Accurate diagnosis is key in treating skin cancer. This study proposes a method to enhance the effectiveness of skin cancer recognition and classification based on Hybrid Learning Model (HLM). HLM architecture combines two deep learning models to extract diverse features from images, employs machine learning for classification, and integrates techniques for parameter optimization, class imbalance handling, and cross-validation, consisting of three main components. First, the base model performs feature extraction from images. This task leverages the strengths of two deep learning architectures: MLP-Mixer and Inception-ResNet-v2. MLP-Mixer focuses on exploiting spatial relationships between patches and color channels through fully connected layers, while Inception-ResNet-v2 uses multi-scale Inception blocks and Residual techniques to capture complex features and mitigate the vanishing gradient problem. Next, the use of hstack to combine features from the base model, then apply recursive feature elimination (RFE) to select important features and create the second dataset. Finally, the meta-model uses the second dataset with the Random Forest (RF) algorithm, leveraging multiple decision tree (DT) to reduce overfitting and increase accuracy through a majority voting method. Additionally, to address data imbalance, class weight adjustment, and StratifiedKFold are applied, with GridSearchCV used to find optimal RF hyperparameters. The proposed method is experimented with the ISIC HAM10000 dataset. The results demonstrate that independent models based on MLP-Mixer, Inception-ResNet-v2 achieve 92.07% and 96.55% accuracy rates, respectively. The fusion methods are based on Max voting, Average voting, Weighted voting result 96.52%, 96.77%, and 96.95% accuracy rates, respectively. Finally, our proposed method achieves 97.79% accuracy. These results highlight its potential for skin cancer diagnosis through imaging.