TL;DR
This paper introduces the Golden Path Analyzer (GPA), a novel algorithm for clustering web user clickstreams by identifying "golden paths" – shortest, successful user journeys. GPA effectively groups users into semantically meaningful clusters, offering actionable insights for website designers to improve user experience and task completion.
What Problem Does This Paper Address?
Understanding how users navigate websites and complete tasks is crucial for effective design. Existing methods for analyzing user clickstreams often struggle with the complexity and variability of user behavior. Many systems either:
- Exclude users who deviate from predefined paths: This leads to a loss of valuable data and an incomplete picture of user journeys.
- Focus on short, fragmented subpaths: These don't provide a complete understanding of a successful task completion.
- Require strict inclusion criteria: This can result in too many small, difficult-to-interpret clusters.
The authors aim to address these limitations by developing a system that can:
- Identify the most common and successful paths users take to complete a specific task.
- Group users into a small number of easily understandable clusters based on these paths.
- Assign a larger proportion of users to these clusters, even if they make minor diversions.
- Provide actionable insights into why users succeed or fail, and where they get distracted.
Methods
The core of the Golden Path Analyzer (GPA) is a divide-and-conquer approach to clickstream clustering. The key innovation lies in using actual successful user clickstreams as "golden paths" to seed the clusters.
Here's a breakdown of the process:
- Identify "Golden Paths": GPA first identifies "golden paths," which are defined as the shortest, successful sequences of page visits that lead a user from a starting page to a page containing the information needed to complete a specific task. These paths are derived from actual user data and must be followed by a minimum number of users to be considered valid.
- Cluster Assignment: Users are assigned to a cluster if their clickstream is a supersequence of a golden path. This means users are included even if they take detours or excursions, as long as they eventually complete the sequence of pages defined by the golden path.
- Divide and Conquer Algorithm: The algorithm iteratively identifies the shortest golden path that explains a significant portion of user behavior. Once a golden path is identified, all clickstreams that are supersequences of this path are removed from consideration, and the process repeats with the remaining clickstreams. This ensures that the discovered paths are distinct and cover different user strategies.
- Additional Analyses: Beyond clustering, GPA provides detailed analyses, including:
- Falloff Analysis: Identifying pages where users deviate from a golden path.
- Fatal Falloff: Pinpointing pages that lead to user failure.
- Distractor Pages: Highlighting pages that divert users from their intended task.
- Non-recognizers and Overshooters: Identifying users who reach the target page but don't recognize it as the end, or who continue browsing after success.
Datasets
The paper uses data from a specific client engagement with FictiousCarPortal.com for a task of finding a dealer quote on a specific car model. A sample of 200 representative users was tracked using Vividence Connector. The paper does not introduce new benchmark datasets but rather demonstrates GPA's application on a real-world scenario.
Key Insights and Ideas
- Supersequence Clustering is Powerful: By defining clusters based on supersequences of golden paths, GPA can include a much larger percentage of users than methods with strict inclusion criteria, leading to more comprehensive analysis.
- Actual Paths as Seeds: Using actual, successful user paths as cluster seeds provides semantically rich and easily interpretable cluster descriptions, unlike abstract or generated paths.
- "Golden Paths" Offer Actionable Insights: The identified golden paths directly highlight optimal user journeys, making it straightforward for designers to understand what works and where improvements are needed.
- Falloff Analysis Pinpoints Pain Points: Distinguishing between different types of falloffs (immediate, path, fatal) allows for precise identification of usability issues, from confusing navigation to unclear end-of-process signals.
- Divide and Conquer for Clarity: The iterative divide-and-conquer approach ensures that the discovered golden paths represent distinct user strategies, leading to a small number of highly informative clusters.
Key Results
- For the FictiousCarPortal.com example, GPA discovered two primary golden paths that explained a significant portion of user behavior.
- 83% of successful users were explained by these two golden paths.
- The analysis identified specific pages that caused a high number of "fatal falloffs" (e.g., page A1 in GP1 and B1 in GP2), indicating critical areas for site improvement.
- The system demonstrated its ability to identify distractor pages and provide insights into user frustration (e.g., lack of qualified dealers).
- GPA is computationally efficient, typically finishing analysis in less than 10 seconds for sample sizes of 400-800 users.
Limitations and Future Work
- Parameter Sensitivity: The effectiveness of GPA relies on the careful selection of parameters (T, I, C) that define what constitutes a "golden path." The paper acknowledges that human intervention is currently required for page aggregation and parameter tuning.
- Validation: The authors suggest that further validation of the clustering approach, perhaps through comparison with other algorithms or expert-created clusterings, would strengthen the theoretical underpinnings.
- Automation: Extending GPA to a fully "hands-off" system, automating parameter selection and page aggregation, is a key area for future work.
- Anomaly Explanation: Enhancing GPA's ability to explain anomalies in user data, potentially by integrating user comments more deeply, is another promising direction.
Summary updated