Recurrent Halting Chain (RHC)
This repository contains the source code for our paper Recurrent Halting Chain for Early Multi-label Classification, published at KDD, 2020. Please contact Tom Hartvigsen ([email protected]) with any questions.
The key idea of our method is to step through the timesteps one by one, and when the time is right for a given class (as determined by a Halting Policy), add it to a growing set of class labels. We prefer both accurate label sets and early predictions for each class. There are many ways to encode this key idea into a classification architecture and in this work we pair an Order-Free Recurrent Classifier Chain (a state-of-the-art multi-label model) with a Reinforcement Learning-based Halting Policy (a state-of-the-art early classifier).
Examples of code use will be coming soon. For now, this model can effectively be used in lieu of an RNN in pytorch:
from model import RHC
RHC = RHC(ninp, nhid, nclasses) # ninp is the number of variables, nhid is the hidden state size of the rnn, nclasses is the number of classes you want RHC to predict
logits, mean_halting_point = RHC(X) # where X is a tensor of time series of shape (timesteps, instances, variables)
If you find this code helpful, feel free to cite our paper:
@inproceedings{hartvigsen2020recurrent,
title={Recurrent Halting Chain for Early Multi-label Classification},
author={Hartvigsen, Thomas and Sen, Cansu and Kong, Xiangnan and Rundensteiner, Elke},
booktitle={Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},
pages={1382--1392},
year={2020}
}
Tom Hartvigsen will be adding train/test scripts soon -- for now, please check out some analogous code for our EARLIEST paper