Interactive Visual Data Exploration with Subjective Feedback

Kai Puolamäki, Bo Kang, Jefrey Lijffijt, Tijl De Bie

TL;DR

This paper introduces a novel interactive system for exploring high-dimensional data that allows users to guide the visualization process by expressing their learned knowledge as intuitive constraints. By updating a "background distribution" based on user feedback, the system can then present projections of the data that maximally differ from this learned knowledge, revealing previously unknown patterns.

What Problem Does This Paper Address?

Exploratory Data Analysis (EDA) is crucial for understanding complex datasets, but current tools often rely on predefined criteria for generating visualizations. This can lead to projections that highlight obvious features or miss subtle, user-specific insights. There's a need for EDA systems that can effectively incorporate a user's evolving understanding of the data, moving beyond static visualizations to an interactive dialogue where the system learns from the user's observations. This paper addresses this gap by proposing a method that formalizes user knowledge as constraints and uses them to drive the search for informative data projections.

Methods

The core of the proposed method is a theoretical model that uses a Maximum Entropy (MaxEnt) distribution to represent the user's current understanding of the data (the "background distribution"). When a user provides feedback, such as identifying a cluster, this is translated into intuitive constraints (e.g., "these points form a cluster"). These constraints are then used to update the MaxEnt background distribution. The system then finds projections (specifically, 2D views) where the actual data and this learned background distribution differ the most. This difference is quantified by comparing the data to the whitened background distribution, often using Principal Component Analysis (PCA) or Independent Component Analysis (ICA) to find directions of maximal variance or non-Gaussianity. The process is iterative: the user is shown an informative projection, provides feedback, the background distribution is updated, and a new projection is generated. The paper also details efficient numerical methods for updating the background distribution and a whitening transformation to highlight deviations from the expected distribution.

Datasets

The paper evaluates its approach using:

  • Synthetic datasets: Generated with varying numbers of data points, dimensions, and clusters to test the runtime and effectiveness of the method.
  • British National Corpus (BNC) data: A large, annotated text corpus used to explore high-level structure and identify genre-related patterns.
  • UCI Image Segmentation data: A dataset with 2310 samples used to demonstrate the system's ability to reveal structure and outliers.

Key Insights and Ideas

  • Subjective Interestingness: The system moves beyond objective measures of "interestingness" (like high variance) to incorporate the user's subjective understanding and evolving knowledge.
  • Intuitive Constraints: User knowledge is captured through simple, visual interactions (e.g., selecting points to form a cluster) rather than requiring mathematical expertise.
  • Iterative Refinement: The interactive loop of visualization, feedback, and re-projection allows for a dynamic exploration process, uncovering deeper insights over time.
  • MaxEnt as a Knowledge Model: Using a Maximum Entropy distribution provides a principled way to model the user's beliefs and update them based on new information.
  • Whitening for Deviation Detection: The whitening operation is key to making deviations between the data and the background distribution visually apparent, especially when using standard projection methods like PCA/ICA.

Key Results

  • The developed system, SIDER, is implemented and demonstrated to be effective in revealing hidden structures in both synthetic and real-world datasets.
  • The runtime experiments show that the optimization process for updating the background distribution scales well with the number of data points and is primarily dependent on dimensionality and the number of constraints, with a default time cutoff of 10 seconds for interactive use.
  • Case studies on the BNC and UCI Image Segmentation datasets successfully identified meaningful clusters and potential outliers, demonstrating the system's utility for domain experts.
  • The method allows for the use of standard projection techniques (PCA, ICA) in a novel, interactive way, adapting them to user-defined criteria.

Limitations and Future Work

  • Convergence Speed: While generally efficient, the paper notes that convergence can be slow in certain "adversarial" cases, particularly with small datasets or overlapping constraints. The authors suggest potential workarounds like data replication or simply cutting off iterations.
  • Generalization to Other Data Types: The current framework is primarily for real-valued data. The authors suggest that generalizing to categorical or ordinal data, or higher-order statistics, is a potential future direction.
  • Modeling Complex Interactions: While the system handles basic constraints well, exploring more complex forms of user knowledge and interaction could be beneficial.
  • Automated Constraint Generation: Future work could explore ways to automatically suggest potential constraints or interesting patterns to the user.

Summary updated

Abstract

Data visualization and iterative/interactive data mining are growing rapidly in attention, both in research as well as in industry. However, integrated methods and tools that combine advanced visualization and data mining techniques are rare, and those that exist are often specialized to a single problem or domain. In this paper, we introduce a novel generic method for interactive visual exploration of high-dimensional data. In contrast to most visualization tools, it is not based on the traditional dogma of manually zooming and rotating data. Instead, the tool initially presents the user with an ‘interesting’ projection of the data and then employs data randomization with constraints to allow users to flexibly and intuitively express their interests or beliefs using visual interactions that correspond to exactly defined constraints. These constraints expressed by the user are then taken into account by a projection-finding algorithm to compute a new ‘interesting’ projection, a process that can be iterated until the user runs out of time or finds that constraints explain everything she needs to find from the data. We present the tool by means of two case studies, one controlled study on synthetic data and another on real census data. The data and software related to this paper are available at http://www.interesting-patterns.net/forsied/interactive-visual-data-exploration-with-subjective-feedback/.
BibTeX
@inproceedings{puolamaki2016interactive,
  author = {Kai Puolamäki and Bo Kang and Jefrey Lijffijt and Tijl De Bie},
  title = {{Interactive Visual Data Exploration with Subjective Feedback}},
  booktitle = {ECML/PKDD},
  year = {2016},
  doi = {10.1007/978-3-319-46227-1\_14},
}
Figure