Temporal structure, similarity, and prediction.
My published work forms three connected themes: representation and decomposition, similarity and retrieval, and prediction and regression. The map places each paper under its main contribution and connects these themes to my current research.
My MRes and PhD work studies representation operators Φ: T → R, their composition, and five properties: information preservation, distance preservation, invariance, reconstructability, and commutativity. This framework connects questions about temporal structure with my earlier work on output dependence and interpretable prediction; each task requires different information to be retained.
Software: DeTime and TSDecompose-Benchmark.
Stationarity-Aware Retrieval-Augmented Time Series Forecasting (SARAF, KDD 2026) selects and aggregates historical segments using temporal similarity, diversity, and stationarity. This connects retrieval directly to forecasting.
Software: EchoTime provides explainable structural comparison of time series and datasets.
These directions extend the published work above.
Current work studies temporal tokens for foundation and world models, including Time Series Language and Foundation Model and Language-Action Time-Series Tokenization for Efficient VLA Policies. See the associated compute allocations.
Literature examples: PAA stores segment means; SAX adds symbolic encoding, while SAA-SAX encodes slope-based aggregates.
Recent work — October 2026: TOMC extends my interest in compact sequence representations to conversational histories. It constructs task-oriented records and combines them with selected source text under a memory budget.
Semantics-Enhanced Retrieval-Augmented Time Series Forecasting is accepted / in press for an ICML 2026 workshop. It extends the retrieval research theme.
Time-series classification extends the representation question to distinguishing classes from temporal structure.
See the publication list for full bibliographic details and the software page for tools.