luzai/Curriculum_Data_Mining

Data Mining, Summer 2017

Collaborative Filter

Dependence

• Python 2.7
  • Tensorflow 1.1
  • Keras 2

Run

• predict_sub_txt.txt is the final results I predicted

• Put train_sub_txt.txt in data\

• Run ensemble.py, generate res.txt in root directory

Project Structure

$ROOT
├── clean.sh
├── data
│   └── train_sub_txt.txt
├── README.md
└── src
    ├── config.py
    ├── data.py
    ├── ensemble.py
    ├── grid_search.py
    ├── logger.py
    ├── model.py
    └── utils.py

Method

• DeepCF re-implemented with reference to [2] and there are following modifications for our problem:
  • Turn it into classification problem rather than regression.
  • EarlyStop to avoid over fitting, rather than eject noise.
  • Others: Model Ensemble.
• Inherited Ideas:
  • Use history ratings as item's raw feature, enable to generalize better to unseen items. with at least 10 ratings. Because the input is raw feature rather than user index.
  • Adapt In-Matrix prediction, split train: test= 9:1, no overlapping between train and test data.
  • Others: Shuffle and Clean Data, Grid Search.

Some Visualizations:

• After Remove Item less than 10.7 ratings, and shuffle User/Item index:
• • 
• Data Distribution is not balanced, so I tried to augment the unbalanced data, but in the future work I may use on-line hard example mining instead.
• • 
• After grid search for latent dim of SVD, I choose 100
• • 
• From grid search result for Deep CF, we can see severe over-fitting phenomenon
• • 
• So I have to apply early-stop, and I find although the accuracy of DeepCF(On Test Set) is just slightly higher than SVD. But after ensemble, final accuracy can higher than both.
• • 
• • 
• The train and test pattern of SVD, not severely suffer from over-fitting
• • 
• • 

TODO list

• On-line Hard Example Mining
• Ensemble SVD and DaulNet
• Whether to Early-stop/ Checkpoint
• Whether to balance training dataset (Aug+Shuffle)
• SVD
  • Grid search
    • latent dim / Dropout / regularize hyper-parameter
  • accuracy on the fly
• DaulNet
  • Grid search
  • Last time result

	name	acc
0	svd_ynobsxylrc	0.438849
1	svd_mtmhmysqcu	0.436098
2	svd_lmtulnapux	0.435675
3	svd_faufhruaaq	0.434405
4	svd_uxvjuhhcnk	0.432234
5	deep_cf_cfrjuudupl	0.443504
6	deep_cf_kjjiarcqqq	0.444774
7	deep_cf_iwmyylneww	0.448159
8	deep_cf_imdhxyutfe	0.448295

Reference

• [1] SVD : https://github.com/mesuvash/TFMF/blob/master/TFMF.ipynb
• [2] DaulNet/DeepCF re-implement from Xiong Y, Lin D, Niu H, et al. Collaborative Deep Embedding via Dual Networks[J]. 2016. (With Some Modifications)