Answers are grounded in this paper and cite their sources.
Learning individual preferences through proactive interaction is constrained by limited user attention and feedback. We present Proactive Dual-Value Task Mining, which estimates task preferences with a dynamic concept graph and selects candidate tasks by combining this estimate with frequency-based novelty and graph centrality. The InteractMine dataset contains 1000 sampled tasks with independent ratings from four student volunteers and four LLM-simulated profiles. A scenario-disjoint offline evaluation compares the method with simple predictors, active selection strategies, and ablations under matched feedback budgets. At 90 observations, its human-rating MAE is 0.2150 and its deployment reward 0.7308, on par with kNN (0.2143 and 0.7302); with ten observations its human-rating MAE is 0.2207, against 0.2309 for kNN, lower in all four human columns. No difference from the baselines is statistically significant, and the centrality term rarely changes the selected task. With the same estimator, the dual-value rule performs like ε-greedy selection, so these results reflect the estimator rather than the selection rule. The concept-level representation thus matches strong simple baselines, is more accurate on human ratings when feedback is scarce, and is designed to keep preferences interpretable and editable. A wearable prototype with a WeChat Mini Program implements the sensing and monitoring steps.
No takes yet. Share an insight, caveat, or question.
Huang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: