Recommendation systems are important in many production systems. While most commonly used in e-commerce or video applications, they're present in a variety of other cases.
Some examples include Netflix and Amazon, which use Carousel based recommendation systems and Linkedin that uses your network to make content recommendations to you.
There are 3 broad categories of personalised Recommendation Systems
- Collaborative Filtering : Where we use a matrix of rankings to determine a vector of User and Item factors. This subsequently helps us to be able to calculate the probability of a successful recommendation
- Content-Based Filtering: Where we use embeddings to learn the semantic meaning of the items and the users. These are learned embeddings and are derived from a separate process as compared to collaborative filtering.
- Hybrid Models: These combine the two of them
We want to use Collaborative filtering systems when we have a good amount of data because they are cheap and fast to train. We want to use content-based filtering when we have no user data and are starting from zero. Ideally, we combine both to get information from user signals and content signals.
Often times we'll see what's known as a two-tower architecture
In this architecture, we learn a separate encoding for the item and another for the user. This is then projected into the same latent space and we can calculate a similarity score based off this.
![[Pasted image 20240417224503.png | 400]] This might be used as the final step as a way to filter out irrelevant items.
When looking at recommendation systems, we want to look at different metrics. Two useful metrics are precision and recall. Often times, we might take a metric@k which means that we perform the calculation on a slice of k items from the predictions.
Say for instance we had 5 known relevant items and fetched 10 predictions, 2 of which were relevant. This would mean that our precision would be
