Data Science

There are two primary properties of customer ratings:
1. Type of feedback (i.e. implicit or explicit)
2. Sparsity
Typically, the goal of a recommender is to make recommendations about a customer's affinity for a given product
- Customer affinity refers to how much a customer likes a product
An explicit rating is the best measure of customer affinity for a given item
- An explicit rating is just a rating explicitly defined by a user
  - E.g. a rating between $1$ - $5$
- They're the best way to measure customer affinity because the customer has defined their affinity for a product themselves
However, we often don't have explicit ratings for products, but still want to make recommendations
In this case, we can use implicit ratings to measure customer affinity for a given item
- An implicit rating refers to a metric we use as a proxy for a true, customer affinity (or explicit rating)
- For example, we may use sales as an implicit rating
- Unlike an actual explicit rating made by the customer, sales doesn't explicitly represent a rating
  - Instead, it's just the best proxy we have
Implicit feedback is usually thought of as confidence in a preference, whereas explicit feedback is thought of as the actual preference

The second important property of a rating matrix is sparsity
A rating matrix is inherently sparse because any single user interacts with only a tiny fraction of the available items
So, each row of the matrix contains only a few known ratings and all other values are missing
This means there is a disproportionally large number of known ratings corresponding to a few of the most popular items
Whereas, niche product ratings are especially scarce

Recommendation methods can be categorized in a number of ways
They're categorized by the type of predictive model and its inputs
The two main families of recommendation methods are:
- Content-based filtering
- Collaborative filtering
Content-based filtering relies on content data
- E.g textual descriptions of items
Collaborative filtering relies on patterns in the rating matrix
Both approaches can use either of the following strategies:
- Formal predictive models
- Heuristic algorithms searching for a neighborhood of similar users or items
In addition to these core methods, there is a wide range of solutions that can be used together to create hybrid models
We can also extend the core methods to account for contextual data in contextual models
Or, we can detach the core methods to account for non-personalized models

recommendationmethods

The hierarchy of recommendation methods look different if we focus on the purpose of them
Specifically, we can categorize the usage of methods in two dimensions:
- Level of personalize
- Usage of contectual information
The core recommendation methods (i.e. content-based and collaborative filtering) are mainly used for personalized and context-unaware models
- They make recommendations based on historical interacations of items between users
- These recommendations are often displayed in sections like:
  - You might also like
  - Your favorites
  - Inspiried by your browsing history
  - Buy it again
Recommendations can be even more personalized by taking into account contextual information
- E.g. user location, purchase day of week, etc.
- These recommendations could be like the following:
  - Restaurants near you
  - Destinations you might like for your next trip
Whereas, non-personalized methods take a different approach by relying on global statistics and item properties (rather than personal profiles)
- These recommendations could be like:
  - Most popular
  - Trending
  - New releases

personalizedrecommendations

Note, personalized and non-personalized recommendations can be combined together in many different ways
For example, personalized recommendations selected based on historical users' interactions can be sorted by popularity
- Alternatively, the most popular items can be selected within a category of products preferred by the user
Lastly, non-personalized recommendations can also be contextualized by user location or marketing channel
- For example, a product page could include recommendations like:
  - Frequently bought together
  - More like this

Quality Evaluation