TL;DR
This paper explores how to effectively group e-commerce customers based on their website activity. The authors demonstrate that grouping customers by the services they request, their navigation patterns, or their resource usage yields surprisingly similar results, suggesting a unified approach to understanding customer behavior for better e-commerce management.
What Problem Does This Paper Address?
In the world of e-commerce, understanding customer behavior is crucial for everything from optimizing server performance and capacity planning to personalizing user experiences and improving revenue. While traditional commerce allows for direct observation of customer groups, online interactions are more abstract. Researchers have previously used clustering techniques to group customers based on various data metrics derived from their website sessions (sequences of requests). However, these studies often analyzed different problems (e.g., resource usage vs. revenue optimization) using separate grouping criteria, leading to a fragmented understanding. The core problem this paper addresses is the lack of a comprehensive comparison of different customer grouping methods and the need to understand if these distinct approaches lead to substantially different customer segments, which would complicate unified e-commerce management.
Methods
The paper introduces a hybrid clustering algorithm that combines the Minimum Spanning Tree (MST) method with the k-means algorithm. This approach aims to leverage the strengths of both: MST helps in selecting appropriate centroids, while k-means efficiently groups sessions around these centroids.
The authors define three key criteria for characterizing customer sessions:
- Services Requested: Sessions are represented by a vector indicating which distinct web pages (services) were requested.
- Navigation Pattern: Sessions are characterized by the sequence of moves (transitions between consecutive page requests). This creates a higher-dimensional space compared to services requested.
- Resource Usage: Sessions are described by their total System Response Time (SRT) and average SRT per request. SRT is approximated using the median of observed response times, assuming a lightly loaded server.
To handle the high dimensionality, especially for navigation patterns, the paper employs techniques like merging web pages by functionality and representing sessions with only non-zero dimensions.
Datasets
The study uses web server logs from an e-rental business. The logs cover a 24-hour period and are in W3C extended log format. The data was preprocessed to filter out browser-generated requests for embedded objects, focusing on explicit user requests.
Key Insights and Ideas
- Interconnectedness of Customer Behavior: A significant finding is that grouping customers by services requested, navigation patterns, or resource usage yields remarkably similar results. This suggests that these aspects of customer behavior are highly correlated.
- Unified Approach is Feasible: Because the groupings are similar, it's often sufficient to use just one criterion for analysis. This simplifies the process of understanding customer segments for various e-commerce management tasks.
- Service-Based Grouping is Recommended: The authors advocate for grouping by services requested as the preferred method. It's simpler to implement, requires less complex data processing, and directly reflects user intent.
- E-rental Site Specifics: The study highlights that the structure of an e-rental site naturally leads to certain customer interactions, like focusing on a single item (vehicle) with multiple parameters, which differs from broader e-commerce sites like bookstores.
Key Results
The study identified four distinct customer group types across all three grouping methods:
- Rate-Checkers: Users who quickly check rates and leave, often driven by external search facilities. They have very short sessions and low interaction depth.
- Browsers: Users who explore information without a clear objective, visiting various informational pages. They have a moderate session length and low resource consumption.
- Confirmers: Users interested in reservations, spending more time browsing details, checking rates, and available vehicles, but with a low reservation conversion rate.
- Buyers: Users who take their time to gather detailed information, search, check rates, and make reservations, exhibiting longer sessions and higher engagement.
Table 3 in the paper quantifies the "degree of association" between groups derived from different metrics. For instance, 95.5% of sessions in the "service-based" group srvG-0 are also found in the "navigation-based" group nvgG-0, and vice-versa for 88.2%. This strong overlap supports the idea that the underlying customer segments are consistent across different measurement approaches.
Limitations and Future Work
- Data Source: The study is based on data from a single e-rental website. While the findings are presented as generally applicable, further verification on other e-commerce sites is needed.
- Resource Usage Approximation: The approximation of System Response Time (SRT) using median values is based on the assumption of a lightly loaded server. The accuracy of this approximation might be affected in scenarios with higher server load.
- Clustering Algorithm Complexity: While the hybrid MST-k-means approach is more efficient than pure MST, the computational complexity, especially for navigation patterns with many dimensions, can still be a challenge.
- Future Work: The authors suggest extending this analysis to more general-purpose e-commerce sites to validate their findings. They also propose further characterizing these session groups by building Customer Behavior Model Graphs (CBMGs) for each group and using this detailed workload characterization for system performance analysis.
Summary updated