Skip to content

Latest commit

 

History

History
72 lines (60 loc) · 6.02 KB

Proposal.md

File metadata and controls

72 lines (60 loc) · 6.02 KB

H&M Personalized Fashion Recommender

1. Introduction

Online shopping offers shoppers an extensive selection of products to browse through. But with too many choices, customers might not quickly find what interests them or what they are looking for, and ultimately, they might not make a purchase. To enhance the shopping experience, product recommendations are key. More importantly, helping customers make the right choices also has a positive implications for sustainability, as it reduces returns, and thereby minimizes emissions from transportation. This project is initiated by H&M as a Kaggle challenge to provide product recommendations based on previous purchases. For this project, we are given the purchase history of customers across time, along with supporting customer and product metadata. Our goal is to predict what articles each customer will purchase in the 7-day period immediately after the training data end date.

2. Preliminary Literature Review

A comprehensive review and classification of existing works on deep learning recommender models was performed by the article "Deep Learning Based Recommender System: A Survey and New Perspectives". The author classified deep learning based recommendation systems(RS) to two catagories: RS with neural building blocks and RS with deep hybrid models. The catagories are composed of different models like MLP, AE, RNN, CNN, GAN and RNN+CNN, AE+CNN respectively.
image
The autor also discussed challenges of current RS researches and proposed future research directions. The propsed future research direction includes joint representation learning, more explainable models for both user and researchers, RS models with more layers, machine reasoning, etc.
To the end, the author urged RS researching community to establish a standard dataset and hidden/blinded test set for better model evaluation. According to the study, the baseline and dataset of RS is not as standarized as the CV/NLP community, which will create an inconsistent reporting of scores, with each autor using his/her own dataset and evaluation method. Thus, there's seemingly no consensus on the general ranking of the models.

3. Objectives and expected contributions

We hereby posit two research questions, formalized in the following, and by leveraging a large H&M transactions and product dataset. We provide empirical evidence as well as theoretical grounding to answer these questions:

  • Research Question 1 (RQ1): What are the SOTA RS models that can be efficiently implemented to this problem?

  • Research Question 2 (RQ2): How do those models perform against each other in the Kaggle test set?

We expect this work can make the following contributions:

  • We provide deep learning approaches to identify customers shopping habits and give accurate prediction to the future purchases.
  • We quantitatively evaluate the SOTA deep learning RS models with the hidden/blinded test set provided by Kaggle. This work can provide evidence for which model works best in real world business situations.

4. Methodology

In this project, we will use the dataset from H&M Kaggle competition: the transaction history from 2018/09/20 to 2020/09/21; the product with price and simple descriptions; images that are corresponding to each article_id; metadata for each customer_id and article_id.
We will analyze existing models from previous research to find the best fit for our problem and data. Then we will use Pythorch to reproduce the models. The successfully reproduced models will be trained with the complete training dataset and then will be tested with the test dataset constructed from the original dataset. The top performers will be submitted to Kaggle to evaluate.
The models are classified as follows:

Recommendation with Neural Building Blocks:

  • Multilayer Perceptron Based Recommendation
    1. Neural Extension of Traditional Recommendation Methods
    2. Feature Representation Learning with MLP
    3. Recommendation with Deep Structured Semantic Model
  • Autoencoder Based Recommendation
    1. Autoencoder-Based Collaborative Filtering
    2. Feature Representation Learning with Autoencoder
  • Convolutional Neural Networks Based Recommendation
    1. Feature Representation Learning with CNNs
    2. CNN-Based Collaborative Filtering
    3. Graph CNNs for Recommendation
  • Recurrent Neural Networks Based Recommendation
    1. Session-Based Recommendation Without User Identifier
    2. Sequential Recommendation with User Identifier
    3. Feature Representation Learning with RNNs
  • Restricted Boltzmann Machine Based Recommendation
  • Neural Attention Based Recommendation
    1. Recommendation with Vanilla Attention
    2. Recommendation with Co-Attention
  • Neural AutoRegressive Based Recommendation
  • Deep Reinforcement Learning for Recommendation
  • Adversarial Network Based Recommendation

Recommendation with Deep Hybrid Models：

  • CNNs and Autoencoder
  • CNNs and RNNs
  • RNNs and Autoencoder
  • RNNs with DRL

5. Brief Project Schedule

Task Assignee
Preliminary research on existing models All
Data preprocessing Lyu Zhang
Constructing models Zongshuo Wei, Hechen Wang, Yufu Liao
Model training All
Performance evaluation All
Research report writing All

6. Links and references

Project: https://github.com/ZonW/HM-Personalized-Fashion-Recommender
Kaggle: https://www.kaggle.com/c/h-and-m-personalized-fashion-recommendations
Reference: https://arxiv.org/abs/1707.07435