Information Technology Reference
In-Depth Information
(a)
Cyclic
and naıve Bayes model.
(b)
KddCup99
and decision tree model.
Fig. 6.
Average accuracy with
Cyclic
ans
KddCup99
problems
5
Conclusions
Starting from the requirements constrained by the unpredictable nature of streaming
data, this paper proposed an adaptive selective ensemble approach for data streams
classification. The aim of this work is represented by a new adaptive behavior for an
ensemble model selection approach. The new feature enables the system to automat-
ically adapt the active model threshold to the current stream status. The idea is not
providing a fixed value of the threshold set up experimentally, but letting its value auto-
matically adapt to the data flow changes. When data are quite stable, the system can use
a large part of the ensemble. On the contrary, when data changes the threshold, it has
to be reduced to disable the not up-to-date models. The preliminary results show that,
with respect to the use of a fixed threshold, our adaptive algorithm provides a slightly
worse performance than the ones using the best value of the threshold. Unfortunately,
the choice of the best value is not always feasible, and if a wrong selection is made,
the system loses its precision. Our adaptive approach does not require any assumption
about active model values and displays good adaptation to the different scenarios. This
work represents a first step to guarantee a system completely adaptable to the different
streaming factors. As future works, our aims are to test our adaptive model in a real
stream application with real data. Moreover, we are currently studying the introduc-
tion of runtime monitoring tools for automatically adapting our system, e.g varying the
number of frame levels, or the models available for each layer, dynamically considering
memory consumption and time response constraints.
References
1. Aggarwal, C.C., Han, J., Wang, J., Yu, P.: A framework for clustering evolving data streams.
In: Proceedings of the 2003 International Conference on Very Large Data Bases (VLDB
2003), Berlin, Germany, pp. 81-92 (2003)
2. Aggarwal, C.C., Han, J., Wang, J., Yu, P.: A framework for projected clustering of high
dimensional data streams. In: Proceedings of the 2004 International Conference on Very
Large Data Bases (VLDB 2004), Toronto, Canada, pp. 852-863 (2004)
