PREFAB: PREFerence-based Affective Modeling for Low-Budget Self-Annotation

Paper Title

PREFAB: PREFerence-based Affective Modeling for Low-Budget Self-Annotation

Publication Info

• Topic area: Affective computing and self-annotation methods.
• Keywords: Affective computing, self-annotation, preference learning, peak-end rule, ordinal modeling, cognitive workload, temporal efficiency, user study, interpolation, affective inflection regions.

Background and Problem

• Problem / challenge: Traditional self-annotation methods require full annotation of affective states, which is cognitively demanding, time-intensive, and prone to fatigue and errors. Existing alternatives often rely on intrusive sensors and do not address non-annotated segments.
• Significance: Reducing the burden of self-annotation is critical for scalable affective data collection, enabling applications in gaming, education, and other domains.
• Motivation and related work: Prior work has explored tools for capturing affective states and selective annotation methods using physiological signals. However, these approaches are limited by reliance on intrusive sensors and cardinal modeling. This paper builds on the peak-end rule and ordinal representations to address these gaps.

Solution

• Proposed approach: PREFAB (PREFerence-based Affective Modeling for Low-Budget Self-Annotation), a retrospective self-annotation method targeting affective inflection regions.
• Novelty:
  1. Preference learning-based selective annotation focusing on peak-end inflections.
  2. Integration of ordinal emotion modeling and the peak-end rule.
  3. Introduction of a preview mechanism to enhance annotation accuracy.
  4. Demonstration of reduced cognitive workload and preserved annotation quality.
• Procedure and key techniques:
  • Step 1: Predict affective changes using a trained model with biographical and gameplay data.
  • Step 2: Detect inflection points using peak-end dynamics and event segmentation theory.
  • Step 3: Annotate only selected inflection regions.
  • Step 4: Interpolate unannotated segments using linear interpolation.

Results

• Concrete findings:
  • PREFAB achieved the highest F1 score (up to 0.685) and lowest Δ TE (0.124) compared to baselines across nine games.
  • Reconstruction consistency with full-annotation traces showed moderate agreement (CCC = 0.67), strong ordinal correspondence (Spearman’s ρ = 0.69), and high morphological alignment (DTW similarity = 0.82).
• Advantage over baselines:
  • Outperformed naïve, heuristic, and cardinal modeling methods in identifying affective inflection regions.
  • Reduced cognitive workload and improved confidence compared to full annotation.
• Experiments / evaluation:
  • Technical evaluation using the AGAIN dataset and PAGAN tool.
  • User study with 25 participants comparing full annotation, PREFAB without preview, and PREFAB with preview.
  • Metrics included F1 score, Δ TE, self-reported workload, confidence, annotation quality, and temporal efficiency.
• Limitations and future work:
  • Limited generalizability due to the focus on a single game and participant demographics.
  • Upfront cost of collecting training data for PREFAB.
  • Restricted interpolation policy to linear methods; future work could explore personalized strategies.

Summary

PREFAB introduces a preference learning-based method for retrospective self-annotation, targeting affective inflection regions to reduce cognitive and temporal burdens. It leverages the peak-end rule and ordinal modeling to focus on key emotional moments, supported by a preview mechanism for enhanced accuracy. Results demonstrate significant improvements in annotation efficiency and quality compared to baselines, with reduced workload and preserved affective trajectories. While temporal efficiency gains are conditional, PREFAB offers a scalable solution for affective computing applications, with potential for further refinement in adaptive preview strategies and interpolation techniques.