How the pandemic readied Alibaba’s AI for the world’s biggest shopping day

The news: While the US has been hooked on its election, China has been shopping. From November 1 to 11, the country’s top e-commerce giants, Alibaba and JD, generated $115 billion in sales as part of their annual Single’s Day shopping bonanza. Alibaba, who started the festival in 2009, accounted for $74.1 billion of those sales, a 26% increase on last year. For comparison, Amazon’s 48-hour Prime Day sales only crossed the $10-billion mark this year.

Pandemic stress test: The sheer scale of the event makes it somewhat of a logistical miracle. To pull off the feat, Alibaba and JD invest heavily in AI models and other technology infrastructure to predict shopping demand, optimize the global distribution of goods across warehouses, and streamline worldwide delivery. The systems are usually tested and refined throughout the year before being stretched to their limits during the actual event. This year, however, both companies faced a complication: accounting for changes in shopping behavior due to the pandemic.

Broken models: In the initial weeks after the coronavirus outbreak, both companies saw their AI models behaving oddly. Because the pandemic struck during the Chinese New Year, hundreds of millions of people who would have otherwise been holiday shopping were instead buying lockdown necessities. The erratic behavior made it impossible to rely on historical data. “All of our forecasts were no longer accurate,” says Andrew Huang, general manager of the domestic supply chain at Cainiao, Alibaba’s logistics division.

People were also buying things for different reasons, which was flying in the face of the platforms’ product recommendations. For example, JD’s algorithm assumed people who bought masks were sick and so recommended medicine, when it might have made more sense to recommend them hand sanitizer.

Changing tack: The breakdown of their models forced both companies to get creative. Alibaba doubled down on its short-term forecasting strategy, says Huang. Rather than project shopping patterns based on season, for example, it refined its models to factor in more immediate variables like the previous week of sales leading up to major promotional events or external data like the number of covid cases in each province. As livestreaming e-commerce (showing off products in real time and answering questions from buyers) exploded in popularity during quarantine, the company also built a new forecasting model to project what happens when popular livestream influencers market different products.

And JD retooled its algorithms to consider more external and real-time data signals, like covid case loads, news articles, and public sentiment on social media.

Unexpected boon: Adding these new data sources into their models seems to have worked. Alibaba’s new livestreaming AI model, for example, ended up playing a core role in forecasting sales after the company made livestreaming a core part of its Single’s Day strategy. For JD, its updates may have also increased overall sales. The company says it saw a 3% increase in click-through rate on its product recommendations after it rolled out its improved algorithm, a pattern that held up during Single’s Day.

Understanding context: Both companies have learned from the experience. For example, Huang says his team learned that each livestream influencer mobilizes its fan base to exhibit different purchasing behaviors, so it will continue to create bespoke prediction models for each of its top influencers. Meanwhile, JD says it has realized how much news and current events influence e-commerce patterns and will continue to tweak its product recommendation algorithm accordingly.

from MIT Technology Review https://ift.tt/3km9dQs