EyeforTravel Europe 2018

June 2018, London

Why Airbnb believes machine learning needs ‘thinking’

New technologies like machine learning are considered to be the next generation in travel pricing, but it doesn’t have to be a black box. Tom Bacon shares the insights he gleaned in Miami

Machine learning is necessary to create and maintain more personalised pricing in a complex, constantly changing marketplace. Revenue management (RM), which is already highly sophisticated, fundamentally based on millions of historic data points and designed to be responsive to market and competitive dynamics, becomes exponentially more complex when modified for ‘personalisation’. Adapting current processes to manage forecasting and optimisation across hundreds of more individualised segments of demand is inconceivable. Against this backdrop, travel suppliers, as delegates at the EyeforTravel Smart Analytics conference in Miami heard recently, need machine learning to manage the larger data and then calculate the potential dizzying number of price changes.

However, many analysts refer to machine learning as a ‘black box’. And, in fact, some merchandising experts actually see the complexity of ML as an advantage. ML can identify unusual, hidden, unintuitive, non-linear relationships that, although potentially incomprehensible to analysts, are more ‘optimal’ than any RM system has ever been able to do before!

But, is ‘incomprehensible’ a good thing? Actually, many experts in the travel industry vehemently disagree; they insist upon ‘explainable’ ML.

In Miami, in fact, data scientists from some large travel suppliers argued that ML can and should be carefully managed. Rather than either-or (either ML, or comprehensible), they defended managing ML in a way that enhances, rather than defies, comprehension.

Voice of intent

Theresa Johnson of Airbnb has a data-intensive background in sensor design, and signal processing for both ground based hypervelocity impact tests, and in-situ satellite measurements of plasma/spacecraft surface interactions. Incomprehenisble? Perhaps! But as a data scientist for Airbnb, Johnson is leveraging her experience with ML, and is a strong advocate for active management of it. In other words, her approach is to first clearly and narrowly define the business problem; ‘thinking’, she argues, is the first step in effective ML management. 

ML should be used to enhance business understanding, not to replace it

Johnson advocates that significant upfront work is needed to predict relationships based on business insight and to capture causative factors. Business problems should be broken down into essential parts and then ML assigned an explicit role, rather than be used as a way to sort through a kitchen-sink of data to find a relationship incomprehensible to analysts. According to Johnson, ML should be used to enhance business understanding, not to replace it. “Data is the voice of the customer,” she says, and Airbnb is set up to listen to that intent.

Meanwhile, Nuno Antonio, chief technology officer of ITBASE, also spoke of his work managing ML, and referenced a tool for managing it by XGBoost. In one recent case, Antonio said that he could have let ML find a stronger and more complex, but potentially non-transparent, relationship among causative factors (a more ‘optimal’ model as measured by a higher R-squared). Instead, he opted for increased understanding instead. ITBASE used ML to produce decision trees of the explanatory factors that explicitly laid out the relationship of the drivers to the outcome. For Antonio then, ML produced not just the business model but helped display the entire process by which it produced the model outcome. Indeed, machine learning developed a totally transparent box!

None of the companies in Miami (and for the record, virtually all were involved with cutting-edge data science in travel) accepted deployment of ML as a black box. Instead each insisted on ‘explainable’ ML.

Tom Bacon has been in the business 25 years, as an airline veteran and now industry consultant in revenue optimisation. He leads audit teams for airline commercial activities including revenue management, scheduling and fleet planning. Questions? Email Tom or visit his website

Download the Miami Data Summit Round-up report for free here to get all the insights from the event. 

Missed Miami? Why not join us one of our upcoming events

Related Reads

comments powered by Disqus