Logo Logo
Help
Contact
Switch Language to German
Ahmadi Fahandar, Mohsen; Hüllermeier, Eyke (April 2021): Analogical Embedding for Analogy-Based Learning to Rank. Advances in Intelligent Data Analysis XIX. IDA 2021, April 26-28, 2021, Porto, Portugal.
Full text not available from 'Open Access LMU'.

Abstract

Learning to rank based on principles of analogical reasoning has recently been proposed as a novel approach to preference learning. The approach essentially builds on a regularity assumption of the following kind: Given objects A B, C, D, if A relates to B as C relates to D, and A is preferred to B, then C is presumably preferred to D. This assumption is formalized in terms of so-called analogical proportions, which operate on a feature representation of the objects. A suitable representation is therefore essential for the success of analogy-based learning to rank. Therefore, we propose a method for analogical embedding, i.e., for embedding the data in a target space such that, in this space, the aforementioned analogy assumption is as valid and strongly pronounced as possible. This is accomplished by means of a neural network with a quadruple Siamese structure, which is trained on a suitably designed set of examples in the form of quadruples of objects. By conducting experiments on several real-world data sets, we provide evidence for the usefulness of analogical embedding and its potential to improve the performance of analogy-based learning to rank.