LISTA DE ABREVIACIONES
CAPÍTULO 2. REVISIÓN DE LA LITERATURA
4.1 Propuesta y Desarrollo de Modelos DIHEQS y SATSIB
7.2.2 Future Work
The data sets presented in chapter 4 contain isolated drifts with different levels of severity and speed. For studies which do not emphasize predictable and recurrent drifts, these data sets are enough, as sequences of drift can be analysed through real world problems. However, studies of recurrent and predictable drifts would benefit from data sets containing different sequences of different types of recurrent and predictable drifts. So, such data sets should be created as future work.
7.3
A Diversity Study in the Presence of Drift
7.3.1 Contributions
Even though ensembles have been used to handle concept drift, the literature did not contain any deep study of why they can be helpful for that and which of their features can contribute or not to deal with concept drift. Such a study is important because a better understanding of the behaviour of ensembles in the presence of concept drift allows better exploitation of their features for dealing with drifts. Chapter 5 presents such a study, answering research question 1.2.2: “when, how and why ensembles can be helpful for dealing with drifts?”
The study shows that, even though low diversity ensembles are likely to be more accurate on the old concept, high diversity ensembles can reduce the drop in accuracy on the new concept right after the beginning of the drift. The analysis also indicates that the higher the severity of the drift the more important the use of higher diversity if no additional strategy is adopted to converge to the new concept. This is particularly helpful when the system does not know that a drift is happening and cannot adopt any strategy to help convergence to the new concept.
Nevertheless, additional strategies are important to encourage convergence to the new con- cept. Simply training on the new concept an old ensemble trained with low diversity on the old concept is helpful during part of the drifting time in gradual drifts, as some of the instances to be predicted still reflect the old concept. The strategy of using an ensemble trained with high diversity on the old concept, but with low diversity on the new concept, is helpful to improve the accuracy when the drifts have low severity or speed. This behaviour is intuitively reasonable, as the old concept can be helpful when the changes are slow or not large. When the drift is fast and with high severity, the best strategy is to simply create a new low diversity ensemble from scratch to learn the new concept.
This study also allows us to answer research question 1.2.3: “can we use information from the old concept to better deal with the new concept? How?” Intuitively, if a concept drift causes few changes to the old concept, information learnt from the old concept should be helpful to aid the learning of the new concept. However, to the best of our knowledge, no approach in the
7.3 A Diversity Study in the Presence of Drift
literature attempts to use information from the old concept in order to aid the learning of the new concept. Chapter 5 shows that using an ensemble trained with high diversity on the old concept, but with low diversity on the new concept, allows the use of information from the old concept to aid the learning of the new concept.
An additional contribution given by chapter 5 is the modified online bagging algorithm. The study indicates that the simple technique of using a pre-defined parameter for the P oisson distribution of online bagging allows us to consistently and directly encourage more or less diversity in the ensemble.
7.3.2 Future Work
Further study of diversity in the presence of recurrent and predictable drifts could be done as future work.
It is also interesting to check whether the difficulty of the concept learnt before a drift influences the learning of the new concept. As explained in section 5.3, the study raised sus- picions that a concept easily learnt may be more difficult to be forgotten, possibly increasing the error on the new concept. Another hypothesis to be checked is if higher diversity can help learning when there are many irrelevant attributes, as the only database in which a λ lower than 1 (λ = 0.1) provided the best test error before the drift was Plane, which contains many irrelevant attributes.
A different direction of research related to this study is to determine whether it is possible to estimate the underlying distributions of a certain concept so as to identify it when it reappears. In this way, information learnt from a certain concept could be reused if and when the concept reappears. The use of naive Bayes as base learners may be helpful for that, as it is a direct way to estimate the posterior probability.
Further directions also include the study of other features which may help to deal with drifts, such as ensemble size, base learners’ individual accuracy and feature selection. Considering feature selection, as an initial study, the effect of ensembles with different sets of features could be analysed. That could have a relationship with diversity, which should be verified.
Another direction would be to investigate the use of mutation to deal with concept drifts. Very high diversity does not allow good convergence to the current concept, allowing the ensem- ble to converge to a new concept at the same time as information from the old concept can be used to aid the learning. Instead of using an ensemble with high diversity, we could mutate the learners. An example of mutation for MLPs would be to add small random values to its weights. The disadvantage would be that different types of mutation would be required for different types of learners.