부산대학교 도서관

Forecasting time series values based on historic covariates has been an active area of research in statistics and machine learning. With the availability of computation resources and big data infrastructure supporting massive volume, velocity and variety, the algorithms have evolved from classic statistical learning to neural-network driven loss minimisation techniques. While state of the art attention and self-attention-transformers have shown promise of improved performance with sufficient training data, most of them fail to generalise to different problems of time-series modelling (such as classification and extremum forecasting) with asynchronously sampled covariates. This paper introduces the concept of a generalised time series embedding and transfer learning for time series (analogous to token-to-vector or image-to-vector embeddings in language and vision models respectively) that allow joint training with a unified interface. The major benefit of this work is a unified embedding model employing multi-attention for feature representation which enables benchmark performance against state of the art models from recent literature.