Tag: #time-series23 results
SERAF is a new framework that improves time series forecasting by combining numerical data with self-generated textual descriptions. By performing dual retrieval, it overcomes the limitations of traditional similarity-based methods in non-stationary environments.
This study compares encoder-only Transformers and LSTMs for streamflow inference in ungauged basins using NOAA National Water Model data. Results indicate that recurrent memory architectures remain superior for upstream sequence reconstruction tasks compared to Transformers.
KA-FCM replaces scalar weights in Fuzzy Cognitive Maps with learnable univariate B-spline functions on edges, enabling non-monotonic causal relationships per the Kolmogorov-Arnold theorem. It outperforms standard FCMs and rivals MLPs across non-monotonic inference, symbolic regression, and chaotic time-series forecasting. The approach preserves graph interpretability for extracting mathematical laws.
Analysis shows GluFormer's self-supervised framework transfers to ecology for unlabeled time-series like FLUXNET flux data. Addresses static metrics' limits, enabling early ecosystem risk prediction without annotations. Tokenization adapts Transformer to multi-scale ecological dynamics.
38-day dataset logs Gemini's daily stock forecasts, rationales, sentiment, and confidence over 10-day horizons. Time-locked and published on Hugging Face with interactive dashboard and Colab notebook. Studies LLM stability, narrative drift, and calibration under uncertainty.
KairosVL introduces the Semantic-Conditional Time Series Reasoning task, blending numerical modeling with contextual semantics for complex analysis. It employs a two-round reinforcement learning framework: first enhancing temporal primitive perception, then semantic-conditioned reasoning. The model delivers competitive results on synthetic and real-world tasks while boosting generalization.
A developer is surveying the embedded ML community to identify the primary pain points in time-series data workflows, such as cleaning, labeling, or deployment. They are building a hardware-agnostic, generative AI-native tool to streamline these processes for sensor-based applications.
A researcher is struggling to train a Transformer-decoder model on non-language series, noting that the model fails to learn auto-regressive behavior and gets stuck on single-token generation.
Tutorial shows building a deep learning model on Amazon SageMaker for solar flare detection. Uses LSTM networks trained on European Space Agency's STIX instrument data. Deploys the model for real-time predictions.
A team is developing a lightweight demand forecasting system for multi-location retail using only manually entered operational data like revenue, covers, and waste. It employs statistical baselines for the first 30 days and light global ML models thereafter, with outlier exclusion and confidence scoring. They seek feedback on global vs. local models for small datasets, outlier handling, and trustworthy confidence intervals.