Recommender System: Who knows You Better Than Yourself?

As Steve Jobs once said, “people do not often know what they want till they see it”.

Nowadays we are the target of innumerable messages which “advise” us goods or service acquisitions we would have never made had they not been pulled under our nose, therefore pushing us to new needs or waking up latent ones.

Some concrete examples*:

• Two thirds of films watched on Netflix are suggested to users by the famous video streaming platform
• Google News: 38% of click-throughs are originated by links suggested by the web giant’s algorithms
• Amazon generates a good 35% of its sales from suggestions to users

(*) Source: VentureBeat.com

Behind these focused suggestions there are systems which can be most complex: the recommender systems. These systems have the following objectives:

• Maintain customer’s loyalty
• Increase service users’ traffic volume
• Provide a user experience (UX) as focused on expectations as possible
• Widen market limits

But, above all, make people buy what we want to sell. From a more technical point of view, the problem a recommendation engine faces is developing a mathematical model or an “objective” fuction which can foresee how much a user will desire a product or service.

For every “u” user, it is necessary to choose the “i” element which maximizes the “objective” function.

There are several recommendation typologies:

1. based on content: “you have read a book, I suggest similar books”
2. based on correlation among users with similar characteristics in their purchase model

And it is right here artificial intelligence joins the game: the recommender systems which take advantage of artificial intelligence were born with the aim to advising users about objects they had never bought before! Similarity among users can be calculated in various ways:

• social network: the assumption is that two «friends» on a social network should have common interests
• past valutations: certain particular patterns are “discovered” from the analysis of past users’ valutations

How is it possible to know whether a recommender system works?
Here below some valutation meters:

• Precision: Which is the difference between a real rating and its prevision?
• Diversity: How diverse are recommendations?
• Coverage: Which is the space percentage for user elements that can be recommended?
• Serendipity: How surprising are pertinent recommendations?
• Novelty: How surprising are recommendations in general?
• Relevance: How relevant are recommendations?

That said, we cannot not consider the fact that recommender systems have a tremendous potential in supporting omnichannel sales, anticipating customers’ wishes.