CONCLUSIONS

{% if main.adsdop %}{% include 'adsenceinline.tpl' %}{% endif %}

CONCLUSIONS

In this chapter, we proposed a new framework for feature selection in supervised and unsupervised learning. In particular, we note that each feature subset should be evaluated in terms of multiple objectives. In supervised learning, ELSA with neural networks model (ELSA/ANN) was used to search for possible combinations of features and to score customers based on the probability of buying new insurance product respectively. The ELSA/ANN model showed promising results in two different experiments, when market managers have clear decision scenario or when they don't. ELSA was also used for unsupervised feature selection. Our algorithm, ELSA/EM, outperforms a greedy algorithm in terms of classification accuracy. Most importantly, in the proposed framework we can reliably select an appropriate clustering model, including significant features and the number of clusters.

We also proposed a new ensemble construction algorithm, Meta-Evolutionary Ensembles (MEE), where feature selection is used as the diversity mechanism among classifiers in the ensemble. In MEE, classifiers are rewarded for predicting difficult points, relative to the other members of their respective ensembles. Our experimental results indicate that this method shows consistently improved performance compared to a single classifier and the traditional ensembles.

One major direction of future research on the feature selection with ELSA is to find a way to boost the weak selection pressure of ELSA while keeping its local selection mechanism. For problems requiring effective selection pressure, local selection may be too weak because the only selection pressure that ELSA can apply comes from the sharing of resources. Dynamically adjusting the local environmental structure based on the certain ranges of the observed fitness values over a fixed number of generations could be a promising solution. In this way, we could avoid the case in which the solution with the worst performance can survive into the next generation because there are no other solutions in its local environment.

Another major direction of future research is related with the scalability issue. By minimizing the communication among agents , our local selection mechanism makes ELSA efficient and scalable. However, our models suffer the inherent weakness of the wrapper model, the computational complexity. Further by combining EAs with ANN to take the advantages of both algorithms, it is possible that the combined model can be so slow that it cannot provide solutions in a timely manner. With the rapid growth of records and variables in database, this failure can be critical. Combining ELSA with faster learning algorithms such as decision tree algorithms and Support Vector Machine (SVM) will be worth pursuing.

ENDNOTES

1. Continuous objective functions are discretized.

2. This is one of main tasks in the 2000 CoIL challenge (Kim & Street, 2000). For more information about CoIL challenges and the data sets, please refer to http://www.dcs.napier.ac.uk/coil/challenge/.

3. If other objective values are equal, we prefer to choose a solution with small variance.

4. This is reasonable because as we select more prospects, the expected accuracy gain will go down. If the marginal revenue from an additional prospect is much greater than the marginal cost, however, we could sacrifice the expected accuracy gain. Information on mailing cost and customer value was not available in this study.

5. The other four features selected by the ELSA/logit model are: contribution to bicycle policy, fire policy, number of trailer, and lorry policies.

6. The cases of zero or one cluster are meaningless, therefore we count the number of clusters as K = κ + 2 where κ is the number of ones and K _min = 2 ≤ K ≤ K _max .

7. For K = 2, we use F _complexity = 0.76, which is the closest value to 0.69 represented in the front.

8. In our experiments, standard error is computed as standard deviation / iter ^0.5 where iter = 5.