FlowPipe Enhances Data Prep with LLM-Driven Generative Flow Networks

Data preparation is a critical yet complex step in machine learning, involving sequential transformations to clean raw data. Current automated methods struggle with the combinatorial nature of operator sequences and inefficient exploration. FlowPipe introduces a novel approach that frames pipeline construction as a conditional probabilistic flow generation problem. This framework leverages Conditional Generative Flow Networks (C-GFlowNets) to link terminal validation rewards with early pipeline decisions, improving long-horizon credit assignment. It also integrates Deep Semantic Modulation using Feature-wise Linear Modulation (FiLM), allowing large language model (LLM) insights to guide the policy based on dataset semantics. Furthermore, FlowPipe incorporates "failure awareness" to avoid invalid states and focus the search on promising regions. Experimental results across 74 real-world datasets demonstrate FlowPipe's superior performance, achieving an average accuracy improvement of 11.96% and 12.5x faster training convergence compared to state-of-the-art baselines. The source code is publicly available.