ENHANCING MULTIVARIATE TIME SERIES FORECAST- ING WITH MUTUAL INFORMATION-DRIVEN CROSS- VARIABLE AND TEMPORAL MODELING

Shiyi Qi, Liangjian Wen, Yiduo Li, Yuanhang Yang, Zhe Li, Zhongwen Rao, Lujia Pan, Zenglin Xu

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: time series forecasting

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

Recent researches have showcased the significant effectiveness of deep learning techniques for multivariate time series forecasting (MTSF). Broadly speaking, these techniques are bifurcated into two categories: Channel-independence and Channel-mixing approaches. While Channel-independence models have generally demonstrated superior outcomes, Channel-mixing methods, especially
when dealing with time series that display inter-variable correlations, theoretically promise enhanced performance by incorporating the correlation between variables. However, we contend that the unnecessary integration of information through Channel-mixing can curtail the potential enhancement in MTSF model performance. To substantiate this claim, we introduce the Cross-variable Decorrelation Aware feature Modeling (CDAM) for Channel-mixing approaches. This approach is geared toward reducing superfluous information by minimizing
the mutual information between the latent representation of a single univariate sequence and its accompanying multivariate sequence input. Concurrently, it optimizes the joint mutual information shared between the latent representation, its univariate input, and the associated univariate forecast series. Notably, prevailing techniques directly project future series using a single-step forecaster, sidelining the temporal correlation that might exist across varying timesteps in the target series. Addressing this gap, we introduce the Temporal correlation Aware Modeling (TAM). This strategy maximizes the mutual information between adjacent sub-sequences of both the forecasted and target series. By synergizing CDAM and TAM, we sculpt a pioneering framework for MTSF, named as InfoTime. Comprehensive experimental analysis have demonstrated the capability of InfoTime to consistently outpace existing models, encompassing even those considered state-of-the-art.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8290