|
Data Science Colloquium Series 2-3 pm, NKN Hall Image based encoding for financial transactions Amit Godbole Fortiate, Pune. 01-11-19 Abstract This invention belongs to the financial field where monetary transactions take place between a merchant and a consumer, between a sender and a beneficiary or between two account entities in a system. Our business objective is to create a product for financial institutions to enable them create a behavioural model for the transacting entity and use the model to score a future transaction done by the entity as it takes place, in real-time. Every payment transaction follows a global standard specification as it travels from the point of origin (i.e. place of demand - where goods or services are required) to the point of destination (i.e. place of supply - where resources/funds are available). There are several data elements in these specifications which carry attributes about the physical situation in which transaction takes place. These data elements are heterogeneous in nature e.g. amounts, date & time, factors etc. Such heterogeneous data requires huge amount of data engineering effort to bring it to a state so that standard machine learning algorithms can be applied to it. Moreover it is also resource intensive in terms of the processing power and data storage required when it comes to training a machine. We intend to propose an alternative way of processing the financial transaction data there by reducing the effort and cost involved in data engineering and modelling. The idea is to encode the transaction data in an “image” form. Literally what we are proposing is to recreate the “situation” in which transaction happens in real world and present that situation in the form of an image instead of conventional way of storing it in digits/numerals. The primary reason for proposing an image encoding is to normalise the data by removing the heterogeneity. We have created a process to convert the data points into an image by referring a standard image/icon library or applying proprietary QR conversion algorithm on the data points. After the data is encoded in the form of an image, we use image processing algorithms to train the machine and create a model to score a future transaction. About the speaker: Amit Godbole is a Jnana Prabodhini alumnus and did his graduation in Mechanical Engineering from Vishwakarma Institute of technology, Pune and started his career in the cards and payments domain. He considers himself lucky to have got stuck with the same domain for about 17 years now and has worked with some of the worlds famous technology and transaction processing companies like Mastercard, First Data(now Fiserv), Capgemini, HP. He has led product roadmaps as a Director in Mastercard and that’s where he kept wondering why is this transaction data so cryptic and how can it be processed in a faster, easier way to derive insights out of it. These insights will help financial institutions in various ways like mapping consumer persona, predicting cardholder behaviour and ultimately leading to offering best services to them and at the same time predict and control fraud. He started on his pursuit recently and is here to present some challenges he is tackling to solve the problem.
|
