Home

Awesome

Meta-Open-World-Learning

code for our TheWebConf (WWW 2019) paper titled "Open-world Learning and Application to Product Classification".

Checkout this article (or this ) for related background.

*** Feedbacks are welcomed to improve this repository ***

Problem to Solve

Classic supervised learning assumes that the classes seen in testing must have appeared in training. However, this assumption is often violated in real-world applications when new topics emerge constantly or in e-commerce new categories of products appear daily. A model working in such an open environment must be able to

(1) reject examples from unseen classes (not appeared during training) and

(2) incrementally learn the new/unseen classes to expand the existing model.

We call this problem open-world learning (OWL).

Environment

This project is tested on Python 2.7 + Keras 2.2.2 with Tensorflow 1.4.0 on Ubuntu 16.04, but it should generally work for other versions.

Data

Download the preprocessed data from here. Save it to amazon/data .

cd amazon/data/
tar -zxf data.tar.gz

This leads to the meta training classes in train1_npz and 3 sets (25, 50, 75 incrementally) of meta testing classes (including both training/testing data for meta testing classes).

We only release the preprocessed data at this stage (although some preprocessing code is available under amazon folder).

Running Script

cd script

Single run with k=5 n=9 (ncls=9+1=10, with 1 positive class and 9 negative classes).

bash run.sh 5 10

Run all settings of k and n with 10 runs (note this will cause a lot of time so it's better to distribute tasks to different GPUs)

bash run_batch.sh

The results are saved in eval.json under each run's folder under different parameter setting.

Citation:

@inproceedings{xu2019open,
  title={Open-world Learning and Application to Product Classification},
  author={Xu, Hu and Liu, Bing, Shu, Lei and Yu, P.},
  booktitle={Proceedings of the 2019 World Wide Web Conference on World Wide Web},
  year={2019},
  organization={International World Wide Web Conferences Steering Committee}
}

TODO: