Heterogeneous graph attention stops at path endpoints no more: TAHAN groups intermediate nodes by type and leads DBLP and IMDB classification means
Synopsis
Attention aggregation on heterogeneous graphs has long treated node types as static labels, and metapath aggregation can overlook the semantic contribution of intermediate nodes. TAHAN instead collects neighbors along metapath prefixes, merges them across metapaths and levels by node type, then fuses them with within-type attention followed by across-type attention. On DBLP and IMDB node classification its mean Macro-F1 and Micro-F1 exceed the listed baselines, with both IMDB gains significant after Holm correction; on ACM it falls below SeHGNN.
Interpretation
Neighbor aggregation on heterogeneous graphs is now organized by node type: neighbors of one type are scored among themselves, and a second attention layer decides how much each type's representation contributes to the final node vector. Earlier heterogeneous graph attention networks mostly used metapaths as aggregation channels and focused on relations between endpoint types, leaving the semantics of node types passed through the middle of a path unused. On DBLP and IMDB node classification, TAHAN's mean Macro-F1 and Micro-F1 exceed the baselines listed in the table, and both IMDB gains have Holm-adjusted p-values below 0.001; on ACM its means are 0.75 and 0.86 percentage points below SeHGNN.
The approach treats metapaths as routes for collecting neighbors rather than as separate aggregation channels: nodes reached along every prefix of each metapath are merged across metapaths and hop counts into type-specific neighborhoods. After merging, each target node and neighbor type share one sampling budget, retained by degree-dependent Bernoulli sampling, so each type keeps at most that budget in expectation and the attention input stays bounded. Experiments cover the DBLP, IMDB, ACM and AMiner classification datasets plus AMiner-Large, PubMed and LastFM, with results reported as means and standard deviations over 10 independent random-seed runs.
Perspective
The evaluation covers the DBLP, IMDB, ACM and AMiner node classification datasets, large-scale classification on AMiner-Large, and link prediction on PubMed and LastFM, with all experiments run on the Toubkal supercomputer.
Merging discards path multiplicity and which prefix a neighbor arrived through, so a same-type neighbor reached at different path positions enters the sampled neighborhood as a single candidate; the type-level softmax gives types a shared weight budget, so raising one type's weight reduces what remains for others. The authors note that replacing attention also changes trainable parameters, so the ablations cannot fully separate the effect of weighting from model capacity.