Broad Learning System (BLS) provides a flexible modeling framework, which is a potential substitute of deep neural network models. Due to its fast adaptive ability of automatic model structure selection and online incremental learning strategies, BLS is referred to as a promising technology in the field of knowledge discovery and data engineering. However, traditional BLS model is mainly aimed at pattern classification tasks with approximately even-distributed data and equal misclassification cost. In real applications, most of pattern recognition tasks are unevenly-distributed, such as credit card fraud detection, network intrusion detection, medical diagnosis, etc. In this paper, a Data distribution-based Cost-sensitive-BLS (DDbCs-BLS) is proposed, aimed at solving the problem of pattern classification tasks with imbalance data and varying misclassification costs on different classes. DDbCs-BLS can achieve the best classification boundary by adopting the cost sensitive BLS learners, and ensure the lossless of the information of sparse classes, so as to ensure the classification performance of BLS classifier in various data sets. DDbCs-BLS was validated on multiple public data sets (including balanced and imbalanced data sets). Extensive validation and comparative results show that DDbCs-BLS can effectively determine the best location of the classification boundary line, consequently, it can achieve better classification performance on both balanced and imbalanced data sets.