The ABACOC Algorithm: A Novel Approach for Nonparametric Classification of Data Streams

Abstract

Abstract-Stream mining poses unique challenges to machine learning: predictive models are required to be scalable, incrementally trainable, must remain bounded in size (even when the data stream is arbitrarily long), and be nonparametric in order to achieve high accuracy even in complex and dynamic environments. Moreover, the learning system must be parameterless -traditional tuning methods are problematic in streaming settings-and avoid requiring prior knowledge of the number of distinct class labels occurring in the stream. In this paper, we introduce a new algorithmic approach for nonparametric learning in data streams. Our approach addresses all above mentioned challenges by learning a model that covers the input space using simple local classifiers. The distribution of these classifiers dynamically adapts to the local (unknown) complexity of the classification problem, thus achieving a good balance between model complexity and predictive accuracy. We design four variants of our approach of increasing adaptivity. By means of an extensive empirical evaluation against standard nonparametric baselines, we show state-of-the-art results in terms of accuracy versus model size. For the variant that imposes a strict bound on the model size, we show better performance against all other methods measured at the same model size value. Our empirical analysis is complemented by a theoretical performance guarantee which does not rely on any stochastic assumption on the source generating the stream.

Personalized Food Image Classifier Considering Time-Dependent and Item-Dependent Food Distribution

“…We modify the frequency-based part of this method to classify novel classes, but it did not perform well. It is also difficult to train ABACOC [27], another incremental learning method, with few samples per class. NCM [22], NCT [23] and 1-NN [28] classifier can learn the user's data incrementally and achieve better performance that CNN, but the speed of personalization of NCM and NCT is slow and 1-NN has a cold-start problem.…”

Active Incremental Recognition of Human Activities in a Streaming Context

“…The number of balls is dependent on the complexity of the classification problem. Unlike (De Rosa et al, 2014), where the balls were always centered on input samples, we extend the AUTO-ADJ version of ABACOC described in (De Rosa et al, 2015). In particular, a K-means-like update step makes the centre of each ball shift towards the average of the training samples that were correctly predicted by the local classifier; in other words, the balls track the feature clusters.…”

“…In order to curb the system's memory footprint, we adopt the simple approach proposed in (De Rosa et al, 2015), which is based on deleting existing balls whenever a given budget parameter on the label query rate is attained. This is crucial for real-time applications, as NN search, used in both training and prediction, takes time logarithmic in the number of balls.…”

“…To the best of our knowledge, our approach is the first one to address all the challenges listed above in a principled manner. Our starting point is a recently proposed local algorithm for classification of data streams (De Rosa et al, 2015) which is incrementally trainable and nonparametric (i.e., the model structure is not specified a priori, but determined by the data in such a way that the number of parameters is not fixed in advance), while exhibiting theoretical guarantees on its performance. Here we leverage on this result, and extend it to the active learning setting.…”

“…Furthermore, we show that by combining FIVER with the robust temporal segmentation algorithm presented in (Fanello et al, 2013), we obtain a system able to deal, in a straightforward manner, with a realistic continuous active recognition scenario. A significant contribution of this work is the extension of (De Rosa et al, 2015) to an active learning setting. This is key to the practical application of incremental learning in streaming settings for at least two reasons.…”