Compositional Zero-Shot Learning (CZSL) aims at recognizing novel compositions by combining concepts learned from seen compositions. The key to tackle CZSL is disentangling highly coupled attribute-object compositions and learning exclusive concepts. Previous works mainly design networks to learn visual concepts from top-layer representations provided by visual backbones. As visual backbones progressively integrate information layer by layer, some low-level but critical information for concept learning may be lost, and the coupling between attribute and object features deepens. To address these issues, we propose to extract multi-scale visual features and fuse them in an adaptive way by Mixture of Experts (MoE) networks. We also employ feature-level similarity and a maximum entropy regularization term to constrain the model to effectively disentangle and learn concepts from multi-scale visual information. Comprehensive experiments on three CZSL benchmark datasets demonstrate that our method significantly outperforms previous SOTA methods in both closed-world and open-world settings.

\beginabstract Dynamic graphs play an important role in many fields like social relationship analysis, recommender systems and medical science, as graphs evolve over time. It is fundamental to capture the evolution patterns for dynamic graphs. Existing works mostly focus on constraining the temporal smoothness between neighbor snapshots, however, fail to capture sharp shifts, which can be beneficial for graph dynamics embedding. To solve it, we assume the evolution of dynamic graph nodes can be split into temporal shift embedding and temporal consistency embedding. Thus, we propose the Self-supervised Temporal-aware Dynamic Graph representation Learning framework (STDGL) for disentangling the temporal shift embedding from temporal consistency embedding via a well-designed auxiliary task from the perspectives of both node local and global connectivity modeling in a self-supervised manner, further enhancing the learning of interpretable graph representations and improving the performance of various downstream tasks. Extensive experiments on link prediction, edge classification and node classification tasks demonstrate STDGL successfully learns the disentangled temporal shift and consistency representations. Furthermore, the results indicate significant improvements in our STDGL over the state-of-the-art methods, and appealing interpretability and transferability owing to the disentangled node representations. \endabstract

The maximum clique problem (MCP) has various applications to reveal the structure and function of graphs. Graphs are constantly updated in the real life. However, no algorithm is specifically designed for dynamic graph. AlthoughMCPin dynamic graphs can be solved by simply invoking a state-of-the-art static approach, such asPMC, when the graph is updated, such an approach of simply re-calculating from scratch is inefficient. The key issue withMCPalgorithm is to find a large clique, namely aseed, as fast as possible. Thus, search space can be pruned based on the seed. Size of the seed greedily found byPMCcannot be guaranteed, as it fluctuates considerably. Moreover, the time required to find a seed underPMCis up to$O(| E| \cdot \Delta (G))$, where$\Delta (G)$is the highest degree inG. In this article, we intend to find a sizable seed by updating the previous maximum clique with the incident vertices of the inserted/deleted edge. Size of the seed now is guaranteed to be no less than$\omega (G^{\prime})\; - \;1$, where$\omega (G^{\prime})$is the size of the maximum clique on the updated graph. Moreover, the seed can be found in a time complexity of$O(\Delta (G)^{2})$. Two other crucial issues related to theMCPin dynamic graphs are refreshing rate and refreshing overhead. After a tight upper bound is imposed on$\omega (G^{\prime})$, the necessity of refreshing is evaluated by comparing the seed with its largest challenger, then unnecessary refreshing is wiped out effectively. The size of the largest challenger is judiciously estimated using a lazy growth strategy. Subsequently, the search space in refreshing is confined on a much smaller subgraph using a local refreshing strategy. Extensive experiments indicate that the proposed approach outperforms the baseline algorithm by approximately one order of magnitude.

Duplicate Question Identification (DQI) improves the processing efficiency and accuracy of large-scale community question answering and automatic QA system. The purpose of DQI task is to identify whether the paired questions are semantically equivalent. However, how to distinguish the synonyms or homonyms in paired questions is still challenging. Most previous works focus on the word-level or phrase-level semantic differences. We firstly propose to explore the asking emphasis of a question as a key factor in DQI. Asking emphasis bridges semantic equivalence between two questions. In this paper, we propose an attention model with multi-fusion asking emphasis (MFAE) for DQI. At first, BERT is used to obtain the dynamic pre-trained word embeddings. Then we get inter- and intra-asking emphasis by summing inter-attention and self-attention, respectively; the idea is that, the more a word interacts with others, the more important the word is. Finally, we use eight-way combinations to generate multi-fusion asking emphasis and multi-fusion word representation. Experimental results demonstrate that our model achieves state-of-the-art performance on both Quora Question Pairs and CQADupStack data. In addition, our model can also improve the results for natural language inference task on SNLI and MultiNLI datasets. The code is available at https://github.com/rzhangpku/MFAE.