Seamlessly Unifying Attributes and Items: Conversational Recommendation for Cold-start Users

Static recommendation methods like collaborative filtering suffer from the inherent limitation of performing real-time personalization for cold-start users. Online recommendation, e.g., multi-armed bandit approach, addresses this by interactively exploring user preference online and pursuing exploration-exploitation (EE) trade-off. However, existing bandit-based model actions homogeneously. Specifically, they only consider items as arms, being incapable handling item attributes, which naturally provide interpretable information user's current demands can effectively filter out undesired items. In work, we conversational users, where a system both ask attributes recommend to interactively. This important scenario was studied in recent work. it employs hand-crafted function decide when or make recommendations. Such separate modeling makes effectiveness highly rely on choice function, thus introducing fragility system. To address limitation, seamlessly unify same arm space achieve their EE trade-offs automatically using framework Thompson Sampling. Our Conversational Sampling (ConTS) holistically solves all questions choosing with maximal reward play. Extensive experiments three benchmark datasets show that ConTS outperforms state-of-the-art UCB (ConUCB) Estimation-Action-Reflection metrics success rate average number conversation turns.