Network intrusion detection systems are playing a significant role in protecting communication infrastructures against sophisticated cyber-attacks. Existing machine learning classification methods are needed to achieve accurate classifications of network data. Network cyber-attacks are diverse and dynamic in nature. Thus, existing methods sometimes fail due to a lack of multi-view learning for cyber-attacks, which captures the complex, high-dimensional, and non-linear relationships inherent in network traffic data. These limitations lead to reduced detection accuracy and poor generalisation for diverse and evolving attack patterns. This paper proposes a Hybrid Feature Representation Framework (HFRF) that integrates multi-stage feature refinement, multi-view transformation, and adaptive feature selection for effective network data classification. The proposed framework initially performs variance- and mutual information-based filtering to retain relevant features, followed by dual transformations using Principal Component Analysis (PCA) and Non-Negative Matrix Factorization (NMF) to capture both global and localised data structures. These complementary representations are then fused and optimised using an ensemble-based feature-importance mechanism to select the most discriminative features dynamically. Extensive experiments are conducted on benchmark datasets, including CICIDS2018 and UNSW-NB15, to demonstrate the efficacy of the proposed method compared to other methods.