LSRFN

eterogeneous graph representation learning is crit-

ical for analyzing complex data structures. Metapaths within this

field are vital as they elucidate high-order relationships across the

graph, significantly enhancing the model’s accuracy and depth of

understanding. However, metapaths tend to prioritize long-range

dependencies of the target node, which can lead to the oversight of

potentially crucial 1st-order heterogeneous neighbors or short-

range dependencies. To address this challenge and circumvent

manual labeling, we propose the Long Short-Range Fusion

Network (LSRFN), an innovative unsupervised approach to

heterogeneous graph representation learning. LSRFN implements

two distinct masking strategies, short-range and long-range, to

obscure the features of the target node and its heterogeneous

neighbors. These representations are then learned independently

under each masking regime. In a subsequent step, features

learned with long-range masking are employed to reconstruct the

metapath-based adjacency matrix. Concurrently, features from

both masking conditions are leveraged to reconstruct the masked

node features jointly, culminating in the representation learning

process. Our experimental results affirm that LSRFN achieves

top-tier performance in node classification and clustering tasks

on the majority of datasets and remains competitive on the rest.