Hello :) Today is Day 108!
A quick summary of today:
- added a field to show generated code by codellama in the text2chart webapp
- covered recommender systems from XCS224W: ML with Graphs
Firstly, about the text2chart webapp
When I create the prompt for codellama, I give it info about the type of columns in the dataframe, using the below logic
It may not be the most robust, but it does fine for now. I thought a problem with the app is that the date column is read as numerical values, and tried to introduce the (now) commented if statement, tested it a bit and it did not seem to do anything.
Then, I decided to show the user what kind of code is being used to generate the shown graph (probably a feature I should have had from the beginning). And this actually made me realise that sometimes the generated code is actually not that good, so a more detailed query might be needed.
Example:
So if a resulting graph is not good, we can see what was wrong in the code, and give a more specific query. However the downside is, is that if this were a final product, it requires python knowledge from the user, which is not ideal.
Secondly, recommendation systems
Covered topics:
Modern recommender systems, top-k recommendations, recall@K, embedded-based models for recommender systems, binary loss, Bayesian Personalised Ranking loss, GNNs for RecSys, Neural Graph Collaboration Filtering, LightGCN, PinSAGE (the paper of which I read on Day 100)
That is all for today!
See you tomorrow :)
p.s. feeling a bit better today