KptnCook data challenge

Objectives

1. Create an endpoint to get the number of times user x listenned to artist y:

   • user_id
   • Number of listens

2. Create an endpoint to update the listens for artists y by user x:

   • user_id
   • Update of Number of listens
   • Save dataset to csv tu use it back later

3. Create an endpoint for each user to get recommandations such as:

   • Random artists: taking a sample of 5 artists
   • Still unknown artists: taking a sample of 5 never listenned artists
   • Artists that are similar to the ones already listenned

Environment configuration

To build the docker image:

docker build -t kptncook-data-challenge .

To check vulnerabilities from docker image:

docker scan kptncook-data-challenge

To run the docker image:

docker run -d --name kptncook-data-challenge_container -p 80:80 kptncook-data-challenge
out: 176b0cb9897582f3c923fd3c179cb700415762ddc96e081176adb726a3a681e5

Commands to run after a code update

docker stop kptncook-data-challenge_container
docker rm kptncook-data-challenge_container
docker build -t kptncook-data-challenge .  
docker run -d --name kptncook-data-challenge_container -p 80:80 kptncook-data-challenge

or simply execute run_docker.bat

To access api : http://localhost

To see the doc: http://127.0.0.1/docs

Recommendations based on similar artists

To recommend similar artists to the ones already listenned, here are the steps:

1. 5 already listenned artist are selected (knwon artists)
2. corrwith() Pandas' function is used to find each 5 previous selected most similar artist, based on a correlation rank.

corrwith() allows us to choose among 3 different correlation ranks : pearson, kendall, spearman.
If no method is precised, corrwith uses Pearson coefficient which works as follows:

Pearson correlation coefficient value	Strength	Direction
Greater than .5	Strong	Positive
Between .3 and .5	Moderate	Positive
Between 0 and .3	Weak	Positive
0	None	None
Between 0 and –.3	Weak	Negative
Between –.3 and –.5	Moderate	Negative
Less than –.5	Strong	Negative

In the results, mosts of recommendations are between 0 and .3. It shows a weak correlation, however not null.
After some tries with the other 2 ranking method, results appears to be sensitively the same.One hypothesis is that users of this dataset have not listened to enough music, which makes ranking coefficients less effective.

Acknowledgments

• Python FastAPI vs Flask
• Introduction to FastApi
• FastAPI in Containers - Docker
• Recommendation engine using pandas
• Pearson correlation coefficient