The 2nd International Workshop on Inductive Reasoning and Machine Learning for the Semantic Web Proceedings

I will describe a novel meta-learning approach to optimizingthe knowledge discovery or data mining (DM) process. This approach hasthree features that distinguish it from its predecessors. First, previousmeta-learning research has focused exclusively on improving the learningphase of the DM process. More specifically, the goal of meta-learning hastypically been to select the most appropriate algorithm and/or parame-ter settings for a given learning task. We adopt a more process-orientedapproach whereby meta-learning is applied to design choices at differentstages of the complete data mining process or workflow (hence the termmeta-mining). Second, meta-learning for algorithm or model selectionhas consisted mainly in mapping dataset properties to the observed per-formance of algorithms viewed as black boxes. While several generationsof researchers have worked intensively on characterizing datasets, littlehas been done to understand the internal mechanisms of the algorithmsused. At best, a few have considered perceptible features of algorithmslike their ease of implementation or their robustness to noise, or the in-terpretability of the models they produce. In contrast, our meta-learningapproach complements dataset descriptions with an in-depth analysisand characterization of algorithms their underlying assumptions, opti-mization goals and strategies, together with the structure and complexityof the models and patterns they generate. Third, previous meta-learningapproaches have been strictly (meta) data-driven. To make sense of theintricate relationships between tasks, data and algorithms at differentstages of the data mining process, our meta-miner relies on extensivebackground knowledge concerning knowledge discovery itself. For thisreason we have developed a data mining ontology, which defines the es-sential concepts and relations needed to represent and analyse data min-ing objects and processes. In addition, a DM knowledge base gathers as-sertions concerning data preprocessing and machine learning algorithmsas well as their implementations in several open-source software pack-ages. The DM ontology and knowledge base are domain-independent;they can be exploited in any application area to build databases de-scribing domain-specific data analysis tasks, datasets and experiments.Aside from their direct utility in their respective target domains, suchdatabases are the indispensable source of training and evaluation datafor the meta-miner. These three features together lay the groundwork forsemantic meta-mining, the process of mining DM meta-data on the basisof data mining expertise distilled in an ontology and knowledge base.