Modeling Fashion Recommendations with Sketch-Based Model

This repository contains our solution to FARFETCH Fashion Recommendations Challenge that achieved 3rd place.

The aim is to predicit the products clicked by a user from a list of selected recommendations. First, we use Cleora - our graph embedding method - to represent products as a directed graph and learn their vector representation. Products are embeded in two relations:

• Only clicked products in a given session (clicked-modality)
• Viewed products in a given session (viewed-modality)

As a results we obtain two embeddings. No all products were clikced or viewed, so for a small number of proucts we do not have a vector representation from Cleora. Next, we apply EMDE to predict the product based on previously clicked and viewed products. We also add some features associated with each user.

Model takes as an input 4 sketches:

1. Sketch of all products clicked in the previous sessions (clicked-modality)
2. Sketch of all products clicked in the current sessions apart from the current query (clicked-modality)
3. Sketch of all products viewed in the current sessions apart from the current query (viewed-modality)
4. Sketch of all products displayed in the current query (viewed-modality)

Model return a sketch of a product clicked in the current query. Output sketch is from viewed-modality as it contains more product embedding/codes. The output sketch is then scored against all product sketch from viewd-modality and click probiablity is obtained.

Requirements

• Download binary Cleora release. Then add execution permission to run it. Refer to cleora github webpage for more details about Cleora.
• Python 3.7
• Install requirments: pip install -r requirements.txt
• GPU for training

Getting Started

1. Create data directory in src folder:

   mkdir src/data

2. Put train.parquet, validation.parquet and test.parquet into src/data folder

3. Change directory to src

   cd src

4. Transform all parquest files to CSV with sequential-like form. It also creates input files to Cleora:

   python transform_to_sequential_data.py --data-dir data          
   
   

   This script will create three CSV files: data/train_original_processed_reproducing.csv, data/val_original_processed_reproducing.csv and data/test_original_processed_reproducing.csv And two input files for Cleora algorithm data/cleoraInput_sessionIdGrouped_viewed, data/cleoraInput_sessionIdGrouped_onlyClicked. Script also creates and saves dict with products2attributes data; at data/products_dict_reproducing.

5. Create datapoints for running the model:

   python create_datapoints.py --data-dir data          
   
   

   This script will create three files: data/train_datapoints_sequential_reproducing, data/validation_datapoints_sequential_reproducing and data/test_datapoints_sequential_reproducing.

6. Compute product sketches using Cleora and EMDE

   python encode.py --data-dir data    
   

   This script will create LSH codes for each product from viewed-modality and clicked-modality. Codes are saved to data/codes_viewed and codes_clicked

7. Run training

   python train.py --data-dir data    
   

   Logs are saved to: src/logs/runs

8. Download trained model checkpoint: https://drive.google.com/file/d/1vnuKZGdEGHzGkBrVUx7JNbyOcqE-OK5o/view?usp=sharing

9. Run test. Use flag checkpoint-path to specify trained model path; model_trained.ckpt by default. Flag --subset-to-use to specify whether to use validation or test subset; test by default.

   python test.py --data-dir data