Jonny Brooks-Bartlett – Data Scientist at deliveroo
They use predictive modelling and machine learning.
Their hypothesis is “if they can show you more relevant places you will spend money”
given a list of restaurants, can they rank them optimally
how do they quantify that?
how do you turn that into a machine learning problem?
list of sessions, 0 if they didnt buy, 1 if they did
use a pointwise approach (ask the question for each item in list)
target variable is ‘how likely are you tobuy’
you then pick attributes to add to a model to see if they have an effect
start simple, doesnt need o be machine learning, can be a heuristic (eg eta + popularity) then itterate
it doesnt have to be perfect
Evaluating models:
offline metrics (evaluating before they in to production)
MRR
precision at k
etc
they use MRR
The actual workflow:
track data, write sql queries, make models, generate MRR, pick the one with the highest MRR
they use circleci and deploy
then they do AB tests and split 50/50, if its better, roll out, else revert.
then itterate. constantly.
they write in python, save models in tesorflow, and then read in production language.
amazon sagemaker – can train models and can deploy at scale too
read googles rules of machine learning.
Leave a Reply
You must be logged in to post a comment.