Collaborative filtering over evolution provenance data for interactive visual data exploration

Highlights

• Proposal of a visual data exploration system with content and collaborative recommendations.

• Proposal of several merge techniques to aggregate users exploration sessions in a multi-user graph.

• Proposal of several optimizations to improve collaborative-filtering recommendation computation.

• Quantitative evaluation of collaborative-filtering recommendations techniques.

• Qualitative evaluation of users experiences when visually exploring data using our system.

Abstract

In interactive visual data exploration, users rely on recommendations on what data to explore next. EVLIN is a system that recommends queries to retrieve these data for the next exploration step, paired with suited visualizations. This paper extends EVLIN by combining its content-based recommendations with recommendations leveraging collaborative filtering to improve the effectiveness of recommendation-based visual data exploration. The recommendations rely on evolution provenance, which tracks users’ interactions during interactive visual data exploration. As more users explore a dataset, the evolution provenance of individual user explorations is incrementally integrated into a multi-user graph, for which we present match and merge algorithms. To compute collaborative-filtering recommendations, we present a search algorithm and optimizations to efficiently search queries similar to a current user’s query in the multi-user graph and give preference to queries that have been previously explored in an exploration step succeeding those similar queries. Our experimental evaluation studies the efficiency and effectiveness of the solutions proposed in this paper and demonstrates that using the full system with both content-based and collaborative-filtering recommendations enabled allows for effective interactive visual data exploration.

Introduction

Interactive visual data exploration systems support users in investigating data by providing facilities to express queries and to visualize their results. In such systems, users interact with visualizations of query results, which triggers further queries to be executed and their results visualized back to the user. Clearly, interactive visual data exploration is an iterative process, where users transition from one data visualization to the next through interactions such as interactive selection or details-on-demand [1].

To assist users by leading the way in transitioning from one interesting data visualization to the next, recent visual data exploration tools have incorporated recommendations to assist users in querying and in visualizing interesting regions of the data. Most tools rely on content-based recommendation techniques that base their recommendations on data available in a user’s exploration session, such as prior queries of the same user [2], [3], [4]. Collaborative-filtering leverages information from previous explorations made by multiple users to identify prior similar exploration queries worth recommending to the current user. A first work that focuses on query recommendations based on collaborative-filtering for data exploration is [5]. However, to the best of our knowledge, no visual data exploration system blends the two aforementioned recommendation techniques.

This paper presents EVLIN＋＋, a system that bridges this gap. Our motivation to combine both recommendation techniques is to improve the quality or effectiveness of the visual data exploration experience, as recommendations then also take into account globally interesting trends instead of solely relying on the often limited local information underlying content-based recommendations. EVLIN＋＋ extends our EVLIN system [6], [7] that relies on content-based recommendations by integrating collaborative recommendations, as introduced in [8]. The collaborative-filtering recommendations rely on evolution provenance, which tracks user actions and selections alongside their exploration. Overall, this paper makes the following contributions:

System architecture. We present the overall architecture of EVLIN＋＋, which, in addition to the components of the original EVLIN system, has components for storing, aggregating, and analyzing evolution provenance for the purpose of collaborative filtering. It further alters the component that quantifies the interestingness of individual recommendations, as it takes into account scores from both the content-based and the collaborative-filtering based recommendations.

Collaborative-filtering. While we have described the general idea of incorporating collaborative-filtering into EVLIN in a short paper [8], we neither discussed detailed algorithms nor optimizations. This paper describes the two steps underlying our collaborative-filtering based recommendations in detail.

In the first step, we merge a description of a user’s exploration session, i.e., its evolution provenance, into a global multi-user interaction graph based on a similarity matching between the local session graph and the global graph. For this, we discuss several techniques to incrementally merge evolution provenance graphs into a global graph. The different approaches trade off merge efficiency and effectiveness.

In the second step, we compute recommendations for next exploration steps by first searching exploration steps similar to a user’s current step in the global graph and then recommending and scoring steps adjacent to those in the global graph. This paper presents our baseline collaborative-filtering query recommendation algorithm and optimizations to improve its runtime.

Implementation and evaluation. We implement the full system and study the performance of both steps of the collaborative-filtering recommendations presented in this paper. We also evaluate the performance of the interactive visual data exploration achieved when combining our content-based and collaborative-filtering-based query recommendations on both synthetic and real data. The results show the efficiency and the effectiveness of our proposed optimized solutions for interactive visual data exploration.

Section 2 introduces the general architecture and processing of EVLIN＋＋. Section 3 describes our evolution provenance model, providing necessary preliminaries and definitions. Merging evolution provenance graphs is discussed in Section 4. Section 5 covers collaborative-filtering recommendation computation. Section 6 discusses how we integrate collaborative-filtering recommendations and the content-based recommendations to obtain a final set of ranked recommendations. Section 7 presents our experimental evaluation. We discuss related work in Section 8 and conclude in Section 9.