Abstract: Formal concept analysis which is an unsupervised learning method for conceptual clustering constitutes an appropriate framework for data mining. However, due to the completeness of concept lattice, the task of constructing the lattice is known to be computationally expensive. The iceberg lattice of context, a substructure of the complete concept lattice, served as a condensed representation of frequent itemsets. And it is well suited for analyzing very large database. And building concept lattice by merging factor lattices drawn from data fragments may be adapted to distributed data mining environment. Inspired by those ideas, a novel algorithm called Icegalamera for iceberg concept lattice assembly from heterogeneous relational tables is presented and is utilized for closed frequent itemsets mining. The completeness of closed frequent itemsets produced by Icegalamera is proved both in theory and in empirical way, and then the merge mapping process is analyzed and implemented from partial iceberg concept lattices to global one. This algorithm avoids computation of structuring the complete concept lattice. Furthermore the merge and pruning strategies are adopted, which makes the algorithm's efficiency outperforms that of the apriori algorithm on generating frequent itemsets under condense and sparse data set.