/KDD-benchmark

ECNU Data Mining Final Project

Primary LanguagePython

KDD_Benchmark

ECNU CS Data Mining, 2021 Spring

✔️ Our work ranked 2 finally. Not bad.

Team Member

@Zhao Yunxiang, @Wei Mingda, @Wu Ronghuan

How to run

python3 trainer.py

you may need to modify the following line in trainer.py based on your python environment.

cmd = "/usr/local/bin/python3 evalution.py %s %s" % (gold_file, pred_file)

Timeline

  • coauthor_1 + coauthor_2 + RandomForestClassifier: 75.08%
  • coauthor_1 + coauthor_2 + stringDistance_1 + stringDistance_2 + RandomForestClassifier: 93.5442%
  • Actually, the demo provided by my teacher performs really well :-)
  • 2021/05/27: After adding some features, the accuracy reached 95.1%.
  • 2021/05/31: After model ensemble, the accuracy reached 95.71%.

Ideas

1. add features

  1. Add feature publication_year. Time-consuming.

  2. Try different ways to calculate text similarity. Jaro Winkler wins.

  3. Add feature keyword of paper titles and keywords. Time-consuming.

  4. Add feature journal_conference_year. Good.

  5. Add feature journal_count. Slow.

  6. Try classifier GradientBoostingRegressor. Result: 0: 44%, 1: 99%, overall: 50%

    Modify parameters. Bad.

    n_estimators=100, learning_rate=1.0, max_depth=1, random_state=0

  7. Add feature affiliation_count. Slow. Nothing changed.

2. select models

RandomForestClassifier performs well.

Finally, we merged different models to generate predictions.

3. model ensemble

According to the accuracy of various models, we assigned different models with different weights.