Index selection is an important part of physical database design. Its goal is to select an appropriate set of indexes to minimize the cost for a given workload under storage constraint. However, selecting a suitable configuration of indexes is a difficult problem to solve. The problem becomes more complex for indexes defined on multiple tables such as bitmap join indexes, since it requires the exploration of a much more search space. Studies dealing with the bitmap join indexes selection problem mainly focused on proposing pruning solutions of the search space by the means of data mining techniques or heuristic approaches. So far, the data mining based approaches have used closed frequent itemsets to reduce the search space for the selection process. These approaches have two notable shortcomings. Firstly, they generate a huge number of indexes with a lot of redundancy that it is very difficult to manage according to the system limitation (number of Indexes per table, storage space constraint). Secondly, when they construct the extraction context for mining frequent sets of attributes, they have used indexable attributes only once for each query in the workload which does not reflect the importance of a given query in the workload. Indeed, the queries in a workload are unlikely to have the same probability of being requested. To overcome these imitations, we propose to combine maximal frequent itemsets and query frequencies to improve the quality of generated indexes. This paper describes an approach that refines the index selection process, incorporating query frequencies in the extraction context for mining frequent set of attributes. We experimentally prove that our approach reduces the storage space and improves the quality of the recommended indexes.