At the fourth edition of the Machine Learning Developers Summit (MLDS), Intuit’s Sidharth Kumar took an in-depth look at the features, metrics, challenges, and potential of building a personal financial management app like Intuit Mint (formerly known as Mint.com). , aimed at users in the US and Canada.
As Intuit’s Principal Data Scientist, Kumar is responsible for augmenting Intuit’s personal finance platform, Mint. At Intuit, he has helped launch several successful AI-powered offerings, including cash flow forecasting, automated subscription discovery, automated budgeting to maximize savings, transaction categorization, abnormal spend detection, and automated customer cohorts.
What is mint?
With 4 million active users, Mint helps its users manage expenses, budgets, subscriptions, etc. As a personal financial management app, Mint has over a decade of user transaction data from tens of millions of users. Transaction data includes bank accounts, money management accounts, retirement or investment accounts, credit cards, merchants, and other financial services. In addition to these, other data also includes clickstream, demographics, geolocation, derived features like aggregates, and streams.
Making decisions and predicting outcomes using this data is easier said than done. Kumar explained the strong and growing competition in the space, along with challenges such as data integrity, the dynamic nature of transaction data, user fatigue, and legacy systems.
“We had to deal with a lot of legacy systems, which we migrated to AWS,” Kumar said. In addition, they used EC2 inference nodes for real-time qualification (tens of millions of transactions qualified per day), along with EMR or node pools for better processing (training up to 100 and millions of ML models in a few hours or several hours). x 1000 core clusters), on-device or federated learning (iOS), etc. For data storage, Intuit Mint currently uses Hive, DynamoDB, Redshift, S3. Additionally, the company leverages Python, Pyspark, R, SQL, and SWIFT (OS) in terms of programming languages.
Essential business metrics include voice of the customer (VoC), retention, click-through rate (CTR), net promoter score (NPS), profit and loss (PnL), relative rankings, and other capabilities with respect to the competition.
AI metrics include: accuracy (cash flow); integrity (recurring transaction discovery); accuracy, recovery and AUC such as credit default and anomaly detection; cross-entropy (categorization); personalized (MCMC) for optimization of savings; AIC, KL-divergence (particularly clustering), etc.
Potential ubiquitous extensions of AI
“Fintech is the next big thing. So, we all know it. Next is to automate customer profiling and relative spend classification,” Kumar said.
The potential (along with the techniques used) use cases include credit default/late payment forecasting (logistics), cash flow forecasting (SVM), anomaly detection/large transaction alerting (GB), transaction categorization (NN), commitment maximization through route recommendation (boost), trader change (GMM +MAB), spending recommendations (boost), trend patterns and user or overall breakdown analysis (statistics), financial well-being improvement (MAB), CLTV (survival) and marketing optimization and ad targeting (NN).
Kumar explained the techniques behind automated customer cohorts and relative spend classification (Auto Encoders + GMM), Invoices or Subscription IDs (FFT), can I afford this? (RF) and automated budgeting (bundle + custom optimizer).
“If we give our customers a simple way to ask Mint if they can afford a certain purchase in a certain category or merchant, then Mint will be able to provide advice based on users’ current spending pattern, helping them make quick purchases. . spending decisions,” she said. Mint uses on-device RF (Swift, CoreML, iOS) for feedback learning and a multivariate statistical model for seeding.