5 Concluding Remarks and Challenges

In this paper we have discussed an approach for discovering some useful knowledge from large amounts of data that are manually and automatically generated during maintenance and operation of commercial aircraft. We have discussed several important issues in analyzing data in this domain. These issues were related to: data format, data complexity, domain information, and presence of contexts. We introduced a knowledge discovery approach that we have developed for this real world application. This approach consists of four steps: (i) identification of the relevant sources of information, (ii) exploration of the selected relevant data, (iii) sampling and data transformation , and (iv) modelling. All of these steps are guided by a specific investigation problem which has to be formulated with the help of domain experts before starting the analysis. The proposed approach helps to guide the analysis through the application of diverse discovery techniques. Such a methodological procedure will help us to address the complexity of the domain considered and therefore optimized our chance to discover valuable information. We presented preliminary results that plausibly confirmed this hypothesis but more experiments are clearly required. During the presentation of the approach, we raised several difficulties that have to be addressed to successfully apply machine learning algorithms in complex real world domains. Among others, we noted the problems related to: the labelling of the instances, the selection of the relevant data, and the use of contextual information. In a long term project, such as the one described here, it may also be very important to address the following three issues: (i) finding an automatic approach to define relevant investigation problems for this domain, (ii) developing tools that are necessary to disseminate discovered knowledge to the end users, and (iii) automating most of the tasks involved in data mining process. Acknowledgments Thanks to Chris Sowerby and Francis Ruest from Air Can-ada to provide domain information and helped us during the evaluation of the results. Thanks to three anonymous referees of the Workshop for their comments on an earlier version of this paper. labelling, and ii) the labelling procedure may be very difficult to automate since it typically involves considerable amount of expertise in the target domain. On the other hand, the complexity of the labelling process may depend on the sampling strategy. More research work on sampling strategies for large data sets [Musick, Catlett, and Russel 1993] is therefore also important. As shown in …