Recommendation System Case Study

• Recommendation System Case Study
  • Objectives
  • Tech Stacks and Team Organization
    • Tech Stacks
    • Team Organization
  • The Case Study Description
    • Data Description
    • References and Hints Students will use collaborative filtering and matrix factorization in Spark or Surprise in this case study. Students are also encouraged to build a simple flask application.

Objectives

There are three aims of the following case study:

1. Improve your big data skills by including as part of your solution a working recommender written in Spark.
2. Demonstrate your knowledge of recommendation systems by using the library Surprise to tweak and customize different algorithms
3. Practice you communication by presenting to a non-technical audience

Tech Stacks and Team Organization

Tech Stacks

You can choose either to use Spark or Surprise for the recommendation system building. You can use both by properly allocating workload if time permits. If you choose to use Spark, the earlier assignments and lecture notes will be useful. However, Surprise is also a good choice and its API is very similar to scikit-learn.

1. Spark and the big data solution

   These are two useful links for the Spark approach.

• The official documentation

• A solution blog

  The second blog post uses MLlib instead of ML it is still a useful resource. ML is the more stable library and it is preferred to MLLib (if possible).

2. Surprise

   Suprise is a Python library that is designed for the recommendation system building. Surprise has a version of the data built in to it.

   Here is a blog post that uses Surprise to predict movie ratings.

   And the documentation is pretty good.

3. Flask

   A sample Flask template app.py file is included in the src directory.

Team Organization

The dataset is pretty clean so parallelization can be utilized to improve the efficiency. For example. If you have a team of 3 Alice, Bob and Carl, you can allocate tasks as following.

1. Alice: Test baseline models in Spark or Surprise and learn the workflow, not necessarily using this dataset.
2. Bob: Prepare and do EDA analysis of the dataset, communicate with Alice to prepare the dataset to feed into models.
3. Carl: Write the skeleton of the Flask application. Define the API to connect your teammates' model to the web interface. Test it with fake data for example.

Later the tasks can be reassigned to create readmes/slides, etc.

The Case Study Description

You and a team of talented data scientists are working for the company, Movies-Legit, who has used a production recommenders for many years now. The recommender provides a significant revenue stream so your managers are heisitent to touch it. The issue is that these systems have been around a long time and your head of data science has asked you and your team members to explore new solutions.

The solution that has been around for so long is called the Mean of Means. We see that some users like to rate things highly---others simply do not. Some items are just better or worse. We can capture these general trends through per-user and per-item rating means. We also incorporate the global mean to smooth things out a bit. So if we see a missing value in cell $R_{ij}$, we'll average the global mean with the mean of $U_i$ and the mean of $V_j$ and use that value to fill it in.

We would like you to use this as a baseline when making your comparisons. The basic code showing you how this is done is provided for you in src/baselines.py.

At the end of the day you are to present your solution to the bosses and the entire product team. You can assume that they have little to no depth of knowledge in statistics and big data technologies.

The main goal here is to improve the RMSE, however, another equally important goal is to present your model and the details of your explorations in a clear, but technically sound manner. You need to convince us that you understand the tools you are using AND that it is a better way forward. After all we are making a lot of money with mean of means and it is a major risk to swap it out with anything else.

We would also like you to include some discussion about how you would move from prototype to production.

Data Description

MovieLens is a classical recommender dataset and you can assume your audience is already familiar with it.

There is a larger version of the dataset available from this (link](https://grouplens.org/datasets/movielens), but we do not expect a production ready recommender in only a day's time so do not worry too much about scale for now. You will find the smaller version of the data as part of this repo. Read the README.txt to understand the data description.

References and Hints

• A challenge with the Spark solution can be to implement a mean of means baseline
• Surprise can be challenging to shape into what you might want, but it has a lot builtin
• Make sure you show an example of a prediction
• The numeric results are not as important here as how well you communicate what you have done
• Finally, the 'bosses' are not real keen thinking too hard about RMSE or MAE. You might be able to explain it to them, but if you report it as percent improvement over mean-of means they are more likely to listen.

Good luck!