heyunh2015 / PsiRecICDM2018

The implementation of PsiRec algorithm from ICDM 2018

PsiRec

Introduction

This is the implementation of PsiRec as described in the paper:

Pseudo-Implicit Feedback for Alleviating Data Sparsity in Top-K Recommendation.
IEEE Conference on Data Mining, 2018.
Yun He, Haochen Chen, Ziwei Zhu and James Caverlee.

The PsiRec algorithm generates pseudo-implicit feedback with indirect user-item relationships to enrich the original extreme datasets. This algorithm focuses on alleviating the data sparsity problem in top-K recommendation for implicit feedback datasets.

Usage

PsiRec learns the user preference from training datasets and then ranks the items for the users in testing datasets.

Input

The input of PsiRec is the set of interactions between users and items from the training dataset:

userId itemId 1.0

1.0 means the user purchased the item.

Output

The output are the evaluation results comparing the ranked items of PsiRec and the groundtruth from the testing dataset. An example is presented as follows:

precision@5: 0.00935503812075
recall@5: 0.0312448221345
precision@10: 0.00683597774572
recall@10: 0.0451534507274

Run

./PsiRec.py --train_file filePath --test_file filePath --walk_length 80 --num_walks 10

Parameters

• -train_file, the training dataset file;
• -test_file, the testing dataset file (or validation dataset file);
• -walk_length, the length of a random walk; the default is 80;
• -num_walks, the number of random walks visiting each user and item; the default is 10;
• -window_size, the context size for sampling the indirect user-item pairs; the default is 3;
• -user_number, the number of users in the dataset;
• -item_number, the number of items in the dataset;
• -train_epoch, the iterations of ALS matrix factorization; the default is 25;
• -lambda_value, the regularization value for ALS matrix factorization; the default is 0.25.
• -latent_factors, the number of latent factors of ALS matrix factorization; the default is 100;
• -validation, the Boolean variable to decide if do the validation on the validation dataset; the default is 0.

Files in folder

PsiRec

• -PsiRec.py, the main function of PsiRec;
• -randomWalks.py, the file to generate random walks from the user-item bigraph;
• -SPPMI.py, the file to calculate Shifted Positive Pointwise Mutual Information (SPPMI) value as the confidence for each pseudo-implicit feedback;
• -alsMF.py, the file to apply latent factors model to predict the item lists for each user based on the dataset enriched by PsiRec;
• -evaludation.py, the file to do the evaluation;

Data

• -preProcessData.py, the file to preprocess the raw datasets, including transferring explicit datasets to implicit datasets and split the dataset into three parts: training, testing and validation;
• -DataInPaper, the datasets exactly the same in our paper, which can be used directly by PsiRec.py;
• -rawData, the raw datasets are too large to be handled here, but you can download according the URLs presented in our paper;
• -preProcessedData, the preprocessed datasets, which can be used directly by PsiRec.py;

Citation

Pending.

Acknowledgement

The technique of randomWalks.py is learned from https://github.com/aditya-grover/node2vec. Th first author is Aditya Grover from Standford University. Thanks to them!