<p>In recent years, significant success has been achieved in developing keyword search algorithms for incomplete graphs. A recent breakthrough work, KS-GNN, has transformed keyword search into an embedding-based similarity matching problem using graph neural network technology. The work effectively mitigates the time complexity issue in search problems and exhibits strong performance. However, the model’s utilization of node neighborhood information remains inadequate. Therefore, in this study, we propose a incomplete graph keyword search model, KS-LIAGNN Ultra, based on local information aggregation and graph autoencoder, achieved by enhancing the model’s aggregation mechanism. We first enhance the data using a local information aggregation module, followed by generating representative embeddings through an autoencoder structure. In the embedding space, we leverage an improved aggregation mechanism to fully utilize neighborhood information. By extensive experimental validation, our model consistently outperforms state-of-the-art method in terms of search accuracy on incomplete graphs.</p>