A Mish-regularized non-negative latent factorization of tensors model for dynamic QoS prediction

In big-data and service-computing applications, dynamic relationships frequently arise, such as the dynamic Quality of Service (QoS) data at the user end of web services. Such data is typically modeled as a high-dimensional sparse (HIDS) tensor, which contains rich knowledge of time patterns. The non-negative latent factorization of tensors (NLFT) model has proven highly effective in extracting such patterns from HIDS tensors. However, it usually encounters overfitting problems where the regularization scheme is inappropriate. To address this issue, this paper proposes an NLFT model regularized by the Mish activation function (MISH_NLFT), allowing the regularization term to dynamically adapt to the underlying structure of the data and enhance the expressive power of the learned latent factors. Experimental studies on two publicly available dynamic QoS datasets show that, compared with the traditionalL2-regularized model(L2_NLFT), the proposed MISH_NLFT model more accurately captures the latent temporal patterns and achieves superior prediction performance in QoS prediction tasks for service computing.