This work, grounded in quantum probability interpretation and catastrophe of orthogonality, investigates the representational and generalization capabilities of quantum-probabilistic tensor networks in machine learning tasks such as generation and classification, through the lens of loss function behavior and accuracy variation. Experiments conducted on the Fashion-MNIST dataset systematically examine the impact of different hyperparameters on model performance, accompanied by corresponding theoretical analyses. Building on these insights, we propose an improved multi-qubit encoding scheme that enhances the performance of tensor networks in classification tasks. The proposed method demonstrates improvements on the Fashion-MNIST dataset.