Success at this year’s Smart Hackathon

I joined Smart as a Senior Data Scientist just over a year ago, right as the 2021 Engineering Hackathon was taking place. I remember thinking that it sounded like an awesome event for a number of reasons:

• 24 uninterrupted hours to be innovative and concentrate on delivering a project
• You are provided with plenty of drinks and snacks (key!)
• Your colleagues are there to cheer you on throughout the event, meaning that the atmosphere is fun and you can meet and interact with a lot of new people

I was very keen to get involved in 2022 to showcase my skills and innovate new processes at Smart. When one of the back end engineers got in touch and described a Hackathon idea with great potential, we quickly formulated both a plan and a team to execute it. 

‍

‍

Our diverse team consisted of two QA engineers, a back end engineer and two data scientists. 

The idea was to use anomaly detection methods to monitor financial transactions processed through the Smart platform, harnessing technology to simplify complex processes. The solution would need to automatically identify any transaction that does not follow the normal path and might need manual help to be resolved.

The approach

What we wanted to address was clearly a data science problem. In fact, there is a lot of value to using machine learning algorithms rather than pre-programmed rules to support this functionality of the platform infrastructure. Using a machine learning algorithm would enable us to develop a solution that is agnostic to changes in business rules. 

In this case, we chose to combine two complementary techniques: unsupervised algorithms to learn new anomalous patterns, and supervised algorithms to continuously learn from past anomalies and enhance performance. 

Importantly, the data was not labelled, which meant that transactions were not identified as “anomaly” or “not anomaly”. Therefore, we needed to use an unsupervised learning algorithm to identify the most common patterns in the data. All records would then be assigned anomaly scores, based on how different they were to the most common patterns. In the event that a score exceeded a set limit, the record would be considered an anomaly. 

Using this process, we could label the data and then pass the new dataset to train a supervised learning algorithm. The resulting anomalies would be presented to the Operations Teams, who would check the anomaly suggestions and label them ‘true’ or ‘false’. This newly-labelled data would be used to improve the supervised learning algorithm, as shown below. 

The implementation

We wanted to make sure that we explored any new patterns in the transaction data, but also used already gathered information on known outliers to make our model continuously smarter. We combined a number of models, and by 9pm, we were happy with how the algorithms were performing and the outliers we found. 

After confirming several outliers with the QA engineers, we had a well-deserved coffee break and started building an interactive report showing the outliers we found. 

‍

The results

We worked through most of the night to stitch it all together, finishing our last team meeting just in time for the 10am Hackathon presentations. When the presentations started, we were all amazed at the standard of the different projects presented. In just 24 hours, teams from across the company had created some incredibly creative and innovative things. 

This made it even more incredible when our team was announced the Smart Hackathon 2022 winner! The judge’s feedback stated that they could see genuine commercial value to the business from our idea and how it had been executed.

It was such a fun experience – we had an amazing team and a fun project to tackle. The adrenalin and excitement meant that the lack of sleep had been worth it!

We’re really lucky to be given the opportunity to contribute to the wider business with our idea and we cannot wait to see where it goes. I also hope this experience will lead to more ideas coming from the business to the Data Team, so we can find other new ways to transform pensions, savings and financial well-being, across all generations, around the world. 

Find out more about the Smart Hackathon 2022 here.