Leveraging Data-Analysis Session Logs for Efficient, Personalized, Interactive View Recommendation
Xiaozhong Zhang, Xiaoyu Ge, Panos K. Chrysanthis · 2019
View recommendation has been recently adopted to assist data analysts in better understanding the data. In order to recommend useful views, existing view recommendation approaches propose a variety of utility functions, each suitable for a different usage scenario. However, as the "interestingness" of a recommended view is user-dependent, no single utility function can represent users' preferences and intentions in all cases. With the richly available choices for utility functions, identifying the most appropriate ones along with their tunable parameters remains a challenge even for expert users. To help identify the most appropriate utility function, existing works have made attempts in two different directions, 1) providing generic recommendations of utility functions by learning offline from historical logs, and 2) providing personalized recommendations of utility functions by learning from the interactions with each particular user. Both proposed approaches exhibit clear advantages and disadvantages. In this work, to benefit from both approaches, we device a novel hybrid interactive view recommendation solution, namely HolisticViewSeeker (HVS), that effectively combines the offline learning with the online interactive learning to provide personalized view recommendation. Our experimental evaluations conducted on real-world data show that HVS outperforms both state-of-the-art online and offline approaches by a significant margin in multiple respects.