LSRFN

eterogeneous graph representation learning is crit-ical for analyzing complex data structures. Metapaths within thisfield are vital as they elucidate high-order relationships across thegraph, significantly enhancing the model’s accuracy and depth ofunderstanding. However, metapaths tend to prioritize long-rangedependencies of the target node, which can lead to the oversight ofpotentially crucial 1st-order heterogeneous neighbors or short-range dependencies. To address this challenge and circumventmanual labeling, we propose the Long Short-Range FusionNetwork (LSRFN), an innovative unsupervised approach toheterogeneous graph representation learning. LSRFN implementstwo distinct masking strategies, short-range and long-range, toobscure the features of the target node and its heterogeneousneighbors. These representations are then learned independentlyunder each masking regime. In a subsequent step, featureslearned with long-range masking are employed to reconstruct themetapath-based adjacency matrix. Concurrently, features fromboth masking conditions are leveraged to reconstruct the maskednode features jointly, culminating in the representation learningprocess. Our experimental results affirm that LSRFN achievestop-tier performance in node classification and clustering taskson the majority of datasets and remains competitive on the rest.