Characteristics of Deep and Skim Reading on Smartphones vs. Desktop: A Comparative Study

Document Title

Characteristics of Deep and Skim Reading on Smartphones vs. Desktop: A Comparative Study

Document Information

• Subject Area: Human-Computer Interaction (HCI), Reading Behavior Analysis
• Keywords: Deep Reading, Skim Reading, Digital Devices, Eye Tracking, Fixation Characteristics, Reading Pattern Classification

Research Background and Issues

• Identified Problems or Challenges:

  • With the proliferation of digital devices, especially mobile devices, reading habits are shifting from deep reading to skim reading.
  • How can deep reading and skim reading patterns be more accurately measured and distinguished? Currently, there is a lack of methods for large-scale implicit measurement of reading patterns on both desktop and mobile devices.

• Importance of the Problem:

  • Deep reading promotes text comprehension, memory retention, and critical thinking, which are essential for developing complex cognitive skills. However, skim reading may undermine these abilities.
  • As younger generations increasingly rely on mobile devices for reading, deep reading skills may gradually be neglected, potentially impacting education, information comprehension, and decision-making abilities.

• Research Motivation and Related Work:

  • Current assessments of reading quality primarily rely on post-reading comprehension tests, which are not easily scalable to real-world reading scenarios.
  • Eye-tracking data is considered closely related to cognitive processes, but its application in distinguishing deep reading from skim reading remains underexplored, particularly on mobile devices.
  • This study is inspired by two prior works (Biedert et al., Kelton et al.), but unlike these studies, it focuses on longer texts and, for the first time, incorporates mobile device usage scenarios.

Proposed Solution

• Proposed Solution:

  • Developed an implicit method to detect deep reading and skim reading using eye-tracking data.
  • Systematically induced deep reading and skim reading behaviors on desktop and smartphone devices, combined with machine learning models to differentiate between the two reading patterns.

• Innovations:

  • Conducted the first comparison of eye-tracking characteristics of deep reading and skim reading on desktop and mobile platforms.
  • Explored model adaptability and feature importance for detecting reading patterns across the two platforms.
  • The method is scalable and implicitly measures reading patterns without imposing additional burdens on users or authors.
  • Selected study texts that closely resemble real-world reading environments, including long texts (>1500 words).

• Implementation Steps:

  1. Experimental Design:
     • Inducing reading patterns: Participants were assigned tasks for deep reading (answering in-depth questions) and skim reading (quickly locating answers).
     • Recorded eye-tracking data and interaction behaviors, calculating features such as fixation duration and saccade length.
     • Tasks were completed in a quiet environment using both desktop and smartphone devices.
  2. Data Preprocessing:
     • Addressed scrolling behavior by realigning fixation data to ensure a vertically continuous reading experience.
     • Extracted low-level and mid-level fixation features within varying time windows (1-120 seconds sliding window).
  3. Machine Learning Training and Evaluation:
     • Trained classifiers such as SVM, Random Forest, and XGBoost.
     • Optimized classification model performance using leave-one-out cross-validation.
  4. Cross-Platform Analysis:
     • Analyzed differences between desktop and mobile devices and evaluated model transferability.

Research Findings

• Specific Findings:

  • Successfully induced deep reading and skim reading behaviors:
    • In deep reading, participants read more slowly, focused on text comprehension, and achieved higher accuracy in answering questions.
    • In skim reading, participants read faster but primarily retained explicit information.
  • Developed machine learning models capable of distinguishing deep reading from skim reading:
    • On desktop devices, the XGBoost classifier achieved an AUC of 0.82 and an accuracy of 88.36%.
    • On mobile devices, logistic regression and XGBoost also performed well, with AUC values reaching up to 0.73.
  • Deep reading characteristics included more short saccades, longer fixation durations, and regression behaviors; skim reading characteristics relied on longer saccades and more concentrated targeting of specific areas.

• Comparison with Existing Solutions:

  • Conducted the first study on reading patterns on mobile devices, addressing the limitations of prior research that focused solely on desktop platforms.
  • The experimental method was closer to real-world reading environments (long texts, multi-page navigation), and classifier performance slightly surpassed existing methods.

• Experimental or Evaluation Results:

  • Deep reading activities were significantly associated with eye-tracking features; the importance of model features varied with time windows, requiring a window of 60-120 seconds to capture effective features.
  • Compared to desktops, feature extraction on smartphones was more challenging, necessitating optimized feature sets specifically for mobile platforms.

• Limitations and Future Directions:

  • Limitations:
    • The experiment was conducted in a laboratory setting, ignoring distractions and environmental factors in natural scenarios.
    • The sample size was small (29 participants), limiting the generalizability of the model.
    • Cross-device model transfer failed due to screen size differences, leading to a lack of universal features.
  • Future Directions:
    • Improve mid- and high-level feature construction to reduce device dependency.
    • Validate the model in natural settings and explore the use of RGB cameras as a lower-cost tracking sensor.
    • Extend research to examine how interface design influences deep reading and propose design principles to promote deeper text engagement.