Trend Encoder with Attentive Neural Process: Node Performance Extrapolation for Non-Time-Series Datasets

Future network infrastructures will support multiple heterogeneous networks to facilitate 6G and network disaggregation. This will require the verification of various types of devices and components. Consequently, the efficiency of performance verification needs to be enhanced for the combinations of numerous network nodes and components, considering unknown network conditions. This study focuses on improving the inference accuracy of network node performance in the extrapolation. To address this issue, we propose a trend encoder for non-time-series datasets, which collaborates with an Attentive Neural Process. Preliminary extrapolation results show that the coefficient of determination of router throughput is improved by paying more attention to the trend information, including the relationship between both router configurations and input traffic influencing router performance for non-time-series datasets.