Cameron McLurcan – 1094458 MSc Stream 1 Machine Learning – Assessment 4 The paper chosen for task 2 of assessment 4 is D. Margineantu & T. Dietterich, (1997), Pruning Adaptive Boosting. The problem addressed is that ensemble learning approaches such as AdaBoosting require large memory storage capabilities – capabilities that may not be feasible for many situations. The authors therefore propose a technique called ‘pruning’ which only selects a subset of the hypotheses. This allows a similar level of performance while using less memory. The principle of pruning is based off of the “accuracy/diversity tradeoff,” where two classifiers of high accuracy would reliably provide similar data. Outliers on the other hand would very likely be inaccurate in comparison. The authors devised five pruning algorithms for testing: Early Stopping, KL-divergence Pruning, Kappa Pruning, Kappa-Error Convex Hull Pruning, and Reduce-Error Pruning with Backfitting. Early Stopping is simply halting production of classifiers after M constructions of classifiers from the AdaBoost algorithm. KL-divergence Pruning is an algorithm that chooses a subgroup of diverse classifiers to discard with the Kullback-Leubler Divergence algorithm to prune. Once measured, the selection algorithm used is: Kappa Pruning measures differences between classifiers with a variation of the Kappa algorithm, which measures the level of ‘agreement’ between them. The variation provides a much lower probability of the classifiers agreeing by chance due to skewed class numbers. Once calculated, the predetermined classifiers are selected for use with the most ‘agreeable’ pairs of classifiers given priority. Kappa-Error Convex Hull Pruning is similar to Kappa pruning, but also takes the error rate of the classifiers into consideration. A scatter plot is drawn of each pair of classifiers, with x being Kappa values, and Y being the error rate. The subgroup of classifiers to be used are then taken from the tip of the cluster with the highest Kappa (which generally also has the lowest error rate). The final pruning procedure is Reduce-Error Pruning with Backfitting, which requires the splitting of the training set into a pruning set and a sub-training set. The sub-training set is used to train the AdaBoost classifier, whereas the pruning set is used to select which of the classifiers to keep and which to discard. A method called backfitting is used in conjunction with the pruning set in order to select the classifiers with the lowest pruning set error rates. It was found that the best pruning methods were Kappa Pruning and Reduce-Error Pruning, where both retained a high level of performance (above 0.8 in relation to the unpruned control group) until pruning percentages of 60% and above. Even radical pruning procedures between 60%-80% proved fairly reliable. It was also interesting to note that with these two methods, pruning the 20% most different classifiers resulted in a performance rate above that of the unpruned control group, having eliminated the outliers in the classifier sets. Asides from the fact that pruning 20% of the classifiers for those methods provided a more reliable result than AdaBoosting on its own, the fundamental weakness of pruning is the reduction of accuracy when the subgroup of classifiers to be pruned is large enough to make a significant difference to memory costs. The paper was written in a time when memory was much scarcer than it is today, and so the value of memory in comparison with accuracy of results has shifted accordingly. Pruning is therefore less necessary in most cases than when the paper was released.