Clustering Analysis of Live E-commerce User Behavior: A Practical Data Mining of Engagement Using K-Means, DBSCAN, and Other Algorithms

This article introduces a user engagement clustering analysis project for Facebook live commerce scenarios. By comparing four algorithms—K-Means, hierarchical clustering, DBSCAN, and Gaussian Mixture Model—it helps e-commerce operators identify different user groups and develop targeted marketing strategies. The project covers the complete process from data preprocessing to model evaluation, providing a practical technical reference framework for practitioners.

With the explosive growth of live e-commerce, understanding user behavior patterns in live streams has become a core operational issue. User engagement in live scenarios is highly real-time and interactive, with users completing the browse-interact-purchase link in a short time. However, under massive data, simple statistics can hardly reveal the inherent differences among users (e.g., highly interactive users may be potential buyers or onlookers; low-interactive users may be prospective new customers or silent buyers), which gives rise to the need for clustering analysis.

This project was developed by Ankita Rani Patro, building a multi-algorithm clustering framework for user engagement data in Facebook live commerce. The characteristics of the four algorithms are as follows:

• K-Means: A classic partitioning method, efficient and simple, suitable for data with spherical distribution, quickly identifying users with similar behaviors.
• Hierarchical Clustering: Builds a tree-like structure, no need to pre-specify the number of clusters, reveals hierarchical relationships, suitable for exploring user group-subgroup structures.
• DBSCAN: Density-based, automatically identifies noise points (e.g., bot accounts), discovers clusters of arbitrary shapes, and identifies core user groups.
• GMM: Probabilistic soft clustering, allows users to belong to multiple clusters with different probabilities, suitable for scenarios with fuzzy boundaries (e.g., users have both potential buyer and content enthusiast attributes).

The project follows the standard data science paradigm:

• Data Preprocessing: Process high-dimensional features (watching duration, interaction count, comment sentiment, sharing behavior, etc.), missing values, and outliers, and perform standardization (due to large differences in feature dimensions and distributions).
• Model Training: Each algorithm requires hyperparameter tuning: K-Means uses the elbow method/silhouette coefficient to determine the optimal K; DBSCAN sets eps (neighborhood radius) and min_samples (minimum number of samples); GMM determines the number of Gaussian distributions and covariance type. The optimal solution is found by comparing different parameter configurations.

Each of the four algorithms has its own advantages and disadvantages:

• K-Means: Its strengths lie in speed and interpretability, suitable for large-scale uniformly distributed data. It is used to identify high-value/ordinary/at-risk-of-churn users in real time, supporting quick decision-making.
• Hierarchical Clustering: Its value lies in the hierarchical structure, which allows exploration from macro to micro levels, helping to develop layered operation strategies.
• DBSCAN: Its uniqueness is in anomaly detection; it can filter bot/brush accounts and identify core fan groups and edge audiences.
• GMM: Soft clustering is suitable for scenarios with fuzzy boundaries, providing probabilistic attribution judgments and facilitating refined operations.

Clustering results can guide business decisions, identifying typical user groups and corresponding strategies:

• Core Purchase Type: Medium watching duration but high conversion rate, low price sensitivity → Maintain loyalty and increase average order value.
• Content Consumption Type: Long watching time, frequent interaction but few purchases → Cultivate purchase intention through content marketing.
• Impulse Purchase Type: Short watching time but quick decision-making, sensitive to limited-time offers → Create a sense of urgency with flash sales.
• Silent Observation Type: Stable watching but low interaction → Need stronger conversion incentives.

This project is for teaching and research purposes; in actual production, real-time data stream processing, temporal evolution of user behavior, user associations across multiple live streams, etc., need to be considered. Future expansion directions: Introduce autoencoders for feature learning, combine time series analysis to capture dynamic changes, and integrate clustering results with recommendation systems to achieve personalized push.