The effect of explanations and algorithmic accuracy on visual recommender systems of artistic images
summary
Study tests a visual recommender built on UGallery item data that ranks by cosine similarity using either AlexNet fc6 4096-d vector (DNN) or 7-dim attractiveness features (brightness, sharpness, colorfulness, etc.). Three interfaces compare no explanation, example-based explanations (similar images), and transparent feature histograms. With N=121 users the DNN gives higher accuracy and benefits most from explanations (higher perceived relevance, explainability, satisfaction and trust), AVF yields greater perceived diversity, and overly detailed transparency can reduce effectiveness.
AI-written summary · about
chapters (16)
- 0:00Start
- 1:18Related Work: Art Recommendation
- 4:33Gaining Inspiration from DARPA XAI
- 6:55Research Questions
- 7:44Materials and Methodology
- 8:45Visual Features: DNN vs. AVE
- 9:43Visual Features: AVF
- 10:14UGallery: Making Recommendations
- 11:03Generating Explanations
- 11:38White-box Explanation (AVF)
- 12:10Black-box explanation (DNN and AVF)
- 12:30Study Procedure
- 13:16explainable, no transparency
- 13:33explainable & transparent
- 17:18Conclusion
- 18:05Future Work
related talks
33m"Are You Sure?": blending product comparisons and recommendations with A.I.
09mVisual, textual or hybrid: the effect of user expertise on different explanations
45mAI-404: Explanation Not Found
09mAre Explanations Helpful? A Comparative Study of the Effects of Explanations in AI-Assisted ...
09mQuantitative Evaluation of Machine Learning Explanations: A Human-Grounded Benchmark
49mCreation as Curation [Keynote]
16mMaking recommendations explainable
37mInteractive Explainable AI