This paper addresses the problem of directly identifying and exploiting statistical arbitrage (StatArb) in options markets using graph learning. The authors propose a two-stage approach. In the first stage, they define a novel prediction target based on synthetic zero-coupon bonds constructed from synthetic-long and underlying instrument positions; deviations from the law of one price represent pure arbitrage opportunities. They develop RNConv, a graph convolutional network that integrates Neural Oblivious Decision Ensembles (NODE) to effectively handle tabular node features while capturing relational structures among options. In the second stage, they introduce SLSA (Synthetic-Long-Short-Arbitrage), a class of positions that are provably minimal risk and neutral to all Black-Scholes risk factors under the arbitrage-free assumption. An SLSA projection method converts arbitrage predictions into such positions. Experiments on KOSPI 200 index options demonstrate that RNConv statistically significantly outperforms standard GNN baselines (GCN, SAGE, GAT, GPS), and SLSA achieves consistently positive returns with an average P&L-contract information ratio of 0.1627. The work bridges gaps in applying deep learning to StatArb in options markets and in incorporating tree-based methods into graph neural networks for tabular features.