customers") to the final prediction (such as "high conversion intention").

[0105] In this invention, the system first aggregates heterogeneous customer behavior data from multiple sources through a unified data pipeline, constructing a comprehensive customer profile that integrates temporal, static, and graph structure features. Next, a hybrid model combining a Long Short-Term Memory (LSTM) network with an integrated attention mechanism and a Graph Neural Network (GNN) mines the evolutionary patterns of behavioral sequences from a temporal dimension and analyzes the correlation features of the complex customer-product network from a spatial dimension. Then, a multi-task learning framework collaboratively outputs multi-dimensional behavioral predictions. Finally, the prediction results are injected into the business system in real time via API, driving automated decisions such as personalized recommendations, precision marketing, and customer retention. Simultaneously, through online learning mechanisms and interpretability analysis, continuous model optimization and a trustworthy and transparent decision-making process are achieved, ultimately forming a complete productivity loop from data perception to intelligent decision-making and system evolution.

[0106] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting e-commerce customer behavior based on big data analysis, characterized in that, Includes the following steps: S1. Collect raw behavioral data of e-commerce customers in real time from multiple heterogeneous data sources. The raw behavioral data includes, but is not limited to, page browsing sequences, product interaction logs, transaction records, social media mentions, and customer service conversation text. S2. The raw behavioral data is fused and preprocessed to construct a unified customer behavior profile dataset with customer ID as the primary key and containing time-series and static features. S3. Based on the business objectives driven by new quality productivity, define the customer behavior tags to be predicted. The behavior tags include short-term conversion intention, long-term value level, churn risk probability and potential demand categories. S4. Construct a hybrid prediction model that integrates multimodal deep learning and graph neural networks. This model uses attention mechanism to process temporal behavior sequences and uses graph structure to analyze the complex relationship between customers, products and customers. S5. Use the processed customer behavior profile dataset to train and optimize the hybrid prediction model, and use online learning technology to continuously and incrementally update the model. S6. Input real-time or batch-processed customer data into the trained hybrid prediction model and output multi-dimensional behavioral prediction results with confidence for each customer. S7. Dynamically inject the prediction results into the personalized recommendation engine, precision marketing system and customer relationship management platform of the e-commerce platform to drive automated and intelligent business decisions.

2. The e-commerce customer behavior prediction method based on big data analysis according to claim 1, characterized in that, In step S1, data is collected in real time from multiple heterogeneous data sources, specifically including: collecting client behavior flow data through the event tracking SDK, collecting server-side transaction and business data through the log collection tool, and synchronizing external data from third-party social media and customer relationship management systems through API interfaces to form a global data lake.

3. The e-commerce customer behavior prediction method based on big data analysis according to claim 1, characterized in that, In step S2, constructing a unified customer behavior profile dataset further includes: S21. Use natural language processing technology to perform sentiment analysis and key information extraction on text data; S22. Use the sliding window technique to construct time-series features with time decay weights for behavioral sequence data; S23. Place the customer in a heterogeneous information network and extract its graph embedding features as a node.

4. The e-commerce customer behavior prediction method based on big data analysis according to claim 1, characterized in that, In step S4, the hybrid prediction model includes a temporal feature extraction module, a graph feature extraction module, and a multi-task learning head; The temporal feature extraction module employs a long short-term memory network with an attention mechanism or a Transformer encoder to capture long-term and key patterns in customer behavior sequences. The graph feature extraction module uses a graph neural network to learn aggregated domain information of customers based on customer-product interaction bipartite graphs or customer social similarity graphs. The multi-task learning head shares the aforementioned features and outputs predictions of multiple behavior labels in parallel, thereby leveraging the correlation between tasks to improve the model's generalization ability.

5. The e-commerce customer behavior prediction method based on big data analysis according to claim 1, characterized in that, In step S5, the model is continuously and incrementally updated using online learning technology. Specifically, a model performance monitoring and data drift detection mechanism is established. When a decline in prediction performance or a significant change in data distribution is detected, the model parameters are automatically fine-tuned using the latest batch of data to achieve autonomous evolution of the model.

6. The e-commerce customer behavior prediction method based on big data analysis according to claim 1, characterized in that, In step S7, driving automated and intelligent business decisions specifically includes: Based on short-term conversion intention predictions, adjust product rankings and ad content in the customer's current conversation in real time; based on long-term value levels and potential demand categories predictions, develop differentiated customer lifecycle management strategies and cross-category marketing plans; based on churn risk probability predictions, automatically trigger high-value customer retention intervention processes.

7. The e-commerce customer behavior prediction method based on big data analysis according to claim 1, characterized in that, It also includes S8, an interpretive analysis module for forecast results, which uses SHAP or LIME technology to perform attribution analysis on individual forecast outputs of the hybrid forecast model and generate understandable customer behavior motivation reports.

8. The e-commerce customer behavior prediction method based on big data analysis according to claim 4, characterized in that, When calculating Long Short-Term Memory (LSTM) network units, for the customer behavior sequence X = (x1, x2, ..., x...), ... T The Long Short-Term Memory (LSTM) network units at each time step t are calculated as follows: Forgotten Gate: f t =σ(W f ·[h t-1 ,x t ]+b f ); Input gate: i t =σ(W i ·[h t-1 ,x t ]+b i ); Candidate memory units: Memory unit update: Output gate: O t =σ(W O ·[h t-1 ,x t ]+b O ); Hidden state output: h t =O t ⊙tanh(C t ); Where σ is the sigmoid function; ⊙ is element-wise multiplication; W and b are trainable parameters; h t The hidden state of the Long Short-Term Memory network at time step t carries the sequence information up to the current moment; C t The long short-term memory network cell state at time step t is responsible for transmitting long-term information in the sequence; To identify key behaviors in the sequence, the hidden state sequence H = (h1, h2, ..., h) output by the Long Short-Term Memory network is used. T Apply attention to: Attention Score: Attention weights: Weighted context vector: Here, vector c is the temporal feature representation that incorporates attention weights, exp(·) is the natural exponential function, and tanh(·) is the hyperbolic tangent activation function.

Images