Chris Drummond

Cost curves: An improved method for visualizing classifier performance

This paper introduces cost curves, a graphical technique for visualizing the performance (error rate or expected cost) of 2-class classifiers over the full range of possible class distributions and misclassification costs. Cost curves are shown to be superior to ROC curves for visualizing classifier performance for most purposes. This is because they visually support several crucial types of performance assessment that cannot be done easily with ROC curves, such as showing confidence intervals on a classifiers performance, and visualizing the statistical significance of the difference in performance of two classifiers. A software tool supporting all the cost curve analysis described in this paper is available from the authors.

Explicitly representing expected cost: an alternative to ROC representation

ABSTRACT This paper proposes an alternative to ROC representation, in which the expected cost of a classi er is represented explicitly. This expected cost representation maintains many of the advantages of ROC representation, but is easier to understand. It allows the experimenter to immediately see the range of costs and class frequencies where a particular classi er is the best and quantitatively how much better it is than other classi ers. This paper demonstrates there is a point/line duality between the two representations. A point in ROC space representing a classi er becomes a line segment spanning the full range of costs and class frequencies. This duality produces equivalent operations in the two spaces, allowing most techniques used in ROC analysis to be readily reproduced in the cost space.

Accelerating reinforcement learning by composing solutions of automatically identified subtasks

This paper discusses a system that accelerates reinforcement learning by using transfer from related tasks. Without such transfer, even if two tasks are very similar at some abstract level, an extensive re-learning effort is required. The system achieves much of its power by transferring parts of previously learned solutions rather than a single complete solution. The system exploits strong features in the multi-dimensional function produced by reinforcement learning in solving a particular task. These features are stable and easy to recognize early in the learning process. They generate a partitioning of the state space and thus the function. The partition is represented as a graph. This is used to index and compose functions stored in a case base to form a close approximation to the solution of the new task. Experiments demonstrate that function composition often produces more than an order of magnitude increase in learning rate compared to a basic reinforcement learning algorithm.

Severe class imbalance: why better algorithms aren't the answer

This paper argues that severe class imbalance is not just an interesting technical challenge that improved learning algorithms will address, it is much more serious. To be useful, a classifier must appreciably outperform a trivial solution, such as choosing the majority class. Any application that is inherently noisy limits the error rate, and cost, that is achievable. When data are normally distributed, even a Bayes optimal classifier has a vanishingly small reduction in the majority classifiers error rate, and cost, as imbalance increases. For fat tailed distributions, and when practical classifiers are used, often no reduction is achieved.

Warning: statistical benchmarking is addictive. Kicking the habit in machine learning

Algorithm performance evaluation is so entrenched in the machine learning community that one could call it an addiction. Like most addictions, it is harmful and very difficult to give up. It is harmful because it has serious limitations. Yet, we have great faith in practicing it in a ritualistic manner: we follow a fixed set of rules telling us the measure, the data sets and the statistical test to use. When we read a paper, even as reviewers, we are not sufficiently critical of results that follow these rules. Here, we will debate what are the limitations and how to best address them. This article may not cure the addiction but hopefully it will be a good first step along that road.

Composing functions to speed up reinforcement learning in a changing world

This paper presents a system that transfers the results of prior learning to speed up reinforcement learning in a changing world. Often, even when the change to the world is relatively small an extensive relearning effort is required. The new system exploits strong features in the multi-dimensional function produced by reinforcement learning. The features generate a partitioning of the state space. The partition is represented as a graph. This is used to index and compose functions stored in a case base to form a close approximation to the solution of the new task. The experimental results investigate one important example of a changing world, a new goal position. In this situation, there is close to a two orders of magnitude increase in learning rate over using a basic reinforcement learning algorithm.

Using a Case Base of Surfaces to Speed-Up Reinforcement Learning

This paper demonstrates the exploitation of certain vision processing techniques to index into a case base of surfaces. The surfaces are the result of reinforcement learning and represent the optimum choice of actions to achieve some goal from anywhere in the state space. This paper shows how strong features that occur in the interaction of the system with its environment can be detected early in the learning process. Such features allow the system to identify when an identical, or very similar, task has been solved previously and to retrieve the relevant surface. This results in an orders of magnitude increase in learning rate.

Robustness of classifiers to changing environments

In this paper, we test some of the most commonly used classifiers to identify which ones are the most robust to changing environments The environment may change over time due to some contextual or definitional changes The environment may change with location It would be surprising if the performance of common classifiers did not degrade with these changes The question, we address here, is whether or not some types of classifier are inherently more immune than others to these effects In this study, we simulate the changing of environment by reducing the influence on the class of the most significant attributes Based on our analysis, K-Nearest Neighbor and Artificial Neural Networks are the most robust learners, ensemble algorithms are somewhat robust, whereas Naive Bayes, Logistic Regression and particularly Decision Trees are the most affected.

Discriminative vs. generative classifiers for cost sensitive learning

This paper experimentally compares the performance of discriminative and generative classifiers for cost sensitive learning. There is some evidence that learning a discriminative classifier is more effective for a traditional classification task. This paper explores the advantages, and disadvantages, of using a generative classifier when the misclassification costs, and class frequencies, are not fixed. The paper details experiments built around commonly used algorithms modified to be cost sensitive. This allows a clear comparison to the same algorithm used to produce a discriminative classifier. The paper compares the performance of these different variants over multiple data sets and for the full range of misclassification costs and class frequencies. It concludes that although some of these variants are better than a single discriminative classifier, the right choice of training set distribution plus careful calibration are needed to make them competitive with multiple discriminative classifiers.

Inner Ensembles: using ensemble methods inside the learning algorithm

Ensemble Methods represent an important research area within machine learning. Here, we argue that the use of such methods can be generalized and applied in many more situations than they have been previously. Instead of using them only to combine the output of an algorithm, we can apply them to the decisions made inside the learning algorithm, itself. We call this approach Inner Ensembles. The main contribution of this work is to demonstrate how broadly this idea can applied. Specifically, we show that the idea can be applied to different classes of learner such as Bayesian networks and K-means clustering.