I sometimes say that business analytics is about making decisions – better. With that I’m trying to convey a vision of business analytics which is broader than just data and computation. It’s about helping organisations to achieve several goals: to understand what has happened, to predict what will happen, and to make the best possible decision now.
So while data and computation are at the core, so are the softer-edged concepts of decision analytics and the less data-oriented, more algorithm methods of optimisation. And business analytics is fundamentally about translation – translating a business problem into the language of algorithms and analysis, and translating a solution back into the language of clients and management.
In this short article I will expand on this description of analytics, and then approach it from another angle by analysing the types of projects undertaken by the students of the UCD Smurfit MSc in Business Analytics.
Viewing analytics as a ladder
It’s common to describe a kind of “ladder” of analytics, progressing from descriptive through predictive to prescriptive analytics.
A well-known example of the ladder model is Gartner’s “Analytic Value Escalator” which is reproduced below. It adds a category, diagnostic analytics, which in other models is merged with descriptive analytics. It suggests a progression up the ladder in terms both of sophistication and of potential value to the business.
(Image source: flickr)
To a first approximation, descriptive analytics means statistics, visualisation, and unsupervised learning; predictive means supervised learning including linear regression, logistic regression, support vector machines, decision trees and forests; and prescriptive analytics means optimisation and decision analytics.
Where do the current buzzwords of big data and deep learning fit into this? Both are at a perhaps surprisingly “low” rung of the ladder. Very few businesses actually deal with big data, and of those that do, the majority only need descriptive analytics on it. Deep learning is a class of methods that usually sit on the predictive rung of the ladder, along with plain old linear regression – a bit more sophisticated, and more powerful, but essentially answering the same types of questions.
For most of these methods, some mathematical understanding is required to avoid making dangerous mistakes, though visualisation and machine learning methods are becoming increasingly commoditized. But for all types of analytics, a much broader array of skills are really needed:
- business insight and domain knowledge;
- communication with clients, management, and domain specialists;
- presentation, visualisation, and written reports;
- understanding of real-world decision-making processes;
- computing chops to munge data and glue algorithms together – even if you never write a line of optimisation or statistical code.
This spectrum from basic to advanced analytics has been proposed to explain the analytics journey that organisations go through, and sometimes that is what happens in practice. However I think a more typical journey, for many organisation and individuals, may start with simple methods being applied at any rung of the ladder, followed by increased sophistication. For example, it is often possible to establish a simple automated classification system based only on domain knowledge and good data quality – no learning models required. This can be used as a baseline and a sanity check against which to measure more sophisticated machine learning methods as the organisation gains the required expertise.
The above could be called an intensional definition of business analytics. But an extensional definition might be useful too: what do we talk about, when we talk about business analytics? In particular, what do analytics organisations talk about, when they come into our classrooms in UCD Smurfit?
To address this question, I have taken a data set consisting of the projects proposed by analytics organisations for collaboration with the students of the UCD MSc in Business Analytics over the past 6 years. I’ve labelled it by industrial sector, by business function, and by analytics method(s) used.
I’ve found that almost all sectors of industry are represented, from obvious candidates like finance through to less obvious ones including sports, agriculture and forestry, and the public sector, as shown below:
Moreover, these analytics projects are being used in pretty much all business functions – from core functions and operations through to sales, CRM, and HR.
Concerning the methods being used, as shown below, the obvious result is that – perhaps predictably – predictive analytics in the form of supervised learning is very popular.
But beyond these a very broad range of methods have been required. Unsupervised learning is popular too. Natural language processing (NLP) and time-series methods are often required either as stand-alone methods or as inputs to other algorithms. Optimisation, including both “classical” and metaheuristic optimisation remain popular, reflecting a broad interpretation of analytics and also reflecting its roots in the field of operations research.
Multi-criteria decision-making (MCDM) is the term I’m using for the field of decision analytics which is not primarily about data and computation, but rather about understanding and improving the decision-making process – based on evidence. Again, this reflects a broader set of concerns than just “data for its own sake”.
Our partners also often suggest more niche methods including spatial analytics, recommender systems, network models, and simulation methods.
We can draw a few conclusions and speculate about the future. Over time, we have noticed an increased interest in extracting value from text, spatial, network and other forms of non-tabular data. This reflects business intuition that value can be extracted from such data, and a recognition that basic tools and approaches – say, linear regression in Excel – are a bad fit.
The dominance of supervised learning suggests that there are opportunities for deep learning and other advanced algorithms which can go beyond the performance of standard regression and classification techniques.
However, “unsupervised learning is believed to be a key for future progress on deep learning towards AI” (Bengio). We believe that unsupervised learning will tend to grow in importance, in comparison with supervised learning. This reflects the increased power of algorithms for unsupervised learning – probably the biggest achievement of deep learning so far – but in particular it reflects the wide availability of large datasets which are suitable for unsupervised learning but not for supervised. Businesses have this data and want to understand and extract value from it.
Although less hyped at the moment, there is continued interest in “the science of better“, that is operations research and optimisation. These are often targeted more at shaving 1% off (say) transport costs, rather than at achieving understanding or prediction. But that 1% can be very meaningful.
Finally, we are seeing an ever-broadening array of applications, and we expect to see our students using analytics in more and more unexpected applications every year.
Next Generation the founders of the DatSci Awards are delighted to partner with UCD Michael Smurfit Graduate Business School to offer 2017/18 applicants (full- & part-time) for the MSc Business Analytics the chance to win a scholarship to cover full-time EU Fees (value in 2017: EUR 13,350*). Successful applicants must meet the entry requirements. Early bird tickets are available until the 31st of May 2017. Learn more