David Jensen,
University of Massachusetts at Amherst

Beyond prediction : Directions for probabilistic and relational learning

Research in learning logical and probabilistic models has greatly increased the range of phenomena that machine learning can address. Recent work has extended these boundaries even further by unifying these two powerful learning frameworks. However, new frontiers await. This talk will survey some recent work in learning probabilistic models of relational data, and discuss several applications of these techniques, including fraud detection in the U.S. securities industry. I will argue that current techniques are capable of learning only a subset of the knowledge needed by practitioners in these domains, and that further unification of probabilistic and logical learning offers a unique ability to dramatically expand the utility of machine learning for science, business, and government.
Paolo Frasconi,
Dipartimento di Sistemi e Informatica
Universita degli Studi di Firenze

Learning with Kernels and Logical Representations

Choosing an appropriate kernel function is a crucial decision for many popular statistical learning algorithms and represent the natural entry point for inserting prior knowledge into the learning process. Inductive logic programming, on the other hand, offers powerful and flexible frameworks for describing existing background knowledge and obtaining additional knowledge from the data. It therefore seems natural to explore the synergy between these two important paradigms of machine learning.

In this talk I will discuss a number of recent kernel-based methods that operate in the ILP setting and present examples of successful applications. In kernels on Prolog proof trees, the representation of an example is obtained by recording the execution trace of a logic program expressing background knowledge. The similarity between two examples is then computed as the similarity between the corresponding execution traces. In declarative kernels, logic programming allows us to specify a broad class of convolution kernels, providing a simple interface for the incorporation of relational background knowledge. In kFOIL, features correspond to the truth values of clauses that are dynamically generated by a greedy search algorithm guided by the empirical risk. Unlike the previous two ideas, kFOIL learns the kernel function from the available data and also returns a corresponding set of clauses that are regarded as features. I will finally present a kernel based on rich combinatorial features described by type extension trees and discuss its application to statistical relational learning with interdependent examples.