Content Based Neural Collaborative Filtering

Our project is derived from NCF: https://github.com/hexiangnan/neural_collaborative_filtering

Environment Settings

Keras with Theano backend.

• Keras version: '1.0.7'
• Theano version: '0.8.0'

Build and Run ContentNCF with Docker

Build docker image

docker build --no-cache=true -t content-ncf .

Run docker image:

docker run --volume=$(pwd):/home content-ncf python ContentNCF.py --dataset content_pin --epochs 5 --batch_size 256 --num_factors 64 --layers [64,32,16,8] --reg_mf 0 --reg_layers [0,0,0,0] --num_neg 4 --lr 0.001 --learner adam --verbose 1 --out 1

Dataset

content_pin dataset - The dataset contains a mapping of boards, which represents users, to their pins, which represents images. This dataset contains 500 users and 24498 user-image interactions.

content_pin.train.rating:

• Train file.
• Each Line is a training instance: userID\t itemID\t rating\t imageURL

test.rating:

• Test file (positive instances).
• Each line is a testing instance: userID\t itemID\t rating\t imageURL

test.negative

• Test file (negative instances).
• Each line corresponds to a line in test.rating, containing additional 99 negative samples.
• Each line is in the format: (userID,itemID)\t negativeItemID1\t negativeItemID2 ...

features_n.pkl

• Image feature respresentations with dimensions n*1
• To change the number of feature dimensions, change filename in ContenNCF.py and set num_factors=64 in docker run command.