FLAG: Faster Learning on Anchor Graph with Label Predictor Optimization

Knowledge graphs have received intensive research interests. When the labels of most nodes or datapoints are missing, anchor graph and hierarchical models can be employed. With an graph, we only need to optimize coarsest anchors, inferred from these anchors in a coarse-to-fine manner. The complexity optimization is therefore reduced cubic cost with respect number anchors. However, obtain high accuracy when data distribution complex, scale this set still needs large, which thus inevitably incurs expensive computational burden. As such, challenge scaling up how efficiently estimate while keeping classification performance. To address problem, propose novel approach that adds label predictor conventional models. In proposed approach, not directly optimized, instead, learn estimates their spectral representations. optimized regularization on all based show its solution involves inversion small-size matrix. Built upon hierarchy, design sparse intra-layer adjacency matrix over simultaneously accelerate embedding enhance effectiveness. Our named Faster Learning Anchor Graph (FLAG) as it improves anchor-graph-based methods terms efficiency. Experiments variety publicly available datasets sizes varying thousands millions samples demonstrate effectiveness our approach.