In this project, I use the Surprise package to create a recommender system using reviews of products in the "Watch" category on Amazon. This project and all the contents were developed and published for educational purposes only.
Many businesses can benefit vastly from implementing some sort of recommendation engine. I wanted to complete this project to show the relatively low amount of effort can go into such a fruitful internal product.
While recommender systems can be built in many different languages, I elected to use Python to build this one. I use a few different popular tools and libraries to complete this project:
- Jupyter Notebook 5.3.1
- pandas 0.21.0
- Python 3.6.2
- scikit-surprise 1.0.6
I'm using a public dataset of Amazon reviews. There's anonymized customer information, but the reviews and products are real. There data sets provided by Amazon are split up by product category. I chose to use reviews of products in the "Watch" category. Since the data is so large, I didn't include it in this GitHub repo, however it is available online.
You can find my complete analysis in my [GitHub repo](Collaborative Filtering Amazon Watch Reviews). The file it titles Collaborative Filtering Amazon Watch Reviews
There were multiple hurdles I faced in this project. I detail them in my analysis, but I'll list them briefly here:
- I used my personal laptop, so optimizing parameters took a lot of time. ALso, I could not run a k-NN on this data because it was so large.
- The data doesn't have any product descriptions, so I could not do any sort of content filtering
- Many products have only been reviewed once, and many users have only reviewed one product, which leads to many sparse recommendations