Causal Structure Learning Algorithm Based on Streaming Features

Traditional causal structure learning algorithms rely on having full access to the full feature space throughout the learning process. However, in many real-world applications, all the features cannot be obtained in advance. We propose a causal structure learning algorithm based on streaming feature called CSBS which can effectively apply to the streaming feature space where the feature flow in one by one over time. For each incoming feature, Markov blanket is built dynamically by relevance analysis with every feature arrived and redundancy is analyzed with every relevant feature available whose Markov blanket is updated. When there is no available feature, according to the Markov blanket, we obtain the skeleton of network structure. Then we perform the greedy hill-climbing search constrained to only consider adding an edge discovered in previous skeleton to orient the edges and get the final Bayesian network structure. Experiment results show that our algorithm has superior accuracy over the state-of-the-art causal structure leaning algorithms.