T1A

ML Trends of 2020

№1 - AutoML

In brief, AutoML paradigm is here to empower any user with  “make one-click to get a good model” functionality, letting business and data scientists focus on what really matters - business cases, experimentation and further optimization of business processes.

While it is commonly positioned as a tool for business analysts, making data science available for people without tech background, it still demands for a good understanding of data and data science process best practices.

Since the market still has huge unfulfilled demand for data-related specialists, the popularity of such tools will certainly grow stronger in the coming years.

For now it is important to remember that it is still too early to claim that such approach will displace traditional ml-tools, especially in larger enterprises, where customization and even fractions of additional accuracy is the key. But it certainly helps to set a benchmark for a new model, retrain existing one and go through a hypothesis confirmation process.

№2 — Explainable AI (XAI)

It is important for a business user to understand the logic behind the decision. This is especially true for areas where easily interpretable classification models like decision trees or logistic regression (e.g. credit risk, target marketing, insurance) have been a preferable choice for decades. It is obvious that using, say, a set of xgboosts will most likely give better results, but business and regulatory processes dictate their own constraints, as well as desire to find hidden and non-obvious relationships in data.

Over the last years this gap between performance and transparency has been closed with various methods like LIME, XSHAPE and others, and looking at the results of the recent research we are more than sure that this methods will spread wide.

№3 — Reinforcement Learning

Reinforcement learning, originally used to teach computer play games (and well known for such SOTAs as AphaGo) , now finds numerous applications in business problems where optimal reasoning has to be paired with optimal exploration of possible decisions. You might say it is like continuous A\B testing on steroids providing you prescriptions for the optimal outcome.

Analysts are coming up with more and more applications for enterprises, optimizing such processes as: Next Best Action in direct marketing, personalization of web pages and delivered content, recommender systems, debt collection and promo optimization.

№4 — Graph Analytics

Graph Analytics - set of methods and methodologies to analyse structure of links and relations between various entities. As an example of such graphs - social network graphs from social media data, links of banking accounts based on transactions between them, structure of ownership for a large group of companies.

In machine learning these approaches are commonly used for identification of new strong predictors(features), that describe “neighbors” and relations of the entity. For example we could answer to “How does counterparties’ credit rating affects credit rating of analysed company, how is it different if we look at the whole supply chain?”

Leveraging graph analytics we can rely not only on direct links between entities but also second and third and n’th order links if needed.

Today, graphs are successfully used for analysis of entities with “natural” net-like structure. In future we will use it more extensively for more unobvious scenarios like analysis of chain of communications and interactions between customer and an enterprise or root-cause analysis for direct marketing management

№5 — MLOps

Enterprises' involvement in AI and ML rises every year. Leaders are actively investing in R&D, ML pilots and advanced analytics of existing data. As a result focus gradually shifts from finding successful use cases to the most efficient way to make it prom pilot to production.
It drives demand in both methodology and tools to automate processes involved in machine learning production cycle. 
Taking its roots in DevOps, MLOps (or ModelOps or AiOps) covers everything from a set of tools to framework and methodology to move models from ideation to production as quickly as possible, bridging the gaps between involved functions and their mindsets.


As of today market is fractured with more than 50 tools and solutions covering different aspects of MLOps, with both open-source (MLFlow, KubeFlow) and proprietary (SAS, IBM, …) software.
In the upcoming days we will witness how this market will grow and mature to unification of approaches and methodologies and technology stack behind.