Customer support · Production

Monzo applies Recurrent Neural Networks to predict customer support questions and detect signup fraud

The problem

Monzo's two-person data team had many promising machine learning applications to evaluate but traditional methods required weeks of data manipulation before a model could be trained, making rapid experimentation impractical.

First attempt

A traditional data exploration approach to fraud detection — manually constructing features and looking at correlations — failed because of too much noise in the data and difficulty finding time-based patterns.

Workflow diagram · grounded in source
1
User app events recorded
trigger
“the time series that documents users' actions within our app”
2
RNN finds temporal signals automatically
ai_action
“they require only the "raw" time series as an input, and they find relevant temporal signals automatically”
3
Support question category predicted
ai_action
“correctly determine the question's category in 30% of cases based on the preceding 200 events”
4
Proactive help page served
output
“we could use this model to build a Help page that would instantaneously answer our users' questions without them even needing to ask”
5
Signup fraud scored
ai_action
“detect 40% of fraudulent accounts with a corresponding false discovery rate of 42% purely based on onboarding events”
Reported outcome

RNNs achieved 30% top-1 and 53% top-3 accuracy predicting customer support question categories, and detected 40% of fraudulent accounts with a 42% false discovery rate.
New ideas can now be tested in one or two days, more than ten times quicker than manually designing signals.

Reported metrics
Support question top-1 accuracy30%
Support question top-3 accuracy53%
Fraudulent accounts detected40%
False discovery rate for fraud detection42%
Show all 6 reported metrics
support question top-1 accuracy30%
support question top-3 accuracy53%
fraudulent accounts detected40%
false discovery rate for fraud detection42%
time to test new ML ideaone or two days
speed vs manual signal designmore than ten times quicker
Reported stack
GRUTFLearnTensorflow
Source
https://monzo.com/blog/2017/05/03/practical-machine-learning-for-startups
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

RNNs achieved 30% top-1 and 53% top-3 accuracy predicting customer support question categories, and detected 40% of fraudulent accounts with a 42% false discovery rate.

What tools did this team use?

GRU, TFLearn, Tensorflow.

What results were reported?

Support question top-1 accuracy: 30%; Support question top-3 accuracy: 53%; Fraudulent accounts detected: 40%; False discovery rate for fraud detection: 42% (source-reported, not independently verified).

What failed first in this deployment?

A traditional data exploration approach to fraud detection — manually constructing features and looking at correlations — failed because of too much noise in the data and difficulty finding time-based patterns.

How is this customer support AI workflow structured?

User app events recorded → RNN finds temporal signals automatically → Support question category predicted → Proactive help page served → Signup fraud scored.