Automated machine learning (AutoML) aims for constructing machine learning\n(ML) pipelines automatically. Many studies have investigated efficient methods\nfor algorithm selection and hyperparameter optimization. However, methods for\nML pipeline synthesis and optimization considering the impact of complex\npipeline structures containing multiple preprocessing and classification\nalgorithms have not been studied thoroughly. In this paper, we propose a\ndata-centric approach based on meta-features for pipeline construction and\nhyperparameter optimization inspired by human behavior. By expanding the\npipeline search space incrementally in combination with meta-features of\nintermediate data sets, we are able to prune the pipeline structure search\nspace efficiently. Consequently, flexible and data set specific ML pipelines\ncan be constructed. We prove the effectiveness and competitiveness of our\napproach on 28 data sets used in well-established AutoML benchmarks in\ncomparison with state-of-the-art AutoML frameworks.\n