L3E-HD: A Framework Enabling Efficient Ensemble in High-Dimensional Space for Language Tasks

pytorch implementation of L3E-HD for Language Tasks.

overview

This repository provides a implementation of the framework for language tasks.

• Code

  • hdc.py
    • implementation of HDC framework, using ngram for encoding and hamming distance for similarity calculation.
  • adaboost.py
    • implementation of adaboost framework, using HdC framework as the classifier.
  • main.py
    • provides different datasets and paraameters for user to choose.

• dataset

  • language
    • Language classification task.
  • SST-2
    • The Stanford Sentiment Treebank from GLUE, for sentiment classification task.
  • ag_news_csv
    • News articles classification task.
  • spam.csv
    • Text spam classification task.
  • Youtube-all
    • Youtube comment spam classification task.

Run model

Example

python main.py --task-id 5 --classifiers 4 --boost-lr 1.0 --dim 2000 --ngram 4 --retrain-rounds 0 --hdc-lr 0.0005

Parameters

• --task-id
  • from 1 to 5, indicating different tasks
• --classifiers
  • the number of classifiers for the boost framework
• --boost-lr
  • learning rate of boost framework
• --dim
  • dimension of HDC framework
• --ngram
  • value of n for ngram encoding method in HDC framework
• --retrain-rounds
  • iterations of retraining in HDC framework
• --hdc-lr
  • learning rate of HDC framework

Citation

We now have a paper, titled "L3E-HD: A Framework Enabling Efficient Ensemble in High-Dimensional Space for Language Tasks", which is published in SIGIR-2022.

@inproceedings{liu2021L3EHD,
 title={L3E-HD: A Framework Enabling Efficient Ensemble in High-Dimensional Space for Language Tasks},
 author={Liu, Fangxin and Li, Haomin and Jiang, Li},
 booktitle={Proceedings of the International ACM Sigir Conference on Research and Development in Information Retrieval (SIGIR)},
 year={2022}
}