Data Analytics Convergence: Business Intelligence(BI) Meets Machine Learning (ML)

Headquartered in London, England, BP (NYSE: BP) is a multinational oil and gas company. Operating since 1909, the organization offers its customers with fuel for transportation, energy for heat and light, lubricants to keep engines moving, and the petrochemicals products.

Business intelligence has always been a key enabler for improving decision making processes in large enterprises from early days of spreadsheet software to building enterprise data warehouses for housing large sets of enterprise data and to more recent developments of mining those datasets to unearth hidden relationships. One underlying theme throughout this evolution has been the delegation of crucial task of finding out the remarkable relationships between various objects of interest to human beings.

What BI technology has been doing, in other words, is to make it possible (and often easy too) to find the needle in the proverbial haystack if you somehow know in which sectors of the barn it is likely to be. It is a validatory as opposed to a predictory technology.

When the amount of data is huge in terms of variety, amount, and dimensionality (a.k.a. Big Data)  and/or the relationship between datasets are beyond first-order linear relationships amicable to human intuition, the above strategy of relying solely on humans to make essential thinking about the datasets and utilizing machines only for crucial but dumb data infrastructure tasks becomes totally inadequate. The remedy to the problem follows directly from our characterization of it: finding ways to utilize the machines beyond menial tasks and offloading some or most of cognitive work from humans to the machines.

Does this mean all the technology and associated practices developed over the decades in BI space are not useful anymore in Big Data age? Not at all. On the contrary, they are more useful than ever: whereas in the past humans were in the driving seat and controlling the demand for the use of the datasets acquired and curated diligently, we have now machines taking up that important role and hence unleashing manifold different ways of using the data and finding out obscure, non-intuitive relationships that allude humans. Moreover, machines can bring unprecedented speed and processing scalability to the game that would be either prohibitively expensive or outright impossible to do with human workforce.

Companies have to realize both the enormous potential of using new automated, predictive analytics technologies such as machine learning and how to successfully incorporate and utilize those advanced technologies into the data analysis and processing fabric of their existing infrastructure. It is this marrying of relatively old, stable technologies of data mining, data warehousing, enterprise data models, etc. with the new automated predictive technologies that has the huge potential to unleash the benefits so often being hyped by the vested interests of new tools and applications as the answer to all data analytical problems.

To see this in the context of predictive analytics, let's consider the machine learning(ML) technology. The easiest way to understand machine learning would be to look at the simplest ML algorithm: linear regression. ML technology will build on basic interpolation idea of the regression and extend it using sophisticated mathematical techniques that would not necessarily be obvious to the causal users. For example, some ML algorithms would extend linear regression approach to model non-linear (i.e. higher order) relationships between dependent and independent variables in the dataset via clever mathematical transformations (a.k.a kernel methods) that will express those non-linear relationship in a linear form and hence suitable to be run through a linear algorithm.

Be it a simple linear algorithm or its more sophisticated kernel methods variation, ML algorithms will not have any context on the data they process. This is both a strength and weakness at the same time. Strength because the same algorithms could process a variety of different kinds of data, allowing us to leverage all the work gone through the development of those algorithms in different business contexts, weakness because since the algorithms lack any contextual understanding of the data, perennial computer science truth of garbage in, garbage out manifests itself unceremoniously here : ML models have to be fed "right" kind of data to draw out correct insights that explain the inner relationships in the data being processed.

ML technology provides an impressive set of sophisticated data analysis and modelling algorithms that could find out very intricate relationships among the datasets they process. It provides not only very sophisticated, advanced data analysis and modeling methods but also the ability to use these methods in an automated, hence massively distributed and scalable ways. Its Achilles' heel however is its heavy dependence on the data it is being fed with. Best analytic methods would be useless, as far as drawing out useful insights from them are concerned, if they are applied on the wrong kind of data. More seriously, the use of advanced analytical technology could give a false sense of confidence to their users over the analysis results those methods produce, making the whole undertaking not just useless but actually dangerous.

We can address the fundamental weakness of ML technology by deploying its advanced, raw algorithmic processing capabilities in conjunction with the existing data analytics technology whereby contextual data relationships and key domain knowledge coming from existing BI estate (data mining efforts, data warehouses, enterprise data models, business rules, etc.) are used to feed ML analytics pipeline. This approach will combine superior algorithmic processing capabilities of the new ML technology with the enterprise knowledge accumulated through BI efforts and will allow companies build on their existing data analytics investments  while transitioning to use incoming advanced technologies. This, I believe, is effectively a win-win situation and will be key to the success of any company involved in data analytics efforts.