Fig. 6

Impact of the number of tasks in training set on each of the algorithms on a synthetic dataset. The X-axis denotes the size of the graph d, and the Y axes are the three types of evaluation metrics, FDR (false discovery rate), SHD (structural Hamming distance), and NNZ (number of non-zero entries). We fix the size of the graph to be 30, and the type of graph to be Erdos-Reni, the number of samples to be 100. We vary the number of patients from 10 to 100 and plot the performance against it. When the number of tasks is smaller than 30, adding more tasks to the problem increases the algorithms’ performance. This is because increasing more similar tasks also increases the number of samples for each algorithm. However, after the number of tasks is greater than 40, adding more tasks does not have an additional benefit for MetaNoTears-L1. Among the five algorithms, our algorithm has the lowest SHD and FDR (except for FDR when the number of tasks = 10), while predicting considerably more edges and closer to the true graph compared to MTL