# Insight Week 2

written by Eric J. Ma on 2017-06-10

This week has been intense, mostly because I knew in advance that I'd be spending two days wearing a fancy hat, funky robe, and yellow sash. Because I was missing two days of Insight, I had to get my Minimum Viable Product (MVP) out by Wednesday - thankfully, I did!

If you've followed my blog post series on Insight (this is the 2nd post), my project is forecasting influenza sequences. This week, I hacked out my MVP and deployed it to Heroku as a hybrid HTML report + dashboard. I also picked up and incorporated a few new things along the way.

The first is the use of tooltips on my Bokeh plots. Bokeh is really powerful, and in some of the exploratory analyses, I desired having tooltips as a UI element to help a reader (who might need some introduction) understand the nature of the problem and the data involved.

The second is further mastery of Bootstrap CSS & JS. Now, Insight's Program Directors have told us clearly that we're not gunning to become front-end designers (and the likes). Keeping that in mind, I still think it's important to know at least one front-end framework well enough to produce pleasant-looking interactive tools or reports - knowing front-end elements potentiates us to communicate with front-end designers on final data products.

For the MVP, I tried further experiments with the Grid layout and Modal JS. The key idea behind Grid layouts was easier to grasp - prioritize rows, then columns.

With the Modal, stepping back for a moment, my goal was to display the science behind the project. However, it gets really technical. My audience is probably going to fall into one of two personas: the "business person" who just wants to see the final result and doesn't really care about the techniques, and the "technical person" who wants to dig deeper. I chose to use the Modal effect to satisfy both. The scientific methods are described at a high level on the main page, and the Modal element is used to show further information, graphics, and the likes.

The third was deployment to Heroku itself! David Baumgold first showed me how to use Heroku at PyCon 2016, but I could never wrap my head around it at first. I think I didn't understand how "deployment" worked. A year later, stuff that DB taught me came to fruition, as I hacked on deploying a minimal Flask app to Heroku with my younger brother. That gave me enough of the Heroku-specific concepts to hack together the necessary requirements.txt and Procfile files to deploy Flu Forecaster to the web.

For next week, these are my plans:

• Solve the problem of "average sequences per quarter" being ~10 amino acids different from actual sequences. Two approaches I'm thinking of:
• Use GPflow to train a Sparse Variational GP regressor on my dataset. This should allow me to scale up the forecasts beyond 67 quarters and into 700+ weeks, which is a greater time resolution and thus will give me greater sequence resolution in forecasts.
• Try forecasting variance in addition to mean coordinates. This will give me a full probability distribution to sample from, which may allow me to stick with PyMC3 and quarterly forecasts.

Still not sure which of the above two approaches are the better one, so I'll be sure to give each a shot.