Orion: Enabling Suggestions in a Visual Query Builder for Ultra-Heterogeneous Graphs

TL;DR

Orion is a visual query interface for ultra-heterogeneous graphs that uses machine learning to assist users in constructing complex queries, significantly improving success rates and efficiency.

Contribution

The paper introduces Orion, a novel visual query builder with a machine learning-based suggestion system, including the RDP algorithm, for ultra-heterogeneous graphs.

Findings

70% success rate in query construction with Orion

RDP outperforms other ML algorithms, requiring fewer suggestions

Significant improvement over baseline systems in user studies

Abstract

The database community has long recognized the importance of graphical query interface to the usability of data management systems. Yet, relatively less has been done. We present Orion, a visual interface for querying ultra-heterogeneous graphs. It iteratively assists users in query graph construction by making suggestions via machine learning methods. In its active mode, Orion automatically suggests top-k edges to be added to a query graph. In its passive mode, the user adds a new edge manually, and Orion suggests a ranked list of labels for the edge. Orion's edge ranking algorithm, Random Decision Paths (RDP), makes use of a query log to rank candidate edges by how likely they will match the user's query intent. Extensive user studies using Freebase demonstrated that Orion users have a 70% success rate in constructing complex query graphs, a significant improvement over the 58% success rate by the users of a baseline system that resembles existing visual query builders. Furthermore, using active mode only, the RDP algorithm was compared with several methods adapting other machine learning algorithms such as random forests and naive Bayes classifier, as well as class association rules and recommendation systems based on singular value decomposition. On average, RDP required 40 suggestions to correctly reach a target query graph (using only its active mode of suggestion) while other methods required 1.5--4 times as many suggestions.