In recent years, Transformer-like models have increasingly underscored their importance in Time Series Forecasting (TSF), playing a pivotal role across various fields. However, their attention mechanisms never consider the dependencies across variables, which can diminish the predictive performance. Moreover, these models overlook the inherent non-stationarity of real-world scenarios. To address these challenges, a Dual-Pathway PatchTST model, namely DualPathTST, is developed, where a de-stationary attention mechanism is employed into the original pathway, while a convolutional pathway is designed to capture the cross-variable dependencies. Then, a gated fusion mechanism is introduced to reconcile information from two pathways, which would dynamically integrate internal relationships and cross-variable dependencies. To effectively mitigate internal covariate shifts and enhance the model’s stability, a batch normalization layer is strategically incorporated into the output module. Extensive testing on multiple datasets, including ETT, Exchanges, Electricity, Weather, and Traffic, has conclusively demonstrated that DualPathTST not only significantly outperforms state-of-the-art models, with an average improvement of 4.6%, but also provides an innovative solution to the problem of capturing dependencies across variables in TSF models.

Event relation extraction plays a crucial role in constructing an event knowledge graph. However, current models only extract trigger words as event ontology representations, and do not consider node type during information aggregation, resulting in low accuracy in event relation extraction. To address these challenges, we propose an event relation extraction model based on heterogeneous graph attention networks and event ontology direction induction. To enhance the completeness of event information, we incorporate argument role information, in addition to trigger words, into the input text. A novel heterogeneous graph attention framework is proposed to reasonably allocate weights to trigger words, argument roles, and text information, and then perform two levels of aggregation, node-level and semantic-level, in sequence. To improve the accuracy of event direction discrimination, we construct an event ontology subgraph that includes trigger words and arguments to aggregate complete event structure information during direction induction. Finally, we evaluate our model on three datasets, TimeBank-Dense, MATRES, and HiEve, and demonstrate that our model outperforms state-of-the-art models by 1.2%, 0.5%, and 0.8%, respectively, in terms of the Micro-F1 score. Our proposed model provides a promising solution for event relation extraction and can be applied in various natural language processing applications.