Chicago Machine Learning Meetup

Rob Lancaster will be speaking on "Survival Analysis for Cache Time-to-Live Optimization."

Summary:

We examine the effectiveness of a statistical technique known as survival analysis to optimize the cache time-to-live for hotel rates in a hotel rate cache. We describe how we collect and prepare nearly a billion records per day utilizing MongoDB and Hadoop. Finally, we show how this analysis is improving the operation of our hotel rate cache.

Abstract:

The Orbitz family of travel sites receives hundreds of thousands of searches each day from consumers looking for hotels. A single consumer initiated request may spawn dozens of individual hotel rate look-ups resulting in many millions of such requests per day. Most hotel inventory is managed by suppliers, often in more antiquated systems not capable of handling a large number of requests. In order to minimize the impact of high consumer traffic volume on these suppliers, Orbitz caches rate information locally. Such caching also helps Orbitz maintain look-to-book ratios with their suppliers and reduces latency experienced by the consumer.

Hotel rate and availability information can change over time causing cached information to go stale. This is not desired since it causes consumers to experience discrepancies between cached and real-time rates. Therefore, each piece of information stored in the cache is given a time-to-live (TTL). Historically, the TTL values have been determined by business intuition and have not been generated via a data-driven approach.

Here we investigate the applicability of predictive modelling to optimize TTL values for rates in our hotel rate cache. Specifically, we examine survival analysis as a means of modelling hotel rate volatility. Survival analysis is a statistical technique which models the time until the occurrence of a particular event. In the context of biological organisms, the event of interest is often the death of the organism (hence the name). In our context, the event of interest is a change in the rate offered by a hotel or in its availability status.

We highlight some of the technical challenges in collecting nearly a billion records each day. This includes how we use MongoDB as a collection mechanism for real-time events emitted by our hotel applications before being transferred to Hadoop for long term storage and processing. We’ll also cover how the data is prepared in Hadoop prior to being made available for use in building and evaluating our predictive models.

Finally, we show how our results have both challenged some and confirmed other of our long-held assumptions on rate volatility and how we’re using these results to improve our look-to-book and cache hit ratios while reducing rate discrepancies experienced by consumers.