DataTerminatorX/MRecLibs

About

This project aims at implementing different models for recommender system, regardless of running speed and performance. Still under construction.

Usage

See examples in test_reclib.py

Models

Tips:

1. Memory based models and matrix factorization models evaluate results on trainY while classification models evaluate on testY. Because the latter contain information of trainY when do modeling, the former don't.

2. Usually, sequence models split dataset into train/test/predict by samples while others by time.

• Memory based Models
  • User/Item CF models (with different similarity mearsurements)
  • Content based models
• Matrix Factorization models
  • Vanilla SVD
  • SVD++
• Classfication Models
  • Naive Bayes
  • Logistic Regression
  • GBDT
• Sequence Models
  • N-gram
  • LSTM
  • Time-aware LSTM
• Ensemble Models
  • Bagging
  • AdaBoost
  • Stacking

Memory based Models

Split dataset by time. Have train and predict mode.

--------data set--------

|-- trainX --|-- trainY --|

|--- predictX ----|

• CF (Collaborative Filtering)

  1. Train: Caculate similarities on trainX, tune parameter K (K is number of neighbours), do recommendation, evaluate on trainY, find best K
  2. Predict: Re-caculate similarities on predictX, fix best K, do recommendation

• CB (Content Based)

  1. Train: Tune parameter D (D is dimension of features), extract features for users and items from trainX, calculate similarities, do recommendation, evaluate on trainY
  2. Predict: Re-extract features for users and items from predictX, calculate similarities, do recommendation

Matrix Factorization Models

Split dataset by time. Have train and predict mode

--------data set--------

|-- trainX --|-- trainY --|

|--- predictX ----|

• Vanilla SVD

• SVD++

  1. Train: Tune parameter D (D is dimension of latent features), learn latent features for users and items on trainX, evaluate on trainY, find best D
  2. Predict: Fix best D, learn latent features for users and items on predictX, do recommendation

Classfication Models

Split dataset by time. Have train, test and predict mode.

--------data set----------------

|-- trainX --|-- trainY --|

|-- testX --|-- testY --|

|-- predictX-|

• NB (Naive Bayes)

  1. Train: Extract features from trainX, caculate possibilities p(y|x) on trainX and trainY
  2. Test: Do classification and evaluate on testX and testY
  3. Predict: Re-caculate possibilities on predictX, do classification

• LR (Logistic Regression)

  1. Train: Extract features from trainX, tune parameters e (e is parameter of Sigmoid function) and C (C is coefficient of regularization term), learn model weights on trainY
  2. Test: Do classification and evaluate on testX and testY
  3. Predict: Re-extract features on predictX, fix best e and C, do classification

• XgBoost

Sequence Models

Split dataset by samples. Have train, test and predict mode.

--------data set----------------

|-- train(seqX,seqY) --|

|-- test(seqX,seqY) --|

|-- predict(seqX)-|

• N-gram

• LSTM

• Time-aware LSTM