This paper reports the development of a methodology for identifying and improving occupant behavior in existing residential buildings. In this study, end-use loads were divided into two levels (i.e. main and sub-category), and they were used to deduce corresponding two-level user activities (i.e. general and specific occupant behavior) indirectly. The proposed method is based on three basic data mining techniques: cluster analysis, classification analysis, and association rules mining. Cluster analysis and classification analysis are combined to analyze the main end-use loads, so as to identify energy-inefficient general occupant behavior. Then, association rules are mined to examine end-use loads at both levels, so as to identify energy-inefficient specific occupant behavior. In order to demonstrate its applicability, this methodology was applied to a group of residential buildings in Japan, and one building with the most comprehensive household appliances was selected as the case building. The results show that, for the case building, the method was able to identify the behavior which needs to be modified, and provide occupants with feasible recommendations so that they can make required decisions. Also, a reference building can be identified for the case building to evaluate its energy-saving potential due to occupant behavior modification. The results obtained could help building occupants to modify their behavior, thereby significantly reducing building energy consumption. Moreover, given that the proposed method is partly based on the comparison with similar buildings, it could motivate building occupants to modify their behavior. (C) 2011 Elsevier Ltd. All rights reserved.