An Iterative Model For Relevance Search Over Knowledge Graphs

Relevance search is an established research topic in data mining research.The task is to search a knowledge graph for answer entities that are most relevant to a query entity.As it may be demanding for a user to formally express her desired semantics of relevance,existing methods infer meta-path based semantics from user-provided example answer entities.However,a user in practice may provide very few examples at a time,even none at the beginning of interaction(i.e.,a cold start),thereby limiting the use and effectiveness of existing methods.To address these limitations,this thesis proposes to study iterative relevance search.In each iteration,a user is allowed to easily label the current answer entities rather than to manually input,and then reward the user with improved answer entities.To enable this new iterative setting,this thesis proposes a generic approach TRAVERS.Firstly,TRAVERS proposes a novel diversity-oriented ranker based on a weighted maximum coverage model which can effectively solve the cold start problem.Secondly,this method learns relevance through an enhanced relevance-oriented ranker based on a pairwise learning to rank model which reduces the impact of unbalanced samples labeled by users.In addition,this method solves the exploration-exploitation dilemma by dynamically integrating two different rankers.Finally,this thesis conducts extensive experiments considering the situation of different noises.TRAVERS outperformed a variety of adapted strong baselines in extensive experiments with simulated and real user behavior.