Decision rule elicitation for domain adaptation
summary
Elicits simple heuristic decision rules from experts and composes them with a trained model via a weighted sum of expert rules (weights theta, similarity-based smoothing sigma) then mixes that with the original predictor using an input-dependent weight alpha. Evaluated on 40 modeled experts and a real user study, and on sentiment domain transfer from IMDB to 25 Amazon topics, logistic regression plus rule feedback often exceeds label feedback and the base IMDB model, improving both training- and test-domain accuracy.
AI-written summary · about
chapters (6)
related talks
34mSynergizing model and human decision-making using augmented machine learning
12mCausal Inference And Data-Driven Decision-Making
40mDecision-Making Under Miscalibration
38mPlaying by the Rules-Based-Systems
23mLearning with Limited Labels & Weak Supervision
37mFOLD-SE: An Efficient Rule-based Machine Learning Algorithm with Scalable Explai...
39mPseudo-Labeling for Covariate Shift Adaptation
52mLanguage-guided Adaptation