Product recommendation for C2C marketplace with collaborative filtering ALS algorithm

Abstract

In this project, a machine learning recommendation model is created for an e-commerce company which runs a customer to customer business. The raw data consisted order reviews, order details, product like event information and product details. The explicit and implicit feedbacks are used together and a rating generation logic per user-product couple is applied to create the source data of the model by using Google Cloud BigQuery tool. The ALS algorithm which uses matrix factorization is applied for predicting the top items which have highest ratings for each user. PySpark which is Apache Spark’s python API is used for implementing the ALS model. The best hyperparameters are determined comparing the root mean square error results by using grid search and cross validation and 0.78 of RMSE is reached. The predictions for the empty ratings are sorted then top rated 10 products are taken as recommendations. The evaluation of the model is done by comparing those recommendations with the user preferences. The user preferences are specified by using averagely top rated product categories and most interacted product categories in count. The recommendations are observed to be consistent with the user preferences.