A Functional Programming Approach to Distance-based Machine Learning

Distance-based algorithms for both clustering and prediction are popular within the machine learning community. These algorithms typically deal with attributevalue (single-table) data. The distance functions used are typically hard-coded. We are concerned here with generic distance-based learning algorithms that work on arbitrary types of structured data. In our approach, distance functions are not hard-coded, but are rather first-class citizens that can be stored, retrieved and manipulated. In particular, we can assemble, on-the-fly, distance functions for complex structured data types from pre-existing components. To implement the proposed approach, we use the strongly typed functional language Haskell. Haskell allows us to explicitly manipulate distance functions. We have produced a SW library/application with structured data types and distance functions and used it to evaluate the potential of Haskell as a basis for future work in the field of distancebased machine learning. 1. General Framework for Data Mining A general framework for data mining should elegantly handle different types of data, different data mining tasks, and different types of patterns/models. Dzeroski (2007) proposes such a framework, which explicitly considers different types of structured data and socalled generic learning algorithms that work on arbitrary types of structured data. The basic components of different types of such algorithms (such as distance or kernel-based ones) are discussed. Taking the inductive database (Imielinski and Mannila 1996) philosophy that proposes that patterns/models are first-class citizens that can be stored and manipulated, Dzeroski proposes to store and manipulate basic components of data mining algorithms, such as distance functions.