Interactive Visualization Recommendation with Hier-SUCB

TL;DR

This work addresses the need for interactive, personalized visualization recommendations in large search spaces where user preferences are revealed progressively. It introduces Hier-SUCB, a hierarchical contextual combinatorial semi-bandit that incorporates a learnable bias term to capture configuration-attribute interactions and employs a two-level decision process (configuration first, then attributes and bias) to accelerate learning. The authors prove an improved regret bound of $Reg(T)=O(\sqrt{T\ln^3(m^2T\ln(T))})$ and demonstrate, through synthetic and Plot.ly-based simulations with human validation, that Hier-SUCB outperforms offline methods and standard bandit baselines, with ablations confirming the value of hierarchy and the bias term. The approach enables fast, real-time personalization of visualizations without heavy initial data and has practical implications for aiding analysts in rapidly discovering effective visual representations of complex datasets.

Abstract

Visualization recommendation aims to enable rapid visual analysis of massive datasets. In real-world scenarios, it is essential to quickly gather and comprehend user preferences to cover users from diverse backgrounds, including varying skill levels and analytical tasks. Previous approaches to personalized visualization recommendations are non-interactive and rely on initial user data for new users. As a result, these models cannot effectively explore options or adapt to real-time feedback. To address this limitation, we propose an interactive personalized visualization recommendation (PVisRec) system that learns on user feedback from previous interactions. For more interactive and accurate recommendations, we propose Hier-SUCB, a contextual combinatorial semi-bandit in the PVisRec setting. Theoretically, we show an improved overall regret bound with the same rank of time but an improved rank of action space. We further demonstrate the effectiveness of Hier-SUCB through extensive experiments where it is comparable to offline methods and outperforms other bandit algorithms in the setting of visualization recommendation.

Figures (7)

• Figure 1: An example of interactive PVisRec: A software engineer seeks a useful visualization for a system log dataset. In each round, the agent recommends a visualization and receives user feedback. If negative, it gathers preferences on attributes and configuration separately to refine its model (R1). It is possible that the user likes attributes or configurations but not the visualization (R2). By learning users' feedback, the agent will recommend high-quality visualization accepted by the user (R3).
• Figure 2: Figure 3 shows the distribution of the number of attributes in different user datasets. Table 1 shows all possible configurations of the processed dataset. The item pool of configuration is relatively small
• Figure 3: (a) Our user study shows that from round 20 to 50, the percentage of users that like either visualization increases, indicating the simulator can help bandit algorithms learn user preference. (b) In the visualization dataset Plot.ly, even if a user prefers a set of attributes and visual configurations, they may not prefer their combination.
• Figure 4: Comparison of the hit rate using C2UCB, LinUCB, Hier-SUCB in the synthetic data over 100 iterations. Hier-SUCB outperforms other algorithms in 200 rounds.
• Figure 5: Comparison of the averaged reward (HR@1) using C2UCB, LinUCB, Hier-SUCB and Neuro-PVR (offline method). Hier-SUCB outperforms other bandit algorithms and exceeds the HR@1 of Neuro-PVR in round 80 and 160.

Theorems & Definitions (4)

• lemma 1
• lemma 2
• lemma 3
• theorem 1