What are the Hypotheses Behind the Main Solutions of Recommender System

Boost sales with machine learning and statistics powered recommender system

Alina Zhang


Image by author

To buy, or not to buy: that is the question.

If you were a manager of a grocery store, what products customers never bought but may like? If you run a restaurant, what is the food clients never ordered but may like? If you worked in sales or marketing, what are the books, movies, software, insurance, or t-shirt users never purchased but may like?

Why business needs recommender systems

Boosting sales to the next level by recommending the right products to the right customers, a hybrid recommender system is a promising approach.

Items bought vs Never bought but may like: the recommender system was born with the mission to figure out the answers. Powered by machine learning and statistics, models are able to discover the hidden association between products from historical orders and the underlying patterns in latent features using the users-items matrix.

The hypotheses behind the main solutions of recommender systems

The main solutions with industry proven results in recommender systems are as follows:

  • Content-based recommender
  • Collaborative filtering
  • Popularity
  • Association rule

The first two methods are mostly based on machine learning algorithms while the rest two are lean on statistics.

Content-based recommender

Content-based recommender assumes what user liked in the past, he or she would like it in the future. A content-based recommender learns customers’ interests based on the historical orders, then recommends similar items to the customer. For example, Mr. Panda ordered lots of bamboo in the past, the content-based algorithms which clustered bamboo shoot as a similar item to bamboo would suggest it to Mr. Panda.