The multimodal knowledge graph reasoning (MKGR) task aims to predict the missing facts in the incomplete MKGs by leveraging auxiliary images and descriptions of entities. Existing approaches are trained with single-target objectives, which neglect the probabilistic correlations of entity labels, especially in non-target entities. Moreover, previous studies incorporate all modalities statically or adaptively, overlooking the negative impacts of irrelevant or misleading information in the incompetent modalities. To address these issues, we introduce a novel Reinforced Multimodal Distillation framework, exploiting the Dark Side of Modalities (DSoM) from two perspectives: (1) Dark knowledge from non-target entities: We propose to train a unimodal KGR model through logit distillation to mimic the multimodal soft labels provided by pre-trained multimodal teacher models. The multimodal soft labels could provide rich supervision signals with subtle correlations among both target and non-target entities from multiple perspectives. We further decouple logits into neighbor entities and non-neighbor entities to divide into two types of correlations. (2) Dark side in unhelpful modalities: To exclude the adverse effects of unhelpful modalities, we introduce a reinforced teacher combination mechanism that dynamically selects the optimal set of multimodal teachers for each triple. The agent is trained to maximize the rewards, which are only assigned to the beneficial multimodal combination strategies for the student model. Comprehensive experiments demonstrate the effectiveness of DSoM framework on 5 MKGR datasets. Codes are available at github.com/OreOZhao/DSoM.

The total electron content (TEC) of ionosphere at low latitudes is significantly influenced by solar-geomagnetic activity and seasonal variations. Traditional ionospheric models often struggle to accurately forecast TEC in low latitudes, which limits the improvement of positioning accuracy for single-frequency GNSS (Global Navigation Satellite System) receivers. This study focuses on the Sichuan and Yunnan areas of China that locate on the northern crest of the equatorial ionization anomaly, utilizing data from 48 stations of the Chinese GNSS network. It employs a Convolutional Long Short-Term Memory network with multi-channel characteristics, combined with the Whale Optimization Algorithm, to construct a WOA-CNN-LSTM model for predicting TEC variations. The results demonstrate that the WOA-CNN-LSTM model outperforms Convolutional Neural Network-Gated Recurrent Unit (CNN-GRU), Bidirectional Long Short-Term Memory (BiLSTM), and Recurrent Neural Network (RNN) models in spatial morphology. In 2015 (a year with geomagnetic storms), the root mean square error (RMSE) values for all four seasons are at or below 1.96TECu, with mean absolute error (MAE) values at or below 1.42TECu, and Pearson correlation coefficients at or above 0.98. In 2019 (a calm year), the RMSE values are all below 0.74TECu, MAE values are all below 0.54TECu, and Pearson correlation coefficients remain at or above 0.95. In terms of temporal variation, the RMSE prediction results for the four observation stations are all at or below 2.75TECu for each of the four seasons in 2015, improving to 1.56TECu in 2019. Therefore, this model significantly enhances ionospheric prediction accuracy in low-latitude regions, benefiting navigation positioning, space environment forecasting, and disaster early warning systems.