Interpretability of results through visualization

Visualization of patterns as well as interaction with the results of mining methods raises a number of problems. Fortunately, hierarchical clustering visualizations, as well as editing them, has been well studied during the last twenty years, and LaBRI has participated extensively in that research.
Leveraging their previous work about visualizations of clustered graphs and their results on interactive editing of clusterings, we will design a fully interactive visualization system for the chemo-informatician.
The goal is to set up a “sensemaking loop” to collect feedback from chemo-informaticians and to dynamically adjust our mining process accordingly. The inner representational sensemaking loop is composed of four steps: (1) search a representation allowing to encode salient features, (2) encoding present data in the representation, (3) identify data that do not fit, (4) adjust the representation to account for that data. In our work, we will extend that loop, since user feedback will modify the information to be represented (output of mining algorithms). In past work, LABRI researchers have applied this conception method to different application domains, e.g. bioinformatics, and investigated the evaluation of visualization techniques by analysing end-user performance.
The novelty of the proposed research is to work on methods that ease the design of specific visualizations by offering automatic testing and evaluation of their effectiveness according to end
user needs, a research topic for which LABRI researchers have recently obtained interesting results. InvolvD offers a perfect test-bed to evaluate and adapt our new approach to design and develop interactive visualization software/techniques. There are three main
phases that we will follow during the project:

  1. Exploring the design space. Thanks to our previous work and results obtained during this project, we will define a set of goal tasks to be performed by chemo-informaticians. As a start, rejecting a pattern, reassigning instances, indicating must-link/cannot-link pairs, and reordering patterns according to their importance are must-have functionalities. In addition,
    several new operations will be explored to fit the needs of chemo-informaticians, specifically methods for manual editing of clusters. Designing an interface for those operations is not
    obvious, and we will explore the set of possible interactive visual metaphors (building blocks for our interface design) for each of them and then evaluate which one fits our needs best.
    Node-link diagrams or containment diagrams may be straightforward choices. Going further, we aim to automatically explore and evaluate the design space to provide robust criteria and quantitative metrics that help to select a solution and argue why it should be used.
  2. Automatic evaluation of static visualizations. Evaluating visualizations is difficult. Designing an evaluation is complex and time-consuming, and the results may lack statistical
    significance. Thus, only few possible visualizations are experimentally evaluated even if the design space is large. The rise of modern efficient computer vision techniques, such as deep convolutional neural networks, may assist visualization researchers to automatically check some hypotheses, and thus select good visualization candidates. In this project, we propose to leverage our novel evaluation methodology and to extend it in line with the specific needs of the project. Based on our recent advances in this new area, we will generalize task-checking automation, as well as integrate the latest deep neural network results. The goal tasks to be addressed in this project are good candidates for that purpose since they
    require addressing open problems in the area, such as the automatic detection/checking of connectivity in the visualization, or the automatic detection of embeddings.
  3. Automatic evaluation of interactions. A big challenge in human-computer interaction is the successful analysis of complex tasks that require several interaction steps, such as zooming into a part of the visualization to detect a specific outlier, or selecting a subset of elements in one visualization to highlight them in another (brushing and linking). Such analysis
    workflows are a lot more complex than static visualizations but are what users need. After defining possible analysis scenarios with the chemo-informaticians, we will focus on evaluations of the effectiveness of the interactive visualization, which will automatically run through those scenarios. Recent progress on recurrent deep neural networks will be the cornerstone of that work. To our knowledge, studying the evaluation of interactive systems with AI is completely new, and success of this research will therefore globally benefit the entire information visualization community.