Applications

ILP is often applied to data mining tasks. The goal is not always just concept learning, as presented above. It is also possible to perform classification, distance based learning, clustering, descriptive learning, kernel based learning, reinforcement learning, and so on. Here are two example applications.

Toxicological databases list molecules and their effects to living organisms. It is possible to use ILP to predict this activity based on the structure of the molecule. The structure can be represented as relational tables containing atoms and bonds. Itemsets are the frequent substructures in the molecules that should be useful in classification. Figure 5.2 shows an example. Helma et al. (2000) test several different ILP systems on predictive toxicology. Graph based molecular data mining (see overview by Fischer and Meinl, 2004) is an active research topic with lots of unsolved questions.

Figure 5.2: Two molecules share a substructure in the bottom left arms. Substructures are useful in classifying molecules.

Intrusion detection systems are used to monitor computer systems for signs of security violations. Normally the alerts are presented to the human analyst, but Pietraszek and Tanner (2005) present an application of ILP for automatic classification of alerts to decrease the number of false alerts. The data contain the time of day and week, source and destination ports and IP addresses of the connection, as well as the amount of traffic for each alert. There is also some background knowledge such as the network topology. Other alerts related to the current one are essential in some cases such as password guessing and port scanning. The induced rules were comprehensible and could decrease the number of false alarms considerably.