WSDM Cup 2017: Valdalism Detection

Vandalism detection (task 2) - WSDM Cup 2017

Introduction

We are a team of 4 from Complex Network Research Group (Murata Laboratory) - Tokyo Insitute of Technology. This repository is our submission to the 2017 WSDM Cup. In summary, the task is to detect vandalism in Wikidata dumps.

Milestones

Preparation: 2016/09/26 - 2016/09/30

Objective

1. Setup personal computer to match each others. (Python 3.5.2, Tensorflow 0.10, scikit-learn 0.17.1, Anaconda virtual env, coding style, etc.)
2. Literature review. (Paper listing, reading, and discussion)
3. Competition score metric analysis.
4. Finalize and present possible approaches.

Daily log

1. 27th: List of vandalism detection papers; setup working environments; study wikidata dumps; analyze WSDM'17 score metrics.
2. 28th: Paper reading; discussion about Random Forest and features selection; focusing on Random Forest model and its variations.
3. 29th: Run the provided reference paper's code on the lab's machine; study related techniques to RR; study NN techiqnues that complement RR.
4. 30th: Review week 1.

Baseline model: 2016/10/03 - 2016/10/07

Objective

1. Preprocess wikimedia data, study previous features extration code.
2. Implement simple random forest model based on [1].
3. Working baseline model and sketch of neural network model.

Daily log

1. 3rd: Features from the baseline model [1] are all hand-picked.
2. 4th: Meeting cancelled.
3. 5th: Meeting cancelled.
4. 6th: Some features are missing compared to the original implementation [1]. Using only 29 available features now yields 0.02 on ROC. This result is extremely low.
5. 7th: