Probabilistic Ranking of Database Query Results

• The first step is to divide the attributes in a table into specified (the ones specified by the query) and the unspecified attributes (the ones not specified by the query). • Every tuple in the database is considered a document and correlation between different attributes in the tuple are found. Correlations are often ignored in high dimensional and sparsely populated data spaces in IR but there are strong correlations between the attribute values in relational data spaces. • The authors make a limited independence assumption the specified (and unspecified) attributes within them are assumed to be independent but correlations between specified and unspecified attributes are allowed. • Two kinds of scores are used to rank the documents. A global score which captures the global score of unspecified attributes and a conditional score that captures the strength of correlations between specified and unspecified attribute values. • In the preprocessing phase of the computation two lists global list and the conditional list containing the global and conditional scores of the attributes for tuples are calculated. These are stored as auxiliary tables in the database. • Instead of pre-computing the Top K results for all possible queries a ranked list of tuples for all atomic queries is calculated. Threshold Algorithm, a well known Top K algorithm is adapted to rank the queries using the computed scores at query time. • The efficient adaptation of the algorithm is due to the limited independence assumption that is novel to this paper • The processing involves two modules: (i) The index module (the preprocessing step), where the global and the conditional lists are constructed, and (ii) the List Merge Algorithm that is used to merge the lists associated with the attributes at query time. • Extensive experiments are carried out on the internet movie database and the MSN home advisor database. • The results are compared to the results from a rival query ranking method.