E-commerce Product Recommender

March 1, 2020

Report Abuse
This experiment shows a recommender system trained to recommend products based on user-interactions and user-product metadata
Scenario -------- The company in this experiment wants to recommend items that they think their customers will like based on three things 1. A customer's rating of existing products (known as the affinity score) 2. Features of a customer (such as hobbies and interests) 3. Features of a product (what product category they belong in) The task is to build a machine learning recommendation system that can learn to predict items that customers would likely rate highly. Data Sets --------- There are three data sets in this demo. First a dataset of customers and their "affinity scores", or their rating of each product. ![Affinity Scores][1] Second is a dataset of customer and the "user features", such as their hobbies or whether they have childern. ![Customer Features][2] Third is a daset of items and their "item features", which is a list of which department the product is listed upder. ![Product Features][3] Content ------- This demo shows how to build a recommender model using the matchbox recommender system provided by Microsoft Azure Machine Learning Studio This system allows you to do both **content-based** and **collaborative-filtering** type recommendations. Take note, in Train Recommender: 1. The more latent traits you use under 'Number of traits' in train recommender, the more accurate your recommendation system can be, at the cost of a longer computational time 2. Number of recommendation algorithm iterations refers to how many times the training algorithm loops over the dataset. The higher this number, the more accurate the recommendation system will be, at the cost of a longer computational time. This is usually set between 1 and 10 3. Number of training batchs refers to how many sub-batches should be trained in parallel. Set this to the number of cores your computer has (4 in this case) Outcome ------- As you can see, in Score Recommender, the top 5 item recommendations for each user can be generated [1]: https://raw.githubusercontent.com/johnangrs/aisgbricks/master/contoso_affinity_scores.png [2]: https://raw.githubusercontent.com/johnangrs/aisgbricks/master/contoso_customer_features.png [3]: https://raw.githubusercontent.com/johnangrs/aisgbricks/master/contoso_product_features.png