Firms such as Amazon and Netflix have devised recommendation systems that help consumers to quickly wade through the mountains of information at their fingertips and zero in on the items they might be interested in buying. These systems provide a valuable service and yet, their potential is still not fully developed.

Rong Zheng of HKUST and his co-authors have been looking at recommendation systems and found that they tend to be based on a two-dimensional paradigm - that of user and item, with recommendations based on the users previous personal ratings of similar items and/or the ratings of others.

"Although the traditional two-dimensional user/item paradigm is suitable for some applications, such as recommending books and music CDs, it is significantly less suitable for 'context-rich' applications, such as travelling," they say.

"For example, when recommending vacations to travelers, one would likely recommend a different vacation to a customer in the winter than in the summer, so the time-of-travel context becomes clearly important."

"Or when a movie recommendation provider recommends a movie, it may also want to consider such additional dimensions as time when the when the movie was seen, the company it was seen with and the place. A completely different movie may be recommended to a student when he wants to see it on a Saturday night with his girlfriend in a movie theater than when he wants to see it on Thursday evening with his parents at home."

In view of this, they have devised a recommendation language - called REQUEST for REcommendation QUEry STatements - that captures multi-dimensional factors in a specialized, vertical way. This differs from other query languages which are more general and cumbersome and may not be able to produce recommendations in a similarly effective or intuitive way as REQUEST.

"REQUEST allows its users to express in a flexible manner a broad range of recommendations that are tailored to their own individual needs and therefore more accurately reflect their targets."

"For example, an application recommending a movie can have the following dimensions - the set of all movies that can be recommended, the people to whom movies are recommended, the movie theaters showing the movies, the times when the movie can or has been seen, and the people with whom one can see the movie. It can also use three rating measures - public rating, personal rating, and whether the user has actually seen the given movie."

Each claim is crunched by the REQUEST application and can provide an output that encompasses something like this: Jane gave a personal rating of 6 to the Aviator after seeing it with her boyfriend in movie theatre 5 on February 19, 2005, while the general public rated the movie 8. This information can then be stored and used in helping to devise future recommendations for Jane and other users.

The information can also be broken down or enhanced to do such things as a recommend the top five movies with public and personal ratings that are higher than 8 to students based on movies they have seen; identify the top two professions that appreciate the movie A Beautiful Mind the most; or recommend the top three movies and best times to see them over the weekend to a couple going on a date.

This kind of specific information has obvious appeal to companies as well as users.

"Such functionality would be useful to, for example, the analysts of a company providing recommendation services, who may want to take advantage of all the knowledge that their recommender system holds and analyse it from a variety of perspectives; in the movie example, perhaps they would want to know such things as the top two movie genres for each user age bracket. Alternatively, an agent in a call centre can use such a system to recommend the best plans and services to the customers of a telecommunications company," they say.

The authors performed extensive mathematical tests and comparisons of their model and found it performed well at its intended purpose: to express complex recommendations in a clear and concise manner, and thus help users to deal better with information overload.