In the past week, I made slight progress with my toxicity model, prepared for Research Showcase, and met up with my mentor. After trying an XGBoost model with a new fingerprint (EState) and a weave model, I found a working combination. Averaging two XGBoost models and a graph convolutional network, I got a mean absolute error of .001 less than without the Estate indices. It’s only a little bit better, but progress is still progress. I still need to implement validation with the XGBoost models—I’ll likely use k-fold cross validation to perform it efficiently on the relatively small dataset. I also went on an etiquette lunch, which was a really nice break to have.
I’m excited for Research Showcase, and I’m generally not worried about explaining my topic because I’ve had practice talking about my work to other students. However, I do feel a bit unprepared to discuss computational chemistry with younger kids. For this, I brainstormed with my mentor about how to approach my topic. I decided on relating the prediction of chemical properties to a common object like Playdoh and its stretchiness. We also discussed the types of ensemble learning as well as next steps: trying to extend my toxicity prediction model to other datasets to see if it’s useful in general and not just for a specific task. I’m looking forward to trying my model on more data and presenting at Research Showcase!
Comments