What is Machine Learning?
A mathematician would define ML as a field that intersects statistical, probabilistic, computer science and algorithmic aspects arising from learning iteratively from data and finding hidden insights which can be used to build intelligent applications. Hence it is nothing but a dynamic mathematical formula generated by a machine to predict a certain result with optimisations built in to maximise success or minimise error(minimising error is equally important because hypothetically any model can be applied to any problem). Whereas A statistician’s definition of ML would be that a machine is able to generalise predictions from its experience in the past for any practical situation.
Machine Learning is a Sub-stream of AI (a 25 yr old technology) which was researched and built for the purpose of making computer systems that can think and act like humans. Evolution of ML from AI started when Neural networks was born specifically Artificial Neural Networks which eventually gave birth to ML.
Broadly ML can be divided into 3 different streams viz a.) Supervised learning- Give machine a labelled training data as input and let it predict outputs in the form of certain features carrying a distinct probability, b.) unsupervised learning- Give machine an un-labelled data, let it figure out the input and output variables to solve a given problem c.) reinforcement learning- where a machine learns from a dynamic environment and performs certain goals.
Another class of ML problems/solutions include Classification(binary/multi-label),regression(something that spits out a continuous output), clustering(find buckets- where groups are not known before hand), density estimation(used widely in market research/healthcare) and dimensionality reduction(NLP).
In the last few years ML is being used worldwide to solve large number of critical hard problems with greater efficiency than the existing legacy solutions. Mostly because of availability of platforms that facilitate faster development of ML algorithms. Same ML solution that took 10k+ lines of code 3-4 years back can be done with less than 1k lines of code using platforms like tensorflow from google, Amazon ML, IBM Watson, Microsoft ML Studio, open.ai(YC), Stanford ML Research Lab , FB(M or wit.ai) and numerous other proprietary self built platforms.
Algorithms driven by ML are all around us, even if we realise it or not. Some real world examples include fb-google-insta news feeds, Gmail tabs and spam filtering, YouTube song suggestions, Amazon homepage recommendations, news inshorts feed, self driving cars, netflix recommendation engine. Google Search results, spam filtering, estimating credit risk, preventing fraud, among a host of other things we consume on a daily basis.
At Slice we are solving lots of hard problems through the use of AI and ML across credit risk, marketing, product and operations teams. Use of mobile and digital footprints to derive alternate data points which act as proxy to traditional credit risk signals enables us to apply Machine Learning and Data Science in the the financial services business to make consumer credit accessible to people who do not have a credit history. They are the same class of People who also don’t fulfill criterias that banks/nbfcs/traditional financial services institutions have like an active bank a/c(90% economy is still cash), income proof/salary slip, credit history in bureau (via cibil, high mark, equifax, transunion etc.) hence credit for them is highly inaccessible.
Some relevant proprietary solutions that we have designed specifically in credit risk management are listed below-
- Credit risk variables (hotspot) profiling for loan default analysis
- Credit risk predictive analytics for predicting default probability and deciding credit limit, interest rate and maximum allowed emi tenures through multiple models primarily using regression, neural networks and random forest models.
- Predictive Mathematical modelling of Financial risk for loan portfolio monitoring- LGD and EAD Model
- Internal Credit scoring and Portfolio Grading
- Detecting Identity Theft(and any other Fraud) at the time of credit line application using image recognition algorithms
- Detecting Fraud at the time of any credit transaction through the use of big data engineering and n dimensional analysis
I hope it gives you a fair idea of how Machine Learning and AI is being used to solve real world problems especially in designing key business solutions to reduce operational costs and in risk management in the financial services industry. I would love to hear your thoughts on the ever increasing applications of Machine Learning algorithms in real world so feel free to leave your thoughts in the comment section.