A Hybrid Graph and Vision Transformer Framework for Irregularly Sampled Multivariate Time Series Classification
Author
Roy, Nipa
Term
4. term
Education
Publication year
2026
Submitted on
2026-06-12
Abstract
Irregularly sampled multivariate time series occur frequently in areas such as healthcare data, sensor monitoring, and human activity recognition. In these settings, measurements are taken at uneven time intervals and missing values are common, which makes it difficult for conventional time-series models that assume regular sampling and complete data to perform accurate classification. This thesis proposes a hybrid framework to classify such irregular multivariate time series by combining graph-based representation learning with image-based transformer classification. First, GraFITi is used to represent the irregular observations as sparse graphs, preserving both the temporal structure and the relationships between variables without relying on interpolation or heavy preprocessing. Next, the learned graph representations are converted into line-graph images, which are then classified using Vision Transformer (ViT) and Swin Transformer architectures. The framework is evaluated on three datasets: PhysioNet 2012 (P12), PhysioNet 2019 (P19), and PAMAP2. The experimental results show that using GraFITi-based graph learning improves classification performance and allows the visual models to capture both temporal patterns and cross-variable relationships effectively. Overall, the study demonstrates that integrating graph representation learning with vision-based classification is a flexible and robust approach to analyzing irregular multivariate time series.
Uregelmæssigt samplede multivariate tidsserier optræder ofte i fx sundhedsdata, sensorovervågning og registrering af menneskelig aktivitet. Her bliver målinger foretaget på ujævne tidspunkter, og der mangler ofte værdier. Det gør det svært at klassificere data med traditionelle tidsserie-modeller, som typisk antager faste tidsintervaller og komplette målinger. I dette speciale præsenteres en hybrid metode til klassifikation af sådanne uregelmæssige tidsserier. Metoden kombinerer grafbaseret repræsentationslæring med billedbaseret klassifikation ved hjælp af transformer-modeller. Først bruges GraFITi til at modellere de uregelmæssige målinger som tynde (sparse) grafer. På den måde bevares både tidslig struktur og relationer mellem variabler uden brug af interpolation eller omfattende forbehandling. Derefter omdannes de lærte grafrepræsentationer til såkaldte line-graph billeder, som klassificeres ved hjælp af Vision Transformer (ViT) og Swin Transformer-arkitekturer. Metoden er testet på tre datasæt: PhysioNet 2012 (P12), PhysioNet 2019 (P19) og PAMAP2. Resultaterne viser, at brugen af GraFITi-baseret graflæring forbedrer klassifikationspræstationen og gør det muligt for de visuelle modeller at fange både tidslige mønstre og sammenhænge mellem variabler. Samlet peger studiet på, at kombinationen af grafrepræsentationslæring og visionsbaseret klassifikation er en fleksibel og robust tilgang til analyse af uregelmæssige multivariate tidsserier.
[This abstract has been rewritten with the help of AI based on the project's original abstract]
Keywords