Meet Hexaly at the 5th Annual EURO Practitioners Forum Conference
Hexaly is delighted to sponsor the new edition of EURO Practitioners’ Forum. The conference will be held from October 14 to 15, 2024, at University of Coimbra, Portugal. For more information, visit the conference website here.
Meet our team to discover the latest features and applications of Hexaly 13.0, and explore our job opportunities. To learn more about Hexaly, be sure to attend the presentation by Julien Darlay, Head of Science at Hexaly. Below is a summary of the presentation he will give on Monday, October 14, 2024, at 5:15 PM during the EURO Practitioners’ Forum.
Effective collaboration between operations research and data science in online advertising
This talk provides feedback about a successful project developed in 2022 for one of Hexaly’s clients with an important online business. It involved the revenue management team, data science experts, and an operations research team from Hexaly. The application has been used daily by multiple users since 2023. The business problem can be described as follows: suppose that you have a database of qualified users (age, gender, location, buying patterns, etc.) and the history of all visits for the past few years, and you want to estimate the number of distinct women aged 20 to 40, located in a specific region who are interested in reading books or cinema and are likely to visit your website next week. Answering this question on past visits can be done with an SQL request on a classical database. Still, it requires considerable memory and several hours of computation time because of the history size. Our client needs a precise estimate in
less than one second.
The data science team was already able to derive a precise forecast of the number of future visits for a particular period. The remaining difficulty was to estimate the proportion of visits satisfying the request’s criteria. We derived a high-speed algorithm with the data science team based on approximate counting and the HyperLogLog algorithms. It requires the offline precomputation of data structures by scanning the visit history, which can be done weekly in a few hours. These data structures allow a fast computation of the approximation of the cardinality of the union of sets and are the key elements in estimating the answers. We later generalize their usage inside standard operations research algorithms and models to solve the MAX-k-COVER and the MIN-k-UNION problems.
The success of this project was a result of the close collaboration between the revenue management team, which defined the industrial needs and explained the data, the data science team, which had all the resources to perform massive data analysis, and finally, an O.R. team, which handled the algorithmic aspects. We found a way to summarize the database using HyperLogLogs, keep this aggregate in memory, and develop a fast web service to answer requests in a few milliseconds.
If you’re interested in Hexaly, you can request a trial license here. In the meantime, feel free to contact us; we will be glad to exchange your optimization problems. Additionally, Hexaly is free for faculty and students, so feel free to request your academic license!