This week at the Association for Computing Machinery’s (virtual) Knowledge Discovery and Data Mining conference (KDD), a team of researchers from Amazon Web Services learned that they had won the best-paper award for the conference’s Applied Data Science track.
The award-winning paper is “Temporal-contextual recommendation in real-time” by machine learning scientist Yifei Ma, senior machine learning scientist Murali Narayanaswamy, applied scientist Haibin Lin, and applied scientist Hao Ding.
The technology described in the paper is one of the core algorithms underlying the Amazon Personalize service used by StockX, Domino’s, and PulseLive, among other companies.
“Personalized recommendation helps with the delivery of lesser-known materials in the long tails,” Ma says. “This paper explores the challenges in building a black-box recommender system that self-adapts to any types of datasets without the supervision of machine learning experts.”
Recommender systems typically base their recommendations on two factors: similarities between customers (based on purchase histories) and similarities between products. But this approach can miss shifts in preference that occur over time — as when, for instance, a new product is released in a particular product category.
The research reported in the paper enables recommender systems to use information about both the order in which purchases occur and the time lapses between them. By providing a theoretical foundation for assigning different weights to such temporal factors and more time-invariant factors — such as general connections between different product categories — the paper also presents a principled approach to solving the “cold-start” problem, or how to provide recommendations to customers with sparse purchase histories.
The paper also provides sampling techniques that allow this approach to generalize accurately and efficiently to product catalogues with millions of entries.
