The K-Nearest Neighbors (KNN) algorithm is widely used in classification tasks due to its computational simplicity, effectiveness, and model-free nature. However, the traditional KNN algorithm and its variant, Fuzzy KNN (FKNN), exhibit two fundamental limitations that constrain their practical utility. On the one hand, the static neighborhood size (k-value) fails to account for local data distribution variations, on the other hand, reliance on a single distance metric proves inadequate for characterizing complex sample similarities in high-dimensional or heterogeneous feature spaces. These inherent shortcomings significantly degrade algorithmic performance in real-world applications. To address these issues, this paper proposes a Multi-Distances Weighted Adaptive Fuzzy K-Nearest Neighbors (MDA-FKNN) algorithm. The algorithm first constructs a multi-distances fusion module. This module integrates Euclidean, Manhattan, cosine, and Mahalanobis distances through weighted combination to improve sample similarity assessment accuracy. Second, it integrates FISTA optimization with multi-model ensemble learning to train a dynamic k-value predictor. This enables data-adaptive neighborhood adjustment by capturing local data characteristics. Finally, Bayesian optimization automates the tuning of weights in the multi-distances fusion module and FISTA hyperparameters, enhancing overall algorithmic performance. To evaluate efficacy, we benchmarked MDA-FKNN against baseline methods on 15 public datasets. Experimental results demonstrate that the MDA-FKNN algorithm significantly outperforms the baseline algorithms across all metrics, including accuracy, precision, recall, and Macro-F1 score. Furthermore, statistical significance analysis confirms that MDA-FKNN achieves the best overall performance, exhibiting superior robustness and adaptability.