Ilknur Icke

Multi-objective genetic programming for visual analytics

Visual analytics is a human-machine collaboration to data modeling where extraction of the most informative features plays an important role. Although feature extraction is a multi-objective task, the traditional algorithms either only consider one objective or aggregate the objectives into one scalar criterion to optimize. In this paper, we propose a Pareto-based multi-objective approach to feature extraction for visual analytics applied to data classification problems. We identify classifiability, visual interpretability and semantic interpretability as the three equally important objectives for feature extraction in classification problems and define various measures to quantify these objectives. Our results on a number of benchmark datasets show consistent improvement compared to three standard dimensionality reduction techniques. We also argue that exploration of the multiple Pareto-optimal models provide more insight about the classification problem as opposed to a single optimal solution.

Automated measures for interpretable dimensionality reduction for visual classification: A user study

A number of visual quality measures have been introduced in visual analytics literature in order to automatically select the best views of high dimensional data from a large number of candidate data projections. These methods generally concentrate on the interpretability of the visualization and pay little attention to the interpretability of the projection axes. In this paper, we argue that interpretability of the visualizations and the feature transformation functions are both crucial for visual exploration of high dimensional labeled data. We present a two-part user study to examine these two related but orthogonal aspects of interpretability. We first study how humans judge the quality of 2D scatterplots of various datasets with varying number of classes and provide comparisons with ten automated measures, including a number of visual quality measures and related measures from various machine learning fields. We then investigate how the user perception on interpretability of mathematical expressions relate to various automated measures of complexity that can be used to characterize data projection functions. We conclude with a discussion of how automated measures of visual and semantic interpretability of data projections can be used together for exploratory analysis in classification tasks.This paper studies the interpretability of transformations of labeled higher dimensional data into a 2D representation (scatterplots) for visual classification.1In this context, the term interpretability has two components: the interpretability of the visualization (the image itself) and the interpretability of the visualization axes (the data transformation functions). We define a data transformation function as any linear or non-linear function of the original variables mapping the data into 1D. Even for a small dataset, the space of possible data transformations is beyond the limit of manual exploration, therefore it is important to develop automated techniques that capture both aspects of interpretability so that they can be used to guide the search process without human intervention. The goal of the search process is to find a smaller number of interpretable data transformations for the users to explore. We briefly discuss how we used such automated measures in an evolutionary computing based data dimensionality reduction application for visual analytics. In this paper, we present a two-part user study in which we separately investigated how humans rated the visualizations of labeled data and comprehensibility of mathematical expressions that could be used as data transformation functions. In the first part, we compared human perception with a number of automated measures from the machine learning and visual analytics literature. In the second part, we studied how various structural properties of an expression related to its interpretability.

Dimensionality reduction using symbolic regression

In this paper, we propose a symbolic regression approach for data visualization that is suited for classification tasks. Our algorithm seeks a visually and semantically interpretable lower dimensional representation of the given dataset that would increase classifier accuracy as well. This simultaneous identification of easily interpretable dimensionality reduction and improved classification accuracy relieves the user of the burden of experimenting with the many combinations of classification and dimensionality reduction techniques

Using Simulation to Evaluate Data-Driven Agents

We use simulation to evaluate agents derived from humans interacting in a structured on-line environment. The data set was gathered from student users of an adaptive educational assessment. These data illustrate human behavior patterns within the environment, and we employed these data to train agents to emulate these patterns. The goal is to provide a technique for deriving a set of agents from such data, where individual agents emulate particular characteristics of separable groups of human users and the set of agents collectively represents the whole. The work presented here focuses on finding separable groups of human users according to their behavior patterns, and agents are trained to embody the groups behavior. The burden of creating a meaningful training set is shared across a number of users instead of relying on a single user to produce enough data to train an agent. This methodology also effectively smooths out spurious behavior patterns found in individual humans and single performances, resulting in an agent that is a reliable representative of the groups collective behavior. Our demonstrated approach takes data from hundreds of students, learns appropriate groupings of these students and produces agents which we evaluate in a simulated environment. We present details and results of these processes.