SuperMIL: Supervised Multiple Instance Learning for Time Series Classification

Chentao Liu, Changchun He, Baohan Mi, Xin Huo

Keywords: Multiple Instance Learning, Time Series Classification

TL;DR: This paper presents SuperMIL, a novel MIL framework boosting TSC by capturing local causal dependencies between instances.

Abstract:

Decision-making requirements in fields like industrial monitoring and healthcare are equally critical for ensuring the accuracy and interpretability of time series classification (TSC) methods.
Multiple instance learning (MIL) with interpretability is a promising framework that decomposes the sequences (bag) into instance-level segments and evaluates their contributions via attention mechanisms.
However, the existing MIL methods introduced in the TSC field have two major limitations:1) degraded instance-level feature learning due to optimizing for bag-level predictions and 2) insufficient modeling of causal dependencies among temporally ordered instances.
This paper proposes supervised multi-instance learning (SuperMIL), which iteratively optimizes instance-level pseudo-supervision (from bag-inherited labels) and bag-level weak supervision, yielding discriminative instance features and robust bag-level predictions.
Moreover, SuperMIL integrates a Hawkes pooling module and a coupled multi-instance loss.
The former captures local inter-instance causality by decomposing excitations into directional similarity and instance differences, and the latter models inter-instance as well as collective loss-global attention interactions to align instance-level and bag-level objectives, both of them synergizing capture local causal causality and global instance semantics.
The SuperMIL framework enhances performance in representative TSC models, outperforming traditional MIL methods, as validated through experiments on the UCR and UEA datasets.
Code is available at this repository: \url{https://anonymous.4open.science/r/SuperMIL}.

Primary Area: learning on time series and dynamical systems

Submission Number: 3741