Are the robots really coming? Pt. 1 (ML Introduction)

Audience: Non-technical Reading time: 5 mins

Its difficult to avoid news articles on a daily basis about artificial intelligence (AI), most likely followed by some sort of implied threat to jobs. Many of us have experienced it in the home in the form of smart speakers which use natural language processing AI to interact with voice. But just how much of an intelligent threat to our working lives is AI ? Assuming that the majority of the workforce is employed by bigger enterprise businesses, how relevant and accessible is AI to these organisations today ?

We live in an age of big data with huge volumes of structured and unstructured data generated by a multitude of devices and systems, such as web page views, mobiles, IoT devices, as well as all the servers behind the scenes. Some organisations, the so-called digital natives like Airbnb, Netflix and Uber, have completely disrupted their industry. Not only have they mastered their entire data landscape but by applying AI (aka Data Science) to their customer journey and operating model, they gain competitive advantage through deeper analytical insight for more accurate and timely business decision-making. 

Let’s have a look at some broad categories of analytics which can be derived from big data along with more widely applicable examples (aka use cases) for context.

1) Descriptive: What happened (valuable insight into the past)

Use Case: Retailer understanding relationships between customer segments, location and product purchases for supply chain and logistics management, as well as marketing.

2) Diagnostic: Why did it happen? (in depth insight into root cause)

Use Case: Omni-channel retailer understanding correlations and patterns between its its own Descriptive (1) insight combined with external sources such as weather, economic or travel. 

3) Predictive: What is likely to happen (Supports business decisions) 

Use Case: Subscription based service providers (ie Insurer or Mobile Network Operator) recognising patterns in customer dissatisfaction to prevent churn and protect revenue.

4) Prescriptive: What should we do (Recommendations can be automated)

Use Case: Banking institutions blocking a credit card as soon as anomalies are detected in order to protect the customer against fraud, as well reduce its own losses.

Descriptive and diagnostic (reporting) analytics can be uncovered with more established Big Data methodologies. However, realising the more advanced outcomes of predictive and prescriptive analytics will require the use of AI. More specifically, they require ‘traditional’ Machine Learning (ML) techniques.

Let’s have a brief look into the realm of the Data Scientist and get some basic ML terminology out of the way. An algorithm is a pre-built method or package and when trained or combined with historical datasets, a statistical model is created. Back to our advanced analytics outcomes, it is a trained ML model which can explain something happening now, or predict what might happen. Models need to be tested (as to their accuracy) and likely tweaked (through parameters), to increase the accuracy of the statistical inference (as opposed to intelligence) they actually provide. Algorithm + Data = Model.

ML algorithms are complex in nature but not new. In fact many have been around since the 1950s. They are also open source and therefore publicly available (wikipedia currently lists around 60 named ML algorithms). In addition to their more established services (ie storage, virtual servers and databases), the Cloud Service Providers (CSPs), ie Amazon Web Services, Google’s Cloud Platform and Microsoft’s Azure, have also launched ML services. These more advanced offerings and associated tooling provide access to common ML algorithms. An enterprise business willing to ingest their data onto these public clouds therefore have a much lower barrier of entry to explore the world of ML for near real-time predictive business insight. 

So we’ve had a brief introduction to ML and how it might be used by Data Scientists (with access to CSP tooling) to generate models. ML models are critically dependant however on large volumes of data. Are clearly catalogued datasets readily available to Data Scientists ? Is the data consistent and of sufficient quality to be useful? How much time must the Data Scientist spend on prepping and wrangling the data in readiness for ML? In Pt. 2, I’m going to have a look at the Enterprise challenge of managing and governing the entire data lifecycle in order to deliver business value from data science.