[ad_1]
Two of the three general chairs at this year’s ACM Conference on Web Search and Data Mining (WSDM) are scientists from Alexa Shopping’s offices in Haifa, Israel: Liane Lewin-Eytan, senior manager of applied research, and David Carmel, a principal applied scientist.
When they volunteered for the role, Lewin-Eytan and Carmel were hoping to bring the conference to Jerusalem. “It was the beginning of March [2020],” Lewin-Eytan says. “We had a beautiful bid, with many, many attractions in Jerusalem, because the steering committee of the conference takes also into consideration what the country can offer to participants. We were very excited to bring it here — it was a big honor to the local community — but because of COVID-19, we ended up online. We still hope to host the conference here one day.”
Despite their disappointment, being general chairs gives Lewin-Eytan and Carmel a privileged perspective on the conference as a whole — on the full gamut of presentations and talks and what they say about the field of Web-based search.
“This is quite a young conference,” Carmel says. “It began only in 2008. I think from the first day I was there. The popular conference at that time, WWW [now known as the Web Conference], and all the other conferences that dealt with web data did not give much attention to search. So the idea was to do a kind of spinoff. And the success was tremendous. In a very short time, the conference took a very strong position in the community, and it became one of the prestigious conferences in this area.”
Having been a regular attendee at the conference from the outset, Carmel notes a few novel trends in this year’s paper submissions.
“One of them is an increased interest in e-commerce search,” Carmel says. “Product search has been around for a while in the Web community, but with more and more large companies researching this area, we see more innovative work being submitted and published at WSDM. Topics include not only searching for products to buy but also asking questions about them, discussing their benefits and attributes, which opens the way for new applications and new research.
“Another emerging topic these last few years has been health, which has become even more timely during COVID. As soon as health data, both structured and unstructured, were made available online, it became relevant to WSDM.”
“When a given domain starts to put data online and open it to users, domain-specific web-mining techniques are invented and new applications and services are launched,” Lewin-Eytan adds. “This mix of a given domain — like health, commerce, or tourism — and the Web fosters creativity and drives innovation for that domain”.
Gauging customer satisfaction
For Lewin-Eytan and Carmel, the emphasis on e-commerce is of particular interest, since that’s the problem they’re working on at Alexa Shopping. As Lewin-Eytan explains, the standard approach to search, which ranks results according to the click-through behaviors of people who were presented those results in the past, doesn’t translate easily to product search.
“People take different actions over products,” Lewin-Eytan says. “For instance, you can add a product to a shopping cart, or order it — conduct an actual transaction. While in the web you have a large number of clicks, because clicking doesn’t cost you anything, when you have to take out your credit card and buy something, it’s a totally different experience, and you don’t do that as often. So a transaction cannot be the only signal of satisfaction. If you didn’t do anything with the search results, it doesn’t mean that the search engine didn’t provide relevant results.”
The difficulty of gauging customer satisfaction is even more acute with voice-based queries. When a customer sees a list of on-screen options, the decision not to select a product is as informative as the decision to select one.
With voice-based product search, by contrast, “you have just one chance to be successful, the first one, as the customer typically doesn’t listen to more than one result,” Lewin-Eytan says. “So the problem is much harder. What also makes it harder is the data you can collect is only on what the customer saw, which is one result. So we have much, much less data to optimize for.”
One way that Lewin-Eytan, Carmel, and their colleagues are tackling this problem is by training machine learning models to predict how customers would have responded to the products they didn’t have the chance to see.
“It’s all about satisfying customers,” Lewin-Eytan explains. “That’s the hard task — both defining satisfaction, because it’s based on a number of subjective signals, and predicting it.”
In a similar vein, the Alexa Shopping team is working on predicting customers’ intents, which can dramatically alter the conclusions drawn from their transaction histories.
“We’re working now on understanding when the customer is being playful or when she has an actionable intent,” Lewin-Eytan says. “Customers find a lot of entertainment in these systems, especially with voice assistants. And sometimes their intentions are really hard to infer.”
That focus on customers, Lewin-Eytan and Carmel say, helps explain why they find participating in WSDM so rewarding. “What’s great about WSDM is that every paper is judged based on its value to Web users,” Carmel says. “That makes it particularly attractive to scientists from a customer-centric company like Amazon.”
[ad_2]
Source link
