VCSR-AβPSO: A Hybrid Feature Selection for Enhanced Android Scrolling Scenario Recognition in High-Dimensional Imbalanced Data

Scrolling scenario recognition in the lower-level Android system faces a high-dimensional imbalance problem. Most current research employs multi-stage feature selection methods for dimension reduction. However, it has some drawbacks: 1) The first stage is based on a combination of different feature importance, ignoring the interactions between the features; 2) Further reduction often traps in the problem of local optima due to fixed search steps in the feature space; 3) The reduced feature dimensions still cannot satisfy the real-time operational requirements of Android terminals. To tackle these issues, this work introduces VCSR-AβPSO, a two-stage hybrid feature selection method. The ELimination Et Choix Traduisant la REalité (ELECTRE) approach combines the importance of random forest features and cost-sensitive symmetric uncertainty to take feature interactions into consideration. Moreover, an adaptive β-hill climbing (AβHC)-enhanced Binary Particle Swarm Optimization (BPSO) algorithm is proposed to increase randomness in the search space, avoiding local optima and further reducing feature dimensions to improve the algorithm's real-time performance in terminal operation. AISSV dataset comprising various Android view features was constructed to validate the effectiveness and real-time capability of the proposed algorithm. Experiments on this dataset and other six public datasets have shown that the method performs exceptionally, achieving a 93.34
^%
F1 score in scrolling scenario recognition, surpassing other baseline algorithms to reach state-of-the-art levels, with model runtime controlled within 0.3 milliseconds, significantly improving the real-time performance.