By David Eschwé, Vjekoslav Bonic and Robert Barcik | Advanced Analytics & Artificial Intelligence Tribe at RBI
The technical developments over the last years have fueled tremendous Advanced Analytics activities around the world. Every day new sophisticated approaches are developed. Within the Advanced Analytics & Artificial Intelligence Tribe we are driven by the opportunity to continuously learn something new. However, these methods are merely tools for solving real life problems, e.g. making cars drive autonomously, finding new vaccines for Covid-19 or increasing the gross income for a certain product. Read more about feeding modelling results back to reality, detecting meaningful patterns and the top trends in Advanced Analytics.
Advanced Analytics (AA) is not a narrowly defined area. Instead, it is a set of disciplines like statistics, machine learning, deep learning, or artificial intelligence. These are used for a specific purpose, namely turning data into information, i.e., making predictions or generating recommendations. For success, it requires not only an experienced data scientist, but we also need an expert from the respective business domain. This required close collaboration with so many different business areas makes our job extremely interesting. There is virtually no area where we couldn’t apply AA. In banking business, this can be Retail business (optimizing and personalizing CRM systems, Transactional Scorecards, Chatbots), Corporate business (optimizing CRM and Early Warning System), Capital Markets (Optimizing the FX Margins and recommending Bonds), Compliance (Anti Money Laundry Systems), Controlling (Data Anomaly Detection) and Operations (making the cash in transit process more efficient). Although it’s a vast and trending topic, it’s not entirely new.
Old methods for new insights
Some of the methods used in Advanced Analytics are old. For example, the first step towards neural networks took place in 1943, when Warren McCulloch, a neurophysiologist, and a young mathematician, Walter Pitts, wrote a paper on how neurons might work. They modeled a simple neural network with electrical circuits. The main reason why AA became so popular over the last few years is the fast development of computational power. In today’s cloud environments provided by Amazon, Google, or Microsoft, you can configure mighty machines to estimate complex neural networks in a reasonable time. Let’s look at an easy example.
Detecting meaningful patterns
You might know this story about a pregnant woman who didn’t know she was expecting a child, but her drugstore already knew by comparing her shopping habits to those of other women. This is an example of supervised learning, which we also apply a lot in RBI Group to boost our gross income for certain products. If you had a formula (classification) indicating which of your customers will accept a product offer in the future, you can optimize the reward from your future campaigns. What you do is the following. From previous campaigns you know which customers accepted an offer and which customers did not. Connecting this data with other data sources (e.g. transactional behavior, socio demographic information, credit bureau, etc.) you can then derive the typical patterns behind customers that have accepted an offer. These typical patterns allow us choosing appropriate customers for the next campaign. However, not every detected pattern is meaningful, and it requires a lot of data science training and a lot of business know-how to identify relevant patterns and relationships.
Wrong conclusions or do storks really bring the babies?
When dealing with big data, you need to distinguish between correlation and causality carefully. Very often, these two terms are confused, thus, leading to wrong conclusions.
Imagine the following example. Strange phenomena occurred in the 1960s, in some smaller cities in Germany. People observed that many babies were born at the same time when a larger number of storks started to occur in these cities. So, do Storks bring the babies? Of course not. Researchers found that a lot of young couples were moving into the cities. These young couples upon settling down build houses and eventually decided to have a baby. The houses they build for themselves provided the storks the ideal environment to build their nests. The observed pattern represented a spurious correlation.
Even though this example seems obvious, it is a constant struggle to avoid similar silly conclusions in our daily work. When feeding modeling results back to reality we need to stay curious but alert when embracing the opportunities AA is providing.
AA top trends becoming part of business practice across industries
Currently, Natural Language Processing (NLP) is one of the top trends in the AA community. Pre-trained language models have made practical applications of NLP significantly cheaper, faster, and more manageable. These models are at first pre-trained on a vast dataset (e.g., learn English from the entire Wikipedia). We can then take this already trained model and quickly fine-tune it to adapt to other NLP tasks. For example, to identify the owner(s) of a real estate property from the cadaster registry. This approach was already successfully implemented in RBI Group and allowed for many applications. These are ranging from classifying documents like emails to speed up back-office processes to extracting structured information from long reports like the cadaster registry or a collateral evaluation report.
Another big trend is conversational Artificial Intelligence (AI), which is becoming an integral part of business practice across industries. More companies are adopting the advantages chatbots bring to customer service, sales, and marketing. In RBI Group, we are working on developing and improving chatbots for various applications. In Austria, we recently launched cHRis, a chatbot answering HR questions. In Ukraine, we developed a chatbot system that can issue digital debit cards. In many more countries, we have chatbots in place supporting our customers. Although chatbots are becoming a “must-have” asset for leading businesses, their performance is still very far from human. Therefore, much research is invested in conversational AI and semantic search.
Finally, we should mention Reinforcement Learning (RL) as a trend. Reinforcement Learning is one of the three basic machine learning paradigms, alongside supervised learning (e.g., logistic regression) and unsupervised learning (e.g., cluster analysis). Many experts recognize Reinforcement learning as a promising path towards Artificial General Intelligence (AGI), or true intelligence. For example, this area brings so-called agent-based learning, which could have many exciting applications for a bank. We could simulate how individuals would behave within an economic system.
Besides the top trends, there is one dominating topic the AA community is watching very closely now: the Covid-19 pandemic.
Crisis emphasizes need for more real-time data
Predictive modeling is difficult now, because of the structural break in the data. We also anticipate a sustainable change in our customers’ behavior, requiring potential modifications of our models and approaches in place. But in this context, the adaptive and agile nature of AA approaches comes in handy. We have short development cycles of ‘implement, measure, adjust’, and the feedback-loops implemented in many machine learning systems. Finally, many AA applications are not affected at all, e.g., think of NLP models, chatbots, or real-time signal detection. Summarizing the current crisis emphasizes the need for more real-time data and requires a more conscious approach to modeling in general. Yet AA continues to bring value to the business.
Why morally relevant questions are appreciated
Over the last years, the advances of AA & AI have triggered many activities on the side of lawmakers. For example, the European Commission recently presented its Digital & Data Strategy comprising among others a European Strategy for Data and a Regulatory Framework for Artificial Intelligence. We see similar activities in our core markets in which we face partially different regulatory opinions regarding data protection, cloud computing, IT security, and AA.
At first sight, this regulatory attention seems like bad news. It might seemingly strip away the freedom you had when developing AA models. On the contrary, we appreciate those activities very much. The growing capabilities of AA & AI raise a lot of morally relevant questions, not speaking of the unknown effects on our societies in general. In this sense, it is good that lawmakers require from us reliable and transparent handling of data (GDPR). Furthermore, we also find it very reasonable to require that the future AA and AI systems always operate under human oversight. These systems should also be robust, safe, transparent, and fair. In this way we can justify the trust our customers are putting into us.