Institutional Repository of Vilnius University
Anomalios difuzijos procesų klasifikacija taikant mašininį mokymąsi /
Abstract
dc:descriptionThis work presents a compact convolutional neural network (SimpleDiffusionClassifier, 26 260 parameters) for classifying four types of stochastic processes: normal diffusion, subdiffusive fractional Brownian motion (fBm, H < 0.5), superdiffusive fBm (H > 0.5), and the Ornstein-Uhlenbeck (OU) process, based solely on trajectory morphology. Trajectories of 300 time steps were generated using the Davies-Harte FFT method, which produces mathematically exact fBm with MSD scaling errors below 1%, and instance normalization was applied to remove amplitude and drift information before classification. The model was trained on 16 000 synthetic trajectories (12 800 training, 3 200 test, balanced across four classes) and achieved a weighted F1 score of 0.95. A comparison of five CNN architectures ranging from 20 460 to 263 908 parameters showed a saturation effect, with all models scoring F1 between 0.94 and 0.95, confirming that a compact architecture is sufficient for this task. Boundary analysis revealed that classification confidence decreases from approximately 98% to 70% in the transition zone (H in [0.4, 0.6]), consistent with the physical indistinguishability of trajectories near H = 0.5. The model was applied to two real-world datasets: Bitcoin daily log-prices (2017--2020, 83 sliding windows) were classified as a random walk with 94.2% mean confidence, consistent with the efficient market hypothesis, while deseasonalized Vilnius daily temperature data (2000--2023, 831 windows) were unanimously classified as an Ornstein-Uhlenbeck process with near-100% confidence, reflecting the mean-reverting dynamics of synoptic atmospheric fluctuations.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Zigmantas, Jovaras,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308117404&prefLang=en_US