The pandemic has radically altered online shopping behavior, throwing a wrench into many AI systems.
What’s new: AI inventory trackers, recommendation algorithms, and fraud detection systems trained on pre-pandemic consumer behavior have been flummoxed by the wildly different ways people now browse, binge, and buy, according to MIT Technology Review.
What’s happening: Companies are scrambling to retrain machine learning systems for the new normal.
- Amazon’s recommender typically promotes items the company itself can ship. With its distribution network under strain, the algorithm seems to be promoting products from sellers who handle their own shipments.
- Featurespace, which provides fraud detection technology for financial and insurance companies, revamped its behavior models to account for surges in demand for things like power tools and gardening supplies. Such spikes used to trigger alerts. Now they’re business as usual.
- AI consulting firm Pactera Edge says an upswing in bulk orders broke a client’s predictive inventory systems. Another client found its public-sentiment analysis software distorted by all the gloomy news.
- Phrasee, a company that generates ad copy using natural language processing, tweaked its algorithm to avoid phrases that could spark panic (“going viral”), raise anxiety (“stock up!”), or promote risky behavior (“party wear”).
Behind the news: E-commerce is one of the pandemic’s few beneficiaries: Growth in online sales in April tripled over the same month last year, according to an analysis by electronic payments processor ACI Worldwide.
Why it matters: Beyond its terrible human toll, Covid-19 is making yesterday’s data obsolete. The AI community must find ways to build resilient systems that can adjust as conditions change.
We’re thinking: In our letter of April 29, we pointed out that AI often suffers from a gap between proofs of concept and practical deployments because machine learning systems aren’t good at generalizing when the underlying data distribution changes. Covid-19 is bringing about such changes on a grand scale, and our systems are showing their brittleness. The AI community needs better tools, processes, and frameworks for dealing with this issue.