A new AI method for retrieving time series data from large data sources
Even as multi-modal AI has advanced rapidly in its ability to analyze massive data quantities, it has shown persistent weakness in analyzing time series data at scale. As discussed in earlier NSDPI research, time series analysis is vital to how analysts reconstruct sequences of events, identify causality, and anticipate future developments.
To bridge this gap, this NSDPI paper explores how AI models can be better trained to retrieve relevant text for given time series and vice versa. Because limited datasets have been a persistent challenge in this type of research, we derive a new dataset from Federal Reserve Economic Data (FRED) focused on blog posts documenting economic processes paired with economic time series.
Using this FRED data, we train and study a new AI method based on Contrastive Language-Image Pre-training called FRED-CLIP, a dual-encoder contrastive learning model trained on FRED plot-text pairs. Our experiments show FRED-CLIP can successfully retrieve relevant text and time series data from large data sources. This could lead to exciting downstream applications like faster data discovery, automated report generation, and cross-modal search.
Key Takeaways:
- Full fine-tuning of both the time series and language encoders in our experiments consistently delivered better results than keeping them frozen. This suggests learning modality-specific embeddings is important in this setting.
- Chart-specific image preprocessing modifications do not consistently improve retrieval performance, although some, such as grayscale and aspect ratio padding, are less harmful than others.