Continuous Web Personalization using Selector-template Pairs
Vagner Figueredo de Santana, M. Cecilia C. Baranauskas · 2019 · Proceedings of the 16th International Web for All Conference (W4A) · doi:10.1145/3315002.3317559
Summary
This paper presents a continuous web personalization approach that automatically adjusts website user interfaces based on observed usage patterns to reduce accessibility barriers and usability problems. Built on the WELFIT (Web Event Logger and Flow Identification Tool) platform, the system uses a concept of "selector-template pairs" where selectors are regular expressions that match patterns in event streams (sequences of user interactions like clicks, mouse movements, and focus changes on specific UI elements) and templates are JavaScript/jQuery code skeletons that make adjustments to the matched UI elements. The approach works through four stages: (1) a client-side jQuery plugin non-intrusively logs all JavaScript and jQuery events during natural browsing; (2) event streams are clustered using the Sequence Alignment Method (SAM/Edit Distance) and agglomerative hierarchical clustering to group similar user behaviors per URL; (3) registered selector-template pairs are matched against clustered event streams using state machines to identify problematic interaction patterns and the specific UI elements involved; (4) concrete adjustments (JavaScript code with actual element IDs replacing placeholders) are deployed as weekly experiments with random experimental/control group splits within each cluster. Adjustments address ordering, size, and color of UI elements — the three UI attributes considered. The approach represents "tailoring of level 3" (extension) — inserting new code to modify the UI at runtime without end-user action.
Key findings
The system was evaluated in a 30-month "in the wild" study on the Todos Nós website (a portal for inclusive environments at a Brazilian university) with 408 participants. During the first 24 months, interaction data was collected and six problematic patterns were identified: (I) repetitive clicks near small clickable elements, (II) repetitive mouse movements over poorly contrasted input fields, (III) repetitive movements over non-clickable areas near clickable elements, (IV) clicks on non-clickable elements, (V) text selection indicating poor contrast, and (VI) repetitive clicks on the same element due to lack of feedback. Six corresponding selector-template pairs were created — for example, Adjustment I increases matched small elements by 20%, Adjustment II changes input field background color for better contrast, and Adjustment IV changes the cursor style over non-clickable elements to prevent confusion. During the final 6 months, 110 experiments were generated from the adjustments. Of the 19 experiments that included returning visitors (enabling before/after comparison), 18 successfully reduced usage incidents in the experimental group compared to the control group (94.7% success rate). The most successful adjustments were I (enlarging small elements — 10 successful experiments), IV (cursor correction for non-clickable elements — 4 successful), and V (contrast improvement — 2 successful). One adjustment (I) was unsuccessful when applied to an element that caused horizontal scrolling when enlarged.
Relevance
This research presents a compelling approach to what might be called "self-healing" web accessibility — automatically detecting and fixing accessibility barriers based on how users actually struggle with a website, rather than relying solely on static guideline compliance checking. For accessibility practitioners, this work offers several important insights. First, it demonstrates that interaction patterns reveal accessibility barriers that automated WCAG checkers cannot detect: small click targets, poor contrast, confusing non-clickable elements that look clickable, and missing feedback are all issues that manifest in user behavior rather than in source code analysis. Second, the selector-template approach is elegantly reusable — a template for enlarging small elements works on any website without modification, since the selector identifies the specific elements and the template uses placeholders. Third, the continuous deployment model (weekly experiments with A/B testing) mirrors modern software engineering practices and ensures adjustments are validated before permanent deployment. The 30-month longitudinal study with 408 real users provides unusually strong ecological validity for an accessibility intervention. The approach is particularly valuable because it can work alongside traditional accessibility evaluation: WCAG audits catch known guideline violations, while this system catches the unforeseen usability and accessibility issues that emerge in real-world use.
Tags: web personalization · accessibility · usability · interaction logging · data mining · JavaScript · adaptive interface · automated evaluation · user experience · assistive technology