A valid and important question is: how do existing classification systems (such as neural networks, decision tree classifiers, rule classifiers, etc) fare in such a domain? Why do we need new tools if existing systems work fine?
Unfortunately, feeding in the raw data to conventional learning algorithms is likely to perform poorly under most circumstances. In some domains it is possible to use extensive domain knowledge to manually create techniques for extracting attributes from the raw data and then feeding these into a classification algorithm. This sometimes works well; but it does not generalise and rarely produces meaningful descriptions of what has been learnt. Also, they may fail to take advantage of heuristics that can be applied.
We first investigate why existing techniques and heuristics for conventional classification are likely to fail in the case of these temporal domains.