With the prevalence of service composition, how to recommend API services that meet the Mashup requirements for developers has become a challenging in the field of service computing. Existing works on Web service recommendation typically conduct service recommendation based on the semantic requirements and interaction relationships of Web services, while ignoring the construction intent of Mashup creators. Moreover, they often only focus on the impact of single intent on recommendation, neglecting the diverse requirements of Mashup creators at different intent-levels. As a result, the recommendation models fail to comprehensively understand the construction intent of Mashup creators, which affects the quality and effectiveness of Web service recommendation. To address this problem, this paper proposes an intent recognition-based API recommendation method, denoted as ARIR. This method utilizes the annotation information of Mashup and API to analyse the creation intent of Mashup and the functional intent of API, thereby providing more accurate API with high-quality to Mashup creators. Firstly, it decouples the representations of Mashup and API at different intent-levels and independently initializes the node representations at each intent-level. Secondly, it trains the vector representations at each intent-level using a decoupled graph convolutional neural network module and obtains the representation vectors of Mashup/API at different intent-levels with attention weights. Then, it aggregates the intents of Mashup and API using the Mashup-API interaction relationships, resulting in the final node representations of Mashup/API. Furthermore, it constructs a similarity heterogeneous network by calculating the Mashup-to-Mashup similarity and API-to-API similarity, updating the node representations by using the Mashup and API feature matrices and adjacency matrices, and obtaining the final recommendation prediction results by using a fully connected layer. Finally, the experimental results conducted on real-world Web service datasets demonstrate that the ARIR outperforms the best-performing baseline method with Recall@20 of 3% and NDCG@20 of 5%.