ICLR 2024

A continuación una lista de los posters relacionados a series temporales, donde destaco los que me parecieron más relevantes con respecto a nuestro interés, de lo que fueron las diferentes sesiones de posters de ICLR 2024.

Pongo todos los de series temporales por si ustedes ven algo que les parezca más interesantes que los que destaqué. También para mostrar la relevancia que tuvieron las series temporales en la sesiones.

Sesión 1

• RobustTSF: Towards Theory and Design of Robust Time Series Forecasting with Anomalies
• Time-LLM: Time Series Forecasting by Reprogramming Large Language Models
• Soft Contrastive Learning for Time Series
• Contrastive Difference Predictive Coding
• Towards Transparent Time Series Forecasting
• STanHop: Sparse Tandem Hopfield Model for Memory-Enhanced Time Series Prediction
• Leveraging Generative Models for Unsupervised Alignment of Neural Time Series Data
• Transformer-Modulated Diffusion Models for Probabilistic Multivariate Time Series Forecasting
• TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting
• Parametric Augmentation for Time Series Contrastive Learning

Sesion 2

• REBAR: Retrieval-Based Reconstruction for Time-series Contrastive Learning
• iTransformer: Inverted Transformers Are Effective for Time Series Forecasting
• Generative Modeling of Regular and Irregular Time Series Data via Koopman VAEs
• Generative Learning for Financial Time Series with Irregular and Scale-Invariant Patterns

Sesión 3

• Pathformer: Multi-scale Transformers with Adaptive Pathways for Time Series Forecasting
• MG-TSD: Multi-Granularity Time Series Diffusion Models with Guided Learning Process
• Feature-aligned N-BEATS with Sinkhorn divergence
• CausalTime: Realistically Generated Time-series for Benchmarking of Causal Discovery
• Explaining Time Series via Contrastive and Locally Sparse Perturbations

Sesión 4

• Rethinking Channel Dependence for Multivariate Time Series Forecasting: Learning from Leading Indicators
• GAFormer: Enhancing Timeseries Transformers Through Group-Aware Embeddings
• Biased Temporal Convolution Graph Network for Time Series Forecasting with Missing Values
• CNN Kernels Can Be the Best Shapelets
• Bayesian Bi-clustering of Neural Spiking Activity with Latent Structures
• AmortizedPeriod: Attention-based Amortized Inference for Periodicity Identification
• VQ-TR: Vector Quantized Attention for Time Series Forecasting

Sesión 5

• Stable Neural Stochastic Differential Equations in Analyzing Irregular Time Series Data
• T-Rep: Representation Learning for Time Series using Time-Embeddings
• TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting
• Copula Conformal prediction for multi-step time series prediction
• TACTiS-2: Better, Faster, Simpler Attentional Copulas for Multivariate Time Series
• ARM: Refining Multivariate Forecasting with Adaptive Temporal-Contextual Learning
• Periodicity Decoupling Framework for Long-term Series Forecasting
• Diffusion-TS: Interpretable Diffusion for General Time Series Generation

Sesión 6

• SocioDojo: Building Lifelong Analytical Agents with Real-world Text and Time Series
• WOODS: Benchmarks for Out-of-Distribution Generalization in Time Series
• Self-Supervised Contrastive Learning for Long-term Forecasting
• Interpretable Sparse System Identification: Beyond Recent Deep Learning Techniques on Time-Series Prediction
• DAM: Towards a Foundation Model for Forecasting

Sesión 7

• FITS: Modeling Time Series with 10k Parameters
• Learning to Embed Time Series Patches Independently
• Multi-Resolution Diffusion Models for Time Series Forecasting
• ModernTCN: A Modern Pure Convolution Structure for General Time Series Analysis

Sesión 8

• Towards Enhancing Time Series Contrastive Learning: A Dynamic Bad Pair Mining Approach
• Inherently Interpretable Time Series Classification via Multiple Instance Learning
• Conditional Information Bottleneck Approach for Time Series Imputation
• TEST: Text Prototype Aligned Embedding to Activate LLM's Ability for Time Series
• Disentangling Time Series Representations via Contrastive Independence-of-Support on l-Variational Inference

Notas:

Primero que nada me llamó la atención la gran cantidad de trabajos relacionados al tema. Por otro lado pude ver que la mayor preocupación, supongo porque es la tarea "más desafiante" o porque es la que más vende, es el forecasting. De todas maneras muchos trabajos se presentan como time-series analysis y prueban su modelos en diferentes tareas como: forecasting, reconstrucción, detección de anomalías, generación de datos, relleno de huecos, etc. Es decir, no atan su trabajo a una sola aplicación, lo mismo para los que se presentan como solo de forecasting, prueban su modelo en diferentes tareas.

También pude ver que dentro de los temas que no hemos visto se utiliza bastante Contrastive Learning aplicado a series temporales. Una referencia que se repetía mucho a lo largo de diferentes trabajos era: TS2Vec: Towards Universal Representation of Time Series