Time series forecasting (TSF) has gained significant attention as a widely explored research area in diverse applications. Existing methods, which focus on improvements in the most common scenarios, focus little on performance in rare cases. Despite their scarce occurrences in the data, these rare samples are more challenging and easily overlooked by models, significantly contributing to the total loss. In this paper, we propose a novel approach (dubbed iBACon) that overcomes this limitation by employing imbalance-aware contrastive learning and trend-seasonal decomposition architecture, specifically designed to solve TSF. To this end, we first introduce the Input-Output Difference (IOD) metric as a pseudo-label and reveal the data imbalance phenomenon in TSF. This label continuity inherently provides a meaningful distance between targets, implying a similarity between nearby targets in both label and feature spaces. Based on this similarity, the proposed imbalance-aware contrastive loss aims to reshape feature embeddings to facilitate knowledge dissemination among challenging samples and learn specific predictive features. Finally, when combined with our trend-seasonal decomposition network, iBACon significantly improves TSF accuracy. Experiments show that iBACon enhances overall average accuracy and substantially improves the 1-3% most challenging samples.

Multivariate time series forecasting (MTSF) has gathered extensive attention in various research areas. Many researchers leverage deep neural networks to explore spatial–temporal relationships for MTSF with great success. Nevertheless, plentifully available data required by deep neural networks often struggle to satisfy practical scenarios. To overcome this limitation, we exploit the internal structure of deep neural networks to automatically learn sample relationships, aiming to mitigate data scarcity by propagating information among different data samples. Consequently, in this paper, we propose a M ulti- R elations aware C onvolutional A ttention N etwork, termed MrCAN , which integrates spatial–temporal relation learning and sample relation learning, ad hoc for addressing MTSF issues. In MrCAN, we propose a novel spatial–temporal attention module for spatial–temporal relationship representation learning . With the proposed batch attention module, we further explore relational modeling among samples in each mini-batch to implicitly enrich the training sample information. In particular, we introduce parameter-sharing regressors located before and after the batch attention module to alleviate the training and testing inconsistency in learning batch invariant representations . Extensive experiments on six practical datasets demonstrate that MrCAN compares favorably to nine baseline models . Larger performance gaps are exhibited, especially on small datasets like the SML2010 dataset. Code is publicly available at https://github.com/JZhangNA/MrCAN .