12.2. Recommendation Systems and Profiles

So, let us look at the following situation, very often when you go to an online store, you find a recommendation. For example, it would say, the customers like you who were interested in books like these or customers who bought this phone also look for this pair of head phones. So, these services are recommended to you based on your profile, your online service maintains some profile information about what you like and what you do not like. It compares what you like and do not like with others and identifies the similar category of people.

Recommendation systems are tailored services that suggest products to users based on their preferences. When you shop online, you may see suggestions like 'customers who bought this item also liked...'. This happens because the system gathers data about your interactions—what you’ve purchased and what you’ve viewed—and compares it with the preferences of other users with similar tastes. The goal is to present you with products or services that you are likely to be interested in, enhancing your shopping experience.

Imagine you go to a library where your friend works. Based on your previous visits and the types of books you've borrowed, your friend might suggest new titles that others with similar interests have enjoyed. The library keeps track of your preferences and uses that information to guide you.

So, one instance of this is when you have preferences over things like movies or books, so suppose over the sequence of time, you have gone to some website and entered your preferences about movies that you watch. So, say there are five movies, let us just call them A, B, C, D, and E which both you and somebody else have ranked on this website.

In this example, you and a friend both rate the same five movies (A, B, C, D, E). Even though you both have an opinion on each film, the ranking may differ. For instance, you might rank movie D as your favorite, while your friend ranks movie B at the top. This variation in ranking helps identify how similar or different your tastes are from others, which is crucial for the recommendation engine to function correctly.

Think of a dating app where users rate potential matches. If User 1 ranks Match A highly but User 2 ranks it low, the app will note this discrepancy in preferences. Just like finding friends with similar tastes in movies, the app tries to match users based on who resonates with whom regarding their ratings.

So, what we are trying to do is measure the dissimilarity in terms of what we call inversion. How many pairs of movies or rank in the opposite way between you and your friend?

An inversion refers to a situation in your rankings where you prefer one movie over another, but your friend does the opposite. For example, if you rank Movie D higher than Movie B but your friend ranks Movie B higher than Movie D, that forms an inversion. By counting these inversions across pairs of ranked items, the recommendation system can evaluate how similar two users' tastes are. The fewer the inversions, the more aligned the preferences are, making it easier for the system to suggest recommendations.

Consider a sports team. If one coach favors aggressive strategies (lets say an offensive play) while another prefers defensive ones (a cautious play), their differing strategies can be viewed as inversions. In sports, having too many inversions might mean the coaches need to work on synchronizing their playbooks to ensure the team functions more effectively.

So, if you and your friend rank every pair of movies in the same order, then your total order of performances must be the same. So, if there are zero inversions, then you have exactly similar in your taste to your friend and the rankings are identical.

If you and your friend have zero inversions, it means you have ranked all movies in the same preference order. This alignment suggests that the recommendation system could reliably suggest movies that both of you would enjoy since your tastes are identical. However, if there are numerous inversions, the system must carefully analyze these discrepancies to make better and more tailored recommendations for each user.

Think of a pair of friends planning to watch a movie. If their preferences align perfectly, they can easily agree on what to watch next. However, if they frequently choose different genres or types of movies, they might have to search more meticulously to find a film both would enjoy.

So, we can formulate this in another way. So, now, we take our ranking and we assume that is the given order. So, we pick the certain order for the movies and we call that the basic ranking 1, 2, 3, 4 up to n.

To further analyze ranking preferences, we can label the original sequence of movies with numbers representing their ranks. When comparing your ranking to a friend’s, we can identify which pairs are inversions by checking if one rank precedes another. By analyzing these ranks more formally, we can establish metrics for recommendations based on the patterns of inversions found during this comparison.

Imagine teachers grading assignments. If Teacher A gives student papers scores of 1 to 10 based on one set of criteria and Teacher B bases scores on a different standard, comparing their grading patterns becomes necessary to spot discrepancies (or inversions) and adjust teaching methods or review sessions accordingly.

--