This paper explores the use of Graph Neural Network models producing node embeddings, in order to solve the not fully addressed problem of detecting similar items stored in a knowledge base. Leveraging pre-trained models for textual semantic similarity, our proposed method PIKA aggregates heterogeneous (structured and unstructured) characteristics of an entity and its neighborhood to produce an embedding vector that can be used in different tasks such as information retrieval or classification tasks. Our method learns specific weights for each information brought by an entity, enabling us to process it in an inductive fashion.