A personalized active learning strategy with enhanced user satisfaction for recommender systems

Lacking of interaction data, generating effective recommendations for new users has been a challenging task known as the user cold-start problem. Active learning strategies are widely used to address this by collecting user preferences through recommendation-rating rounds. Typically, the active learning phase ends after a fixed number of ratings, which may not be optimal. This rigid setting can result in premature exits or user fatigue. Observing disadvantages of such a rigid setting, this paper designs a dynamic and personalized strategy, in which not only the recommended items, but also the length of the active learning phase. The phase ends based on the system's knowledge of the user evaluated by rating prediction. Extensive experiments on different datasets show that, the proposed Profile Length Confirming Strategy (PLCS) can effectively shorten the active learning phase (profile length) for various algorithms. For instance, on the Coat dataset, our method achieves accuracy comparable to that of the fixed approach, while reducing the number of interaction rounds by at least 3 for 58.47% of new users. Moreover, a new metric, Variance Via Prediction (VP), is also proposed for selecting seed items during the active learning phase, which outperforms the widely-adopted entropy-based metric by 5.6 % in accuracy and 0.9 % in efficiency on the Coat dataset. The proposed PLCS strategies can be stably applied to a variety of active learning algorithms, enhancing new users' experience for recommender systems in different contexts.