Better-tailored internet suggestions

Researchers develop methods to make product suggestions by e-commerce stores more precise and efficient.

By Elton Alisson

Agência FAPESP – When navigating through virtual stores, consumers are bombarded by ads for products, services and information that, the majority of the time, are vastly different from what they are seeking at that moment.

This situation may change thanks to a post-doctoral study developed at the Universidade de São Paulo’s Computer Intelligence Laboratory (LABIC) within the Mathematical Sciences Institute. Funded by FAPESP, the post-doctorate study intends to help e-commerce sites and content providers improve the efficiency of their recommendation systems.

The researchers are developing methods to allow the computer systems of virtual stores to make product, service and content recommendations based on other variables and not solely on the individual’s access or purchase history.

“The idea is for the store recommendation systems to utilize information from the user’s momentary context while surfing, such as the date, time and geographic location, rather than just previous behavior when accessing a given site, like pages visited,” explained Marcos Aurélio Domingues, the LABIC researcher who conducted the study under the supervision of Professor Solange Resende.

According to Domingues, e-commerce sites and content providers use two different levels of recommendation systems.

In the first, called a “simple or naïve” recommendation, the computer system suggests the same products, music and films that the user bought or accessed on their most recent visits.

In the second type of system, adopted by e-commerce companies such as Amazon, the recommendation system cross-references the product the user is clicking on at the time of the visit with products accessed or bought by people who visited the same pages as the user in the last few days or hours to suggest a product that the user has not seen and may be interested in.

The two systems, however, still do not incorporate the momentary context, such as the day of the week, the time and the user’s location. These are factors that influence the type of recommendation that the user should receive, notes Domingues.

“If the user accesses an electronic shopping site from Monday to Friday during the hours of 8:00 a.m. to 6:00 p.m., he is most likely at work and should receive different types of recommendations than those given on a weekend, for example,” he said.

“Using this context information could help to improve the recommendation systems, making them more refined and precise,” Domingues affirmed.

According to the researcher, one of the main obstacles is the lack of algorithms (command sequences) that make it possible for computer systems to obtain data automatically.

Another gap involves the exploration context information other than the date, time and user’s location.

“The objective of the study is to identify and utilize other types of context information and to develop algorithms that allow for automatic identification by recommendation systems so that the user makes suggestions that are coherent with the moment of access,” Domingues explained.

Examples of applications

Some of the first results of the study conducted by Domingues were utilized in a case study on a German site that suggests baby names to future parents in real time.

During the European Conference on Machine Learning and Principles and Practices of Knowledge Discovery (ECML PKDD), held on September 13 in Prague (Czech Republic), the administrators of the German site challenged programmers and researchers from around the world to develop an algorithm to improve its baby name recommendation system.

Domingues and his colleagues at LABIC and Universidade Federal do Mato Grosso do Sul (UFMS) developed an algorithm that cross-references the user’s location data at the moment that they are surfing with a list of the popular names in the user’s respective country, identified by accessing users in the same region to produce recommendations.

The project took fourth place in the international competition, and the results were published in an article in the annals of the conference.

“Using very simplistic contexts, such as access time and location, we can adapt the site’s name recommendation system to social and cultural contexts and the languages of the users of several countries,” affirmed Domingues.

“With our system, the results of a baby name search conducted by a user from Brazil, for example, are different from the suggestions made to a user in another country or who is also in Brazil but who accessed the site some months later, for example,” he said.

In another case study, the LABIC and UFMS researchers developed an algorithm to hierarchically and automatically organize the content of one agricultural news site released by the Brazilian Agricultural Research Corporation (Embrapa) by topic.

The sets of pages grouped by topic were utilized as contextual information to improve the precision of the recommendation system to users of the agricultural news portal.

The idea is that when a user surfs through published news on the site about soy, for example, the visit will be identified by the recommendation system, which will suggest news items related to soy that are categorized in other topics of the portal, such as articles on monitoring pests that attack crops.

The case study has been submitted for presentation during the 22nd International Conference on Pattern Recognition, which will occur in late August in Stockholm.

“We have obtained much better results with this type of contextual information using the user’s time and location data,” commented Domingues.

According to the researcher, the organization and categorization of information by subject is already utilized by news sites with the objective of facilitating user navigation.

The difference, however, is that the systems of these portals operate based on the premise that users know exactly which subjects they want to read and click directly on the topics related to their preferred themes.

The context-sensitive recommendation system that they intend to install presupposes that the user is navigating through domains that are still unknown and does not know exactly what she is looking for.

“The recommendation system attempts to help the user find something that he likes but does not know exactly how to find. By making a recommendation, the system attempts to guess what the user wants,” Domingues explained.