Ranking database queries with user feedback: A neural network approach

Currently, websites on the Internet serving structured data allow users to perform search based on simple equality or range constraints on data attributes. However, to begin with, users may not know what is desirable to them precisely, to be able to express it accurately in terms of primitive equality or range constraints. Additionally, in most websites, the results provided to users can be sorted with respect to values of any one particular attribute at a time. For the user, this is like searching for a needle in a haystack because the user's notion of interesting objects is generally a function of multiple attributes. In this paper, we develop an approach to (i) support a family of functions involving multiple attributes to rank the tuples, and (ii) improve the ranking of results returned to the user by incorporating user feedback (to learn user's notion of interestingness) with the help of a neural network. The user feedback driven approach is effective in modeling a user's intuitive sense of desirability of a tuple, a notion that is otherwise near impossible to quantify mathematically. To prove the effectiveness of our approach, we have built a middleware for an application domain that implements and evaluates these ideas.