Clever Capsaicin is a recommendation engine built to bring you out of your comfort zone. Find and discover new hot sauces that were made just for your taste buds!

• Introduction
• Data Collection
• Content-based vs Collaborative Filtering
• K-Nearest Neighbors
• Cosine Similarity
• Clustering with K-Means and t-SNE
• Results / Demo
• What's Next?
• References

All I knew was Sriracha (or Rooster Sauce as I've seen it called) growing up. Whether I'm having a bowl of noodles or a bacon, egg, and cheese breakfast sandwich, you'd know I was putting Sriracha on it. As I began to venture deeper into the hot sauce world, I became more familiar with the household names like Tabasco, Frank's or Cholula, but, in my opinion, none of them were quite the same. I always found myself circling back to Sriracha.

I recently discovered a Youtube series called "Hot Ones." The underlying theme of the show is an interview of celebrities while eating chicken wings covered in hot sauce that become progressively hotter and hotter. It was only natural me to open this Pandora's box as the show had introduced many new "artisanal" sauces I had never seen before.

That brings me to the motivation of this project. If any hot sauce enthusiast out there is like me, then you're always struggling to fight the masochist in you while seeking a balance of heat and flavor. How do I find hot sauces like Sriracha but hotter and lesser-known?

Even if you don't have a high tolerance for spicy food or suddenly feel like stepping out of your comfort zone, you only have to like 1 sauce for a recommendation. After all, everybody could use a little extra spice in life.

The images and metadata used in the this project were scraped from Heat Hot Sauce Shop and Amazon and stored via MongoDB.

The features used include:

• Ingredients (type of peppers used, fruits, etc.)
• Flavor profile (smoky, sweet, tangy, etc.)
• Heat level (subjective - 5 levels: Extreme, Extra Hot, Hot, Mild, Medium)
• If the sauce was featured on a show/blog

1. A challenge I came across when cleaning the data was the ingredients section had different names for the same ingredient. An example of this would be the infamous ghost pepper. The ghost pepper is also known as bhuk jolokia. Another example is the hatch chile pepper. While some would argue the following are different because of the location they're grown, for the purpose of this project, hatch chiles are synonymous with Anaheim peppers or New Mexico chiles.
2. The world of hot sauce is a very niche market. IMBD exists for movies reviews and Yelp for restaurant reviews, however, there isn't a single source with abundant user data for hot sauce.
3. How do you evaluate the performance of a recommender system? Unlike some other machine learning models, there aren't necessarily observed values to compare the outputs of the system with. What does it really mean to have a correct or incorrect recommendation?
4. The biggest challenge a recommender system faces is the "Cold Start Problem."
   • How do recommender systems handle new users?
   • What about new items?

The 2 common ways to approach the design of a recommender system are: comparing the item-to-item similarity and user-to-user similarity. This is also known as content-based and collaborative filtering, respectively.

Examples:

• Pandora Radio

• Spotify

• Netflix - Hybrid

• explain both types and differences and why content is better

• hot sauce recommendations using content vs collaborative

• Vectorization
• TF-IDF Vectors

It would be difficult to combine user reviews from different websites, for example, Amazon and Heat Hot Sauce Shop, as the two population of users are not homogenous. Heat Hot Sauce Shop allows users to review the hot sauces in their database whereas Amazon has a different collection of hot sauces and is not limited to hot sauce reviews. In this case, giving a recommendation based on hot sauces a user has tried before is more practical than giving a recommendation based on other users with similar taste. I will elaborate further on this in the next section.

• http://infolab.stanford.edu/~ullman/mmds/ch9.pdf
• https://beckernick.github.io/law-clustering/
• http://blog.untrod.com/2016/06/simple-similar-products-recommendation-engine-in-python.html
• https://datajobs.com/data-science-repo/Recommender-Systems-%5BNetflix%5D.pdf