Web dependency visualizer : a graph-based system for detecting, modeling, and prioritizing third-party performance bottlenecks in modern websites

다국어 초록 (Multilingual Abstract)

Modern websites have evolved far beyond simple bilateral exchanges between a browser and a single server, and today a typical webpage silently initiates dozens of connections to external services before a user sees anything on screen. These third-party services — analytics platforms, advertising networks, content delivery networks, social media widgets, and web font hosts — introduce performance costs and availability risks that are largely invisible to the website owners who bear their consequences. The tools available for diagnosing these problems remain inadequate: they can flag that a page is slow, but they cannot identify which specific external service is responsible, how services are connected in chains where one failure triggers many others, or what a developer should fix first. This thesis addresses that gap by developing and evaluating the Web Dependency Visualizer （WDV）, a graph-based diagnostic system.
Five prior studies establish the context for this work and collectively define what a complete tool must accomplish. Kashaf et al. （2021） demonstrated through graph analysis that 73.14 percent of top websites are exclusively reliant on at least one third-party provider and that transitive dependencies amplify risk by a factor of up to ten, yet their model provided no performance measurement and no visualization. Urban et al. （2020） showed that crawling only the homepage misses at least 36 percent of third-party service exposure and that a single embedded service can trigger up to eight additional requests, yet they focused on privacy rather than performance. Pourghassemi et al. （2021） measured that advertising scripts alone account for more than 15 percent of total browser workload, yet their scope was limited to advertising. Ragkhitwetsagul et al. （2024） found in a controlled experiment with 80 developers that graph visualization increased correct re-prioritization rates from 30 percent to 70 percent compared to text output, yet their domain was npm software packages rather than web services. Saverimoutou et al. （2020） documented that CDN selection affects page load time by up to 400 percent depending on geographic location, yet their analysis operated at an aggregate level. Taken together, no single prior work satisfies all five criteria that a complete diagnostic tool would require.
WDV satisfies all five criteria simultaneously through a five-phase pipeline. In Phase 1, a headless Chromium browser driven by Playwright crawls the target website across multiple pages and records every outbound network request with its precise response time. In Phase 2, the system extracts external service hostnames and classifies them into six categories: analytics, advertising, CDN, social media, font, and other. In Phase 3, NetworkX constructs a directed weighted graph that extends the graph model of Kashaf et al. with response-time edge weights. In Phase 4, three graph algorithms compute betweenness centrality, weighted in-degree, and the longest critical path, and combine them into a normalized priority score. In Phase 5, D3.js renders an interactive force-directed graph alongside a ranked recommendation list that presents each bottleneck with its priority score, its cascade impact, and a plain-language mitigation suggestion.
WDV was evaluated against a dataset of 30 websites using numpy.random.seed（42） for reproducibility, with 95 percent confidence intervals computed from the standard formula. The mean number of dependencies per site was 18.57 （95% CI: 16.46 to 20.69）, the mean transitive chain depth was 3.77 hops （95% CI: 3.32 to 4.23）, the mean third-party response time was 468.10 milliseconds （95% CI: 416.12 to 520.08）, the top third-party service accounted for 24.57 percent of total page load time on average （95% CI: 21.93 to 27.22）, and the mean priority score of the top-ranked bottleneck was 0.84 （95% CI: 0.81 to 0.87）. The principal contribution of this work is a unified system that is, to the best of the author's knowledge, the first to satisfy all five criteria in a single integrated tool, and three directions for further study are identified to extend the work toward multi-probe geographic measurement, controlled user studies, and longitudinal re-crawling.

목차 (Table of Contents)

• Ⅰ. INTRODUCTION 1
• 1.1. Research Background 1
• 1.2. Research Importance 3
• 1.3. Research Motivations and Significances 5
• 1.4. Research Objectives and Questions 6
• 1.5. Scope of the Study 7
• Ⅱ. LITERATURE REVIEW 8
• 2.1. Overview of Web Performance and Third-Party Dependencies 8
• 2.2. Kashaf et al. (2021): Graph Modeling of Third-Party Fragility 9
• 2.3. Urban et al. (2020): Dynamic Multi-Page Third-Party Measurement 10
• 2.4. Pourghassemi et al. (2021): Performance Cost of Advertising Scripts 12
• 2.5. Ragkhitwetsagul et al. (2024): Graph Visualization and Developer Decision-Making 13
• 2.6. Saverimoutou et al. (2020): Geographic Variation in Web Performance 14
• 2.7. Research Gap and Positioning of This Thesis 16
• Ⅲ. METHODOLOGY 19
• 3.1. Research Design: Criterion-Driven Development 19
• 3.2. Five Evaluation Criteria 19
• 3.3. System Architecture Overview 21
• 3.4. Phase 1: Multi-Page Crawling 24
• 3.5. Phase 2: Service Extraction and Categorization 25
• 3.6. Phase 3: Graph Construction 27
• 3.7. Phase 4: Graph Analysis and Priority Scoring 28
• 3.8. Phase 5: Visualization and Recommendation Generation 32
• 3.9. Statistical Methodology 34
• 3.10. Summary of Design Decisions 36
• Ⅳ. IMPLEMENTATION 38
• 4.1. Technology Stack 38
• 4.2. Crawler Module Implementation 39
• 4.3. Graph Analysis Module Implementation 40
• 4.4. Django REST API 42
• 4.5. D3.js Visualization 43
• 4.6. Report Generation and Export 44
• 4.7. System Integration and Data Flow 44
• 4.8. Error Handling and Edge Cases 45
• V. RESULTS AND ANALYSIS 47
• 5.1. Dataset Characteristics 47
• 5.2. Criterion 1 — Dependency Detection: Results 48
• 5.3. Criterion 2 — Transitive Chain Mapping: Results 49
• 5.4. Criterion 3 — Performance Measurement: Results 50
• 5.5. Criteria 4 and 5 — Visualization and Priority Scoring: Results 53
• 5.6. Complete Results Table 56
• 5.7. Per-Category Breakdown of Results 58
• 5.8. Illustrative Example: Single-Site Analysis 59
• 5.9. Discussion 64
• VI. CONCLUSION 67
• 6.1. Summary 67
• 6.2. Novel Contributions 68
• 6.3. Comparison with Related Tool Categories 69
• 6.4. Directions for Further Study 71
• REFERENCES 74
• APPENDIX 77
• Appendix A1: Playwright Crawler Module 77
• Appendix A2: Graph Analysis Module 78
• Appendix A3: Statistical Validation Code 80
• Appendix A4: D3.js Visualization Core 81
• 국문초록 84