The abuse of encryption and obfuscation in malware poses a significant threat to cybersecurity. Dynamic API call sequences, which directly reflect malware behavior and are hard to falsify, offer more robust and reliable features for classification and detection than static ones. Based on our analysis, we identify the following key characteristics in API call sequences: (1) The implementation of malicious functionality often involves the allocation and interaction of resources. API calls to the same resource object, such as files or registries, typically exhibit contextual dependencies, regardless of whether they are adjacent; (2) Multi-process interleaved execution is common in malware, and API sequences can be organized by execution order or process grouping. The sorting method can impact model performance, especially for multi-process malware; (3) API sequences often contain many consecutive repeated API names, but their parameters may differ. Therefore, we can distinguish these repeated calls by their parameters, rather than simply removing redundancy through truncation. Based on these observations, we propose a malware classification ensemble model that integrates multi-dimensional API sequence features. Specifically, we train separate classification models based on three different feature perspectives: the API resource graph, multi-process API sequence representation, and parameter-enhanced API name sequences. The outputs of these three base models are then aggregated using K-Nearest Neighbors (KNN) soft voting. Training and evaluation on three classification tasks demonstrate that the three base models outperform existing API sequence-based detection techniques, and the ensemble model further enhances the detection performance.