Abstract
Modern web browsing often results in a cluttered history filled with redirects, error pages, and transient visits, complicating the process of revisiting valuable content. This paper introduces a system that segments browsing history by task and employs machine learning to predict pages with intrinsic revisit value. By combining heuristic filtering with Support Vector Machine (SVM) classifiers, the system achieves approximately 78% accuracy in identifying relevant pages, significantly reducing digital noise for the user.
Introduction
Despite over 50% of web traffic consisting of revisits, current tools for returning to previously viewed pages remain inadequate. Manual bookmarking is often neglected, while standard chronological history lists suffer from "information overload," containing every URL touched, including ads and accidental clicks. This research addresses the challenge of building a "smart" history list that automatically retains only project-relevant pages.
Methodology
The proposed approach involves two primary stages: task segmentation and revisit prediction.
Task Segmentation and Heuristic Filtering
The system integrates with TaskTracer to associate browsing activity with specific user-defined tasks. Before applying machine learning, three heuristic rules are used to strip away obvious noise:
- Error Removal: Filtering pages containing keywords like "404" or "unavailable."
- De-duplication: Retaining only a single entry per unique URL.
- The 5-Second Rule: Deleting pages visited for less than five seconds, assuming they were merely navigational or immediately rejected.
Machine Learning Classifiers
Two types of models—Naive Bayes (NB) and Support Vector Machines (SVM)—were trained to classify pages as "keep" or "delete." Features included dwell time, link-following behavior, "copy" command frequency, and word frequencies. Semantic weighting was applied to words appearing in task descriptions, page titles, and URLs to enhance predictive power.
Experimental Results
The study utilized 21 datasets from 8 participants. Ground truth was established by having users manually review and clean their history after task completion.
Model Performance
On average, users deemed 34.3% of their task-specific history as "junk." The SVM model consistently outperformed Naive Bayes. While the baseline SVM achieved 74.7% accuracy, incorporating feature selection (using Information Gain for the top 100 features) improved accuracy to 78.02%.
User Variability
A key finding was the high subjectivity of "revisit value." Deletion rates varied significantly between participants, with some wanting to keep nearly all history and others trashing the majority. This highlights the necessity for personalized models that adapt to individual user behavior.
Conclusion and Future Work
The research demonstrates that task-centric history management combined with simple heuristics and ML classifiers can effectively filter web history. Future work will explore incorporating navigation paths and link structures, as well as addressing the asymmetric cost of "false negatives"—where the system accidentally deletes a page the user intended to keep.
Summary updated