Trans-Ising Improves Ising Model Estimation with Transfer Learning

Estimating high-dimensional Ising models is often hampered by limited target sample sizes, making it challenging to leverage auxiliary binary datasets of unknown relevance. This research introduces Trans-Ising, a novel transfer learning approach designed to overcome these limitations. Trans-Ising operates in two main stages. First, it employs a loss-based source screening rule that uses held-out target pseudolikelihood to identify informative auxiliary sources, thereby preventing "negative transfer" where irrelevant data degrades performance. Second, it computes an initial estimator using pooled nodewise L1-regularized logistic regression, followed by a target-only correction step that applies a folded-concave penalty. The method has strong theoretical guarantees, including fixed-node L2 and L1 error bounds and exact graph selection consistency. Extensive simulations and real-data analyses confirm that Trans-Ising achieves lower estimation errors compared to both target-only estimation and naive data pooling, demonstrating its effectiveness in leveraging auxiliary data.