Komparasi Arsitektur Deep Learning Berbasis Attention untuk Prediksi Quality of Service (QoS) Jaringan

Abstract

Quality of Service (QoS) is an important indicator in computer network management as it directly affects the quality of services experienced by users. Accurate QoS prediction can assist network administrators in making proactive decisions and improving network performance. This study compares the performance of four deep learning architectures, namely Conv1D, LSTM, Conv1D+Attention, and LSTM+Attention, for time-series QoS prediction. The dataset consists of 95,611 observations collected from a GNS3-based network simulation environment, including three QoS parameters: traffic_in_bps, traffic_out_bps, and latency_ms. Prior to model training, the data were processed using Min-Max Scaling and a sliding window approach to generate sequential inputs. Model performance was evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the coefficient of determination (R²). Experimental results show that the LSTM+Attention model achieved the best performance with an MAE of 0.003224, an RMSE of 0.013578, and an R² of 0.943872. Meanwhile, the Conv1D model demonstrated competitive performance with results close to the best-performing model. Further analysis indicates that the attention mechanism improved the performance of the LSTM architecture but did not provide similar benefits when applied to Conv1D. These findings suggest that the effectiveness of attention mechanisms is influenced by the characteristics of the underlying architecture used for QoS prediction.