Richard Brendon Kirkby

New ensemble methods for evolving data streams

Advanced analysis of data streams is quickly becoming a key area of data mining research as the number of applications demanding such processing increases. Online mining when such data streams evolve over time, that is when concepts drift or change completely, is becoming one of the core issues. When tackling non-stationary concepts, ensembles of classifiers have several advantages over single classifier methods: they are easy to scale and parallelize, they can adapt to change quickly by pruning under-performing parts of the ensemble, and they therefore usually also generate more accurate concept descriptions. This paper proposes a new experimental data stream framework for studying concept drift, and two new variants of Bagging: ADWIN Bagging and Adaptive-Size Hoeffding Tree (ASHT) Bagging. Using the new experimental framework, an evaluation study on synthetic and real-world datasets comprising up to ten million examples shows that the new ensemble methods perform very well compared to several known methods.

Weka-A Machine Learning Workbench for Data Mining

The Weka workbench is an organized collection of state-of-the-art machine learning algorithms and data preprocessing tools. The basic way of interacting with these methods is by invoking them from the command line. However, convenient interactive graphical user interfaces are provided for data exploration, for setting up large-scale experiments on distributed computing platforms, and for designing configurations for streamed data processing. These interfaces constitute an advanced environment for experimental data mining. The system is written in Java and distributed under the terms of the GNU General Public License.

New options for hoeffding trees

Hoeffding trees are state-of-the-art for processing high-speed data streams. Their ingenuity stems from updating sufficient statistics, only addressing growth when decisions can be made that are guaranteed to be almost identical to those that would be made by conventional batch learning methods. Despite this guarantee, decisions are still subject to limited lookahead and stability issues. In this paper we explore Hoeffding Option Trees, a regular Hoeffding tree containing additional option nodes that allow several tests to be applied, leading to multiple Hoeffding trees as separate paths. We show how to control tree growth in order to generate a mixture of paths, and empirically determine a reasonable number of paths. We then empirically evaluate a spectrum of Hoeffding tree variations: single trees, option trees and bagged trees. Finally, we investigate pruning.We show that on some datasets a pruned option tree can be smaller and more accurate than a single tree.

Optimizing the Induction of Alternating Decision Trees

The alternating decision tree brings comprehensibility to the performance enhancing capabilities of boosting. A single interpretable tree is induced wherein knowledge is distributed across the nodes and multiple paths are traversed to form predictions. The complexity of the algorithm is quadratic in the number of boosting iterations and this makes it unsuitable for larger knowledge discovery in database tasks. In this paper we explore various heuristic methods for reducing this complexity while maintaining the performance characteristics of the original algorithm. In experiments using standard, artificial and knowledge discovery datasets we show that a range of heuristic methods with log linear complexity are capable of achieving similar performance to the original method. Of these methods, the random walk heuristic is seen to outperform all others as the number of boosting iterations increases. The average case complexity of this method is linear.

Multiclass alternating decision trees

The alternating decision tree (ADTree) is a successful classification technique that combines decision trees with the predictive accuracy of boosting into a set of interpretable classification rules. The original formulation of the tree induction algorithm restricted attention to binary classification problems. This paper empirically evaluates several wrapper methods for extending the algorithm to the multiclass case by splitting the problem into several two-class problems. Seeking a more natural solution we then adapt the multiclass LogitBoost and AdaBoost.MH procedures to induce alternating decision trees directly. Experimental results confirm that these procedures are comparable with wrapper methods that are based on the original ADTree formulation in accuracy, while inducing much smaller trees.

Stress-testing hoeffding trees

Hoeffding trees are state-of-the-art in classification for data streams. They perform prediction by choosing the majority class at each leaf. Their predictive accuracy can be increased by adding Naive Bayes models at the leaves of the trees. By stress-testing these two prediction methods using noise and more complex concepts and an order of magnitude more instances than in previous studies, we discover situations where the Naive Bayes method outperforms the standard Hoeffding tree initially but is eventually overtaken. The reason for this crossover is determined and a hybrid adaptive method is proposed that generally outperforms the two original prediction methods for both simple and complex concepts as well as under noise.

Handling numeric attributes in hoeffding trees

For conventional machine learning classification algorithms handling numeric attributes is relatively straightforward. Unsupervised and supervised solutions exist that either segment the data into predefined bins or sort the data and search for the best split points. Unfortunately, none of these solutions carry over particularly well to a data stream environment. Solutions for data streams have been proposed by several authors but as yet none have been compared empirically. In this paper we investigate a range of methods for multi-class tree-based classification where the handling of numeric attributes takes place as the tree is constructed. To this end, we extend an existing approximation approach, based on simple Gaussian approximation. We then compare this method with four approaches from the literature arriving at eight final algorithm configurations for testing. The solutions cover a range of options from perfectly accurate and memory intensive to highly approximate. All methods are tested using the Hoeffding tree classification algorithm. Surprisingly, the experimental comparison shows that the most approximate methods produce the most accurate trees by allowing for faster tree growth.