Multiple Instance Learning with Bag-Level Randomized Trees

Knowledge discovery in databases with a flexible structure poses a great challenge to machine learning community. Multiple Instance Learning (MIL) aims at learning from samples (called bags) represented by multiple feature vectors (called instances) as opposed to single feature vectors characteristic for the traditional data representation. This relaxation turns out to be useful in formulating many machine learning problems including classification of molecules, cancer detection from tissue images or identification of malicious network communications. However, despite the recent progress in this area, the current set of MIL tools still seems to be very application specific and/or burdened with many tuning parameters or processing steps. In this paper, we propose a simple, yet effective tree-based algorithm for solving MIL classification problems. Empirical evaluation against 28 classifiers on 29 publicly available benchmark datasets shows a high level performance of the proposed solution even with its default parameter settings. Data related to this paper are available at: https://github.com/komartom/MIDatasets.jl. Code related to this paper is available at: https://github.com/komartom/BLRT.jl.