Robyn Dash

Inspiration

Selecting a Robyn model is partly statistical, but also partly contextual. For instance, traditional channels usually bought in large, fixed packages, which limits their flexibility. Marketing activity such as scheduled product launches, sales and limited edition items can also affect media performance.

An interactive dashboard for Robyn can enable marketers to better select models that reflect the real-world business context.

What it does

Robyn Dash contains three plots:

1. NRMSE against decomp.RSSD to evaluate model accuracy, colored by cluster for context on how each model evolved.
2. Per-channel ROAS, calculated by total result vs total ad spend, for a high-level overview of channel performance for each model
3. Marketing activity effect, calculated by marketing activity effect mapped to activity dates, to understand how each activity affected performance.

How we built it

Built on RStudio after deep diving into the source code to understand the process for collecting inputs, processing them, creating plots and the contents of each output file. After that, I identified the main areas I wanted to get more insight into based on the available data. For instance, evaluating models beyond the top 1 per cluster, better side-by-side comparison and visualizing marketing activity by category.

Challenges

There were some initial challenges getting Robyn to run at first, due to an issue with dates in the experimental version, but it was fixed a few weeks back.

I also had zero experience with R before this hackathon.

Accomplishments

I achieved most of what I wanted to do with Robyn Dash, especially the third plot which can provide insights to create better marketing campaigns to support media spend.

The feature to compare models beyond the top 1 per cluster is also helpful.

Learnings

I learned a lot about R while creating the dashboard. It's pipeline handling is much more efficient than mass chained functions in other languages, and I can see why it's the language of choice for many people working with data.

I also learned a lot about Plotly and Dash functionality. The added interactivity really helps with navigating complex datasets.

Finally, I learned a lot about machine learning Robyn itself, as well as Prophet and Nevergrad.

What's next for Robyn Dash

Get some real-world testing first, then see how to refine it further.

Additionally, Dash was surprisingly powerful and easy to work with, and I can see it potentially powering an end-to-end dashboard all the way to calibration and future planning.

Built With

• dash
• plotly
• r