COSTOMIZE is a university project (start-up mockup) that tries to help certain American business in the predicted number of clients they are going to have on a specific day, along with predicted income and necessary number of employees to cover the daily workforce. The model will be created from data patterns by SafeGraph company.
For that purpose we have chosen four representative American business a Subway, Walmart, Starbucks and Old Navy, Due to the fact that we only had 2020 and 2021 data we wanted to choose a place where COVID period was not extremely significant so we could use both years for model training. Our selection was Houston, Texas since the restrictions there were not as strict and prolonged as in other states and data patterns were note barely affected by COVID.
Mockup Success stories: Entre Datos
Dashboard examples:
• Entre Datos
• Entre Datos
• Entre Datos
• Entre Datos
The data displayed in the dashboard it’s part of the test data, so the predicted visits are from the test set (not seen by the model). We have limited the data to only one month because it’s just for mockup purposes.
Code repository: GitHub - angel-langdon/Project2021: Project from Project III subject from third year of Data Science Degree.
(Datasets have been removed from the repository)
DISCLAIMER: It is a university project, we are not going to further continue with it and we are not making any money with it. It is more oriented to build a Portfolio. However, if SafeGraph is interested we could work for you and explain how everything works.
I have already asked, but, if you consider that this kind of data should not be public I could make authentication in order to be only seen by my teachers and maybe Safegraph company.
@Angel_Langdon this is great, thanks for sharing. we’d like to explore turning this into a content marketing piece for SafeGraph. would your group be ok with that? we could turn it into a white paper or case study type of piece. or would you like to present in a webinar or video that we host on our website?
Yes, we are okay turning this content into marketing piece for SafeGraph.
We would prefer doing a webinar and we are available from June 20 ( right now we are doing uni exams)
@Angel_Langdon +1 this is awesome work!
I’m really curious to understand the details of your predictive models. Do you have any methodology posted anywhere? Glancing at your github it looks like you are using Lasso regression, but hard to parse what are all of the covariates / variables you are using.
Just curious:
How did you select your variables?
Did you quantify how good your predictions are?
Did you do any analysis to see which variables were most predictive? Any surprises?
Thanks so much for sharing!
@Ryan_Fox_Squire_SafeGraph Hi, thank you very much!
Here you have all our project detailed (University study) COSTOMIZE - Predicting the number of visits of a store by day
This may sound a little unprofessional, however at the moment we don’t know any better way to do it:
We tried different variables based on intuition and also model performance.
Yes, we did, quantify how good were the predictions using two metrics, R2 and MSE based on a validation set (never seen by the model)
Yes we did perform a basic analysis of feature importance using the coefficients of the models. There weren’t any surprises, the number of visitors is a highly correlated with the past number of visits of that store and also a bit correlated with the weather.
Right now our team is very busy with University exams and part-time jobs, however, around the 20th of June we could arrange a meeting with you and explain all the details if you want.
@Angel_Langdon I think your work is great. A couple of questions: (1) Have you noticed an increase in R2 or a decrease in MSE over the past month? (2) Have you tried using something like Facebook’s Prophet package to do the forecasting?