This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
In this paper, we present the Time-IMM dataset, which is specifically designed to capture causal irregularities, to address the problems of irregular, multimodal, and messy time series data (different sampling rates, asynchronous modes, and widespread missing data) that frequently occur in real-world applications such as healthcare, climate modeling, and finance. Time-IMM represents nine types of time series irregularities classified into trigger-based, constraint-based, and artifact-based mechanisms. Along with this dataset, we introduce IMM-TSF, a benchmark library for irregular multimodal time series forecasting. IMM-TSF includes specialized fusion modules, including a timestamp-text fusion module and a multimodal fusion module, and supports both averaging and attention-based integration strategies considering recency. Experimental results show that explicitly modeling multimodality in irregular time series data significantly improves forecasting performance. Time-IMM and IMM-TSF provide a foundation for advancing time series analysis in real-world conditions. The dataset is available at https://www.kaggle.com/datasets/blacksnail789521/time-imm/data and the benchmark library is available at https://anonymous.4open.science/r/IMMTSF_NeurIPS2025 .
We provide a new dataset (Time-IMM) and benchmark library (IMM-TSF) that effectively addresses irregular, multimodal, and missing value problems in real-world time series data.
◦
We empirically demonstrate that explicitly modeling irregularity in multimodal time series data significantly improves forecasting performance.
◦
A dataset that includes various types of irregularities, overcoming the limitations of existing regular time series data-based research and contributing to solving real-world problems.
◦
Asynchronous data integration and practical evaluation possible with timestamp-text fusion and multi-modal fusion modules.
•
Limitations:
◦
Generalizability of the Time-IMM dataset: Generalizability on real time series data containing other types of irregularities than the nine presented is required.
◦
Scalability of the IMM-TSF benchmark library: Scalability needs to be improved by adding more diverse algorithms and evaluation metrics.
◦
Dataset size and diversity: There is a need to improve generalization performance by including data of larger size and from more diverse domains.
