A Study on the Relative Importance of Convolutional Neural Networks in Visually-Aware Recommender Systems

This is the official implementation of our paper A Study on the Relative Importance of Convolutional Neural Networks in Visually-Aware Recommender Systems published in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021.

Authors: Yashar Deldjoo, Tommaso Di Noia, Daniele Malitesta*, Felice Antonio Merra.
*corresponding author

(a) AlexNet

(b) ResNet50

Table of Contents:

• Requirements
• Run and evaluate recommendations
• Datasets
• Parameters for Image Feature Extractors
• Visual Recommenders
• Configuration Files
• The Authors

Requirements

To begin with, please make sure your system has these installed:

• Python 3.6.8
• CUDA 10.1
• cuDNN 7.6.4

Then, install all required Python dependencies with the command:

pip install -r requirements.txt

Finally, you are supposed to structure the dataset folders in the following way:

./data
  amazon_baby_vgg19/
    original/
       images/
        0.jpg
        1.jpg
        ...
  amazon_boys_girls_alexnet/
    original/
      images/
        0.jpg
        1.jpg
        ...

N.B. The dataset folder structure requires the notation <dataset_name>_<cnn_name>, even though the different dataset folders contain the exact same files. This is due to the fact that, when training and evaluating state-of-the-art visual-based recommender systems on these datasets through Elliot, they need to be recognized as different datasets.

Run and evaluate recommendations

To reproduce the results discussed in the paper, please follow these three steps:

1. Extract visual features from item images. You can refer to this GitHub repository.
2. Train and evaluate the visual-based recommenders through this version of Elliot (TO BE MERGED INTO THE MAIN BRANCH SOON).
3. Evaluate the visual diversity (VisDiv). Again, you can refer to this GitHub repository.

Datasets

Dataset	k-cores	# Users	# Products	# Feedbacks
Amazon Baby*	5	606	1,761	3,882
Amazon Boys & Girls*	5	600	2,760	3,910

* https://jmcauley.ucsd.edu/data/amazon/

Parameters for Image Feature Extractors

Fully-connected layers

CNN	Output Layer (script)	Output Shape
AlexNet	5	(1, 4096)
VGG19	fc2	(1, 4096)
ResNet50	avg_pool	(1, 2048)

Convolutional layers (e.g., ACF)

CNN	Output Layer (script)	Output Shape
AlexNet	Not necessary	(36, 256)
VGG19	block5_pool	(49, 512)
ResNet50	avg_pool	(49, 2048)

Visual Recommenders

Model	Paper
Visual Bayesian Personalized Ranking (VBPR)	He and McAuley
Deep Style	Liu et al.
Attentive Collaborative Filtering (ACF)	Chen et al.
Visual Neural Personalized Ranking (VNPR)	Niu et al.

Configuration Files

• Amazon Baby
  • AlexNet
  • VGG19
  • ResNet50
• Amazon Boys & Girls
  • AlexNet
  • VGG19
  • ResNet50

The Authors

• Yashar Deldjoo ([email protected])
• Tommaso Di Noia ([email protected])
• Daniele Malitesta* ([email protected])
• Felice Antonio Merra ([email protected])

*corresponding author