Research on Optimization of Adaptive Data Dimensionality Reduction Algorithm for Internet of Things

The IoT has achieved interconnected communication, generating large-scale complex data that contains a large amount of noise and abnormal data. The concealment of abnormal data exacerbates the “curse of dimensionality”, affecting the processing time and recognition accuracy of model training. Kernel Principal Component Analysis (KPCA) uses nonlinear transformations to map data space to high-dimensional linear space and uses high-dimensional PCA to extract main components. However, KPCA faces the challenge of a single kernel function selection mechanism that cannot meet the characteristics of large amounts of IoT data, such as overfitting, high complexity, and applicability. And the KPCA kernel function parameters are usually empirical fixed values, which cannot meet the requirements of data time-varying. This paper proposes a mixed optimization algorithm that combines kernel functions adaptive selection and parameters adaptive estimation. Firstly, adaptive selection of kernel functions was achieved through minimum deviation, thereby improving its universality. Secondly, kernel function parameters were founded based on real-time estimation from dataset to meet time-varying nature, enhancing algorithm’s real-time adaptability. Thirdly, simulation analysis was conducted on the optimization algorithm’s feature values, number of features, computation time, and recognition accuracy. Results indicate that the adaptive hybrid kernel function optimization algorithm retains more eigenvalues and quantities; The computation time is equivalent to the Perception kernel function, which is about 90s and 160s higher than the Gaussian kernel function and Polynomial kernel function, respectively; The recognition accuracy is 5%, 11%, and 8% higher than that of Perception kernel function, Gaussian kernel function, and Polynomial kernel function.