Towards Guidelines for Designing Human-in-the-Loop Machine Training Interfaces
summary
Pilot study builds a recommender training UI with four interface variants (like/dislike, feature toggles, comparative nearest‑neighbour, and labeled-reason revision) and asks users to train the system to recommend a red chair. Measures include NASA TLX, self-reported engagement, number of inputs and time, and downstream learning efficiency. Authors fit a linear interaction quality model that yields a single Q score, find richer expression improves experience, recall is costly, and more time does not guarantee better labels.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 0:16Challenges
- 0:44Unique Problems
- 1:05Pilot Study
- 1:44Experiment
- 2:06Interfaces
- 2:47Qualitative Feedback
- 3:36Comparative Interaction Quality
related talks
12mHeterogeneity in Customization of Recommender Systems By Users with Homogenous Preferences
18mLearning How To Cook...
20mTowards human-guided machine learning
15mCreating General User Models from Computer Use
09mEvalAssist: Insights on Task-Specific Evaluations and AI-Assisted Judgment Strategy Preferences
44mMachine Learning + Human
28mDesigning for Success: UX Principles for Internal Developer Platforms
12mHands-On Tutorial for Defining and Scoping Recommendation Systems