Workshop AIL Active and Incremental Learning

Most machine learning techniques assume, either explicitly or implicitly, that the data-generating process is stationary. This assumption guarantees that the model learnt during the initial training phase remains valid over time and that its performance is in line with our expectations. Unfortunately, this assumption does not truly hold in the real world representing, in many cases, a simplistic approximation of the reality. The talk will describe the Just-In-Time (JIT) approach that is a flexible tool implementing the detection/adaptation paradigm to cope with evolving processes. Solutions following this approach improve the knowledge about the model in stationary conditions by exploiting additional information coming from the field during the operational life. Differently, in nonstationary conditions, as soon as a change in the data-generating process is detected, the learnt model is discarded and a suitable one activated to keep the performance. As a valuable and challenging application of the proposed approach, JIT classifiers for concept drift will be detailed and discussed. Incremental Decision Tree based on order statistics Christophe Salperwyck1 and Vincent Lemaire2 Abstract. New application domains generate data which are not persistent anymore but volatile: network management, web profile modeling... These data arrive quickly, massively and are visible just once. Thus they necessarily have to be learnt according to their arrival orders. For classification problems online decision trees are known to perform well and are widely used on streaming data. In this paper, we propose a new decision tree method based on order statistics. The construction of an online tree usually needs summaries in the leaves. Our solution uses bounded error quantiles summaries. A robust and performing discretization or grouping method uses these summaries to provide, at the same time, a criterion to find the best split and better density estimations. This estimation is then used to build a naı̈ve Bayes classifier in the leaves to improve the prediction in the early learning stage. New application domains generate data which are not persistent anymore but volatile: network management, web profile modeling... These data arrive quickly, massively and are visible just once. Thus they necessarily have to be learnt according to their arrival orders. For classification problems online decision trees are known to perform well and are widely used on streaming data. In this paper, we propose a new decision tree method based on order statistics. The construction of an online tree usually needs summaries in the leaves. Our solution uses bounded error quantiles summaries. A robust and performing discretization or grouping method uses these summaries to provide, at the same time, a criterion to find the best split and better density estimations. This estimation is then used to build a naı̈ve Bayes classifier in the leaves to improve the prediction in the early learning stage.