Abstract
Many data stream clustering algorithms operate in two well-defined steps: (i) online statistical data collection stage; and (ii) offline macro-clustering stage. The well-known k-means algorithm is often employed for performing the offline macro-clustering step. The conventional k-means algorithm assumes that the number of clusters (k) is defined a priori by the user. Given the difficulty of defining the value of k a priori in real-world problems, we describe a new approach that allows estimating k dynamically from streams with variable number of clusters, which is a common scenario in data with a non-stationary distribution. In addition, we combine our dynamic approach with two different strategies for initializing the centroids during the offline clustering. Analysis of results suggest that, using the dynamic approach, the method k-means++ for centroids initialization present better results. © 2012 Authors.

Abstract
Decision rules are one of the most interpretable and flexible models for data mining prediction tasks. Till now, few works presented online, any-time and one-pass algorithms for learning decision rules in the stream mining scenario. A quite recent algorithm, the Very Fast Decision Rules (VFDR), learns set of rules, where each rule discriminates one class from all the other. In this work we extend the VFDR algorithm by decomposing a multi-class problem into a set of two-class problems and inducing a set of discriminative rules for each binary problem. The proposed algorithm maintains all properties required when learning from stationary data streams: online and any-time classifiers, processing each example once. Moreover, it is able to learn ordered and unordered rule sets. The new approach is evaluated on various real and artificial datasets. The new algorithm improves the performance of the previous version and is competitive with the state-of-the-art decision tree learning method for data streams. © 2012 ACM.

Abstract
The tutorial presents the state-of-the-art in mobile and ubiquitous data stream mining and discusses open research problems, issues, and challenges in this area. © 2011 Authors.

Abstract
As in batch learning, one may identify a class of streaming real-world problems which require the modeling of several targets simultaneously. Due to the dependencies among the targets, simultaneous modeling can be more successful and informative than creating independent models for each target. As a result one may obtain a smaller model able to simultaneously explain the relations between the input attributes and the targets. This problem has not been addressed previously in the streaming setting. We propose an algorithm for inducing multi-target model trees with low computational complexity, based on the principles of predictive clustering trees and probability bounds for supporting splitting decisions. Linear models are computed for each target separately, by incremental training of perceptrons in the leaves of the tree. Experiments are performed on synthetic and real-world datasets. The multi-target regression tree algorithm produces equally accurate and smaller models for simultaneous prediction of all the target attributes, as compared to a set of independent regression trees built separately for each target attribute. When the regression surface is smooth, the linear models computed in the leaves significantly improve the accuracy for all of the targets. © 2011 ACM.

Abstract

Abstract
Learning from data streams is a research area of increasing importance. Nowadays, several stream learning algorithms have been developed. Most of them learn decision models that continuously evolve over time, run in resource-aware environments, detect and react to changes in the environment generating data. One important issue, not yet conveniently addressed, is the design of experimental work to evaluate and compare decision models that evolve over time. There are no golden standards for assessing performance in non-stationary environments. This paper proposes a general framework for assessing predictive stream learning algorithms. We defend the use of Predictive Sequential methods for error estimate - the prequential error. The prequential error allows us to monitor the evolution of the performance of models that evolve over time. Nevertheless, it is known to be a pessimistic estimator in comparison to holdout estimates. To obtain more reliable estimators we need some forgetting mechanism. Two viable alternatives are: sliding windows and fading factors. We observe that the prequential error converges to an holdout estimator when estimated over a sliding window or using fading factors. We present illustrative examples of the use of prequential error estimators, using fading factors, for the tasks of: i) assessing performance of a learning algorithm; ii) comparing learning algorithms; iii) hypothesis testing using McNemar test; and iv) change detection using Page-Hinkley test. In these tasks, the prequential error estimated using fading factors provide reliable estimators. In comparison to sliding windows, fading factors are faster and memory-less, a requirement for streaming applications. This paper is a contribution to a discussion in the good-practices on performance assessment when learning dynamic models that evolve over time.