Devs and Ops don't mix? New data analytics tool could fix that
Descriptive analytics -- looking into what has happened in the past -- is commoditized. Traditional BI, reports, and dashboards are well understood and used in one way or another across a wide spectrum of organizations to keep an eye on their operations.
Predictive analytics -- trying to predict what will happen in the future -- is in the limelight. Today many organizations are scrambling to utilize predictive models leveraging data and Machine Learning (ML) in the hope that this will empower them to anticipate future developments.
Prescriptive analytics is different. Neither what it is nor how it's supposed to work are entirely clear. So when faced with questions, would turning to Google help answer them? Let's have a look at YouTube, the most massive and widely used media channel on the planet, as a case study to examine prescriptive analytics.
Everyone, meet prescriptive analytics
Gartner's definition of prescriptive analytics is "a form of advanced analytics which examines data or content to answer the question "What should be done?" or "What can we do to make _______ happen?", and is characterized by techniques such as graph analysis, simulation, complex event processing, neural networks, recommendation engines, heuristics, and machine learning."
Cryptic much? Lisa Kart, Gartner analyst, offered a more tangible approach by stating that "Prescriptive analytics is about applying logic and mathematics to data, with the goal to specify a preferred course of action- unlike other types of analytics the output is a decision."
So the result of applying prescriptive analytics is not information about the past, not predictions about the future, but decisions. But establishing this as the definition of prescriptive analytics is just the onset for more questions.
To begin with, while the definition of predictive analytics does not really expand on the technical approaches of applying it, the definition of predictive analytics goes to some lengths to list an array of potential techniques.
Why the difference? And does using any of the listed techniques in and of itself qualify an analytics solution as "prescriptive analytics"?
One interpretation for the emphasis on technologies in Gartner's definition is that unlike predictive analytics, for which using ML seems to be a given, prescriptive analytics is fluid, less anchored, and perhaps less understood at this point.
At any rate, solutions using any of the technologies listed should not be awarded the "prescriptive" moniker based on this criterion alone. There are other questions to ask that can perhaps serve better as criteria for classification.
Bottomless bowls, infinite feeds, and autoplay
Other questions such as, which decision exactly are we talking about here? What are the options to choose from, and who makes the decision? For sure, being able to predict the future helps with decisions. But doesn't a decision imply action in the future anyway? So where exactly are the boundaries between predictive and prescriptive?
When faced with overwhelming questions, people often turn to Google. This could possibly help in this case too, although not in the way Google is often used. Google is -- among other things -- the proud owner of YouTube. An iconic example of a data-powered product by a data-powered organization.
The product, in typical Google fashion, is advertisements paid for by third parties and targeting consumers -- users viewing YouTube content. Every YouTube video is viewed alongside advertisements, which consumers occasionally click, bringing advertisement revenue to Google.
This business model has some interesting implications. It implies that the more videos users will watch, the more advertisements they will be shown, and the higher the chances some of them will be clicked. Therefore, the more videos users consume, the more revenue for Google.
So answers to the question "What can we do to make people watch more videos" are of great interest for Google. Apparently "show them more videos they will like" ranked high among other options.
Tristan Harris, ex-design ethicist at Google, elaborated on design hacks used to this effect, including what he called "bottomless bowls, infinite feeds, and autoplay." But at the end of the day, if the content is not relevant, how much good can such hacks do? Case in point -- there has been a fair amount of criticism on the relevance of YouTube recommended videos.
Google, meet the analytics continuum
Google responded by revamping YouTube's recommendation engine. Previously, recommended videos on YouTube have been served based on the same principles driving Amazon's recommendation engine. The architecture and techniques utilized in YouTube's recommendation engine have been outlined in a paper recently published by Google engineers.
Google uses two layers of deep neural networks to narrow down the immense stock of YouTube videos to a ranked list of recommendations. Recommendation generation is modeled as (extreme) multiclass classification, producing an initial pool of recommendations subsequently ranked using logistic regression.
So how would that be classified then? Definitely not descriptive analytics, but is it predictive or prescriptive? Depending on how one looks at it, it might as well be either.
As per another definition, "the prescriptive approach analyzes potential decisions, the interactions between decisions, the influences that bear upon these decisions and the bearing all of the above has on an outcome to ultimately prescribe an optimal course of action in real time."
One cited example is airline ticket pricing systems that use prescriptive analytics to sort through complex combinations of travel factors, demand levels, and purchase timing to present potential passengers with prices designed to optimize profits but also not deter sales
One could argue that this is about making predictions about what users will like, so it's predictive. One could also argue that this is about helping Google answer the "What can we do to make people watch more videos" question, and sorting through complex combinations of video factors, so it's prescriptive.
One could probably argue over definitions for a long time, which goes to show that using he descriptive-predictive-prescriptive classification can be hard. Not only does it look and feel more like a continuum than a set of discrete classes, but the progression in the continuum is not entirely straightforward either.
Great -- but does it work?
So how well is the new recommendation engine performing? Opinions vary apparently. Striving for improvement, Google has explicitly sought after user feedback. Spontaneous, public user feedback however appears to be mostly negative at this time.
On the other hand, there have also been some positive remarks, specifically about the next to watch feature. As this presumably is the top-ranked result of the engine, maybe there is light at the end of the tunnel.
Anecdotal evidence from people with tastes in music ranging from eclectic to mainstream also points towards a sometimes surprising aptitude to auto-compile engaging playlists as of late. So the jury on what this is and whether it works is still out.
However there are still many aspects regarding the intricacies and alternatives in implementing YouTube's call-them-what-you-will analytics that are worth exploring, and we will continue to do so on our tour in the land of the analytics continuum.
Why building IoT solutions remains a challenge for developers: