Stable and reliable operation of high-speed railway requires continuous and reliable power supply of traction substation, and isolating switch is one of the most important electrical devices in high-speed railway traction substation. In this paper, to address the issue of isolating switch accurate localization and state recognition simultaneously, we present an automatic isolating switch segmentation and state recognition framework called ISSSR-Net using multitask learning that consists of two stages. First, an isolating switch segmentation network called ISS-Net is proposed for isolating switch pixel-level segmentation precisely, of which a new structure containing a strip pooling module, a channel attention and three pyramid pooling modules is designed to greatly improve the segmentation and recognition performance, even in complex conditions such as rain, snow and fog. Second, to improve state recognition accuracy, the segmentation map yielded from the ISS-Net and the feature map from the shared backbone are together fed into isolating switch recognition network called ISR-Net to recognize its three states. In addition, a global context block is integrated into ISR-Net to further improve state recognition accuracy. Extensive experimental results on a self-collected dataset from Heishan traction substation corroborate that this paper provides an effective and robust method to achieve isolating switch segmentation and state recognition simultaneously. The MIoUand${F}~1$-score of segmentation reach 0.93 and 0.94 respectively, which is better than U-Net and its variants, and other SOTA. The${F}~1$-score of recognition reach 1.00, which is higher than HOG+SVM and other deep learning methods that have been experimented.

Non-parallel text style transfer aims to rephrase a sentence with another style while reserving its content relying on non-parallel data. Most existing methods can be divided into two groups: 1) separating content and style of the input text and 2) directly modeling the style transfer process. To the best of our knowledge, all these existing works lack fine-grained supervisory signals during training, which leads to difficulty in achieving a good balance between content preservation and style satisfaction. However, the study on supervisors which could provide fine-grained supervisory signals for training transfer models has received relatively less attention. Thus, we propose a self-attentional discriminator and a training strategy for training an attentional transfer model by leveraging the fine-grained supervisory signals from the proposed discriminator. Specifically, our discriminator provides token-wise style/content weights by performing self-attention between the sentence vector and the token embeddings, which forms a shortcut in back-propagation leading to more accurate gradients. The style/content weights also pose a better content alignment constraint and improve the interpretability of the training procedure by identifying stylized tokens. The training of our transfer model is end-to-end via Gumbel-Softmax with the pre-trained discriminator. Experiments on two datasets with automatic and human evaluations as well as theoretical and empirical analysis demonstrate the effectiveness of our method.1

The popularity of online social coding (SC) platforms such as GitHub is growing due to their social functionalities and tremendous support during the product development lifecycle. The rich information of experts' contributions on repositories can be leveraged to recruit experts for new/existing projects. In this paper, we define the problem of collaborative experts finding in SC platforms. Given a project, we model an SC platform as an attributed heterogeneous network, learn latent representations of network entities in an end-to-end manner and utilize them to discover collaborative experts to complete a project. Extensive experiments on real-world datasets from GitHub indicate the superiority of the proposed approach over the state-of-the-art in terms of a range of performance measures.

Learning the latent representations of entities based on their relationships and the data associated with them is an essential task in many applications such as ranking, recommendation systems, graph-based team formation, keyword search, and many more. However, the majority of existing techniques learn the latent representations of either network or textual data. Structural embedding techniques suffer from the sparsity of real-world networks. Attributes of nodes are a source of rich information to ameliorate network embedding vectors which are overlooked in the literature. Thus, most existing network representation learning tools capture structural information. This paper introduces an open-source toolkit called OpenAttHetRL to learn the latent representations of entities based on their both network and textual data in an end-to-end fashion. OpenAttHetRL is easy to employ and adapt for a variety of tasks including ranking, recommendation systems, and expert finding. OpenAttHetRL aims to provide a unified toolkit for data pre-processing, building and training models, and performing predictions for a downstream task. It employs a graph convolution network to capture the relationships among entities and a kernel pooling technique to preserve the similarity of their textual data in the embedding space. We use expert finding in community question answering systems to demonstrate how OpenAttHetRL can be trained to get latent representations of questions, their askers, tags, and answerers and find potential answerers of new questions.

Cross-domain recommendation aims to leverage the rich interaction information in the source domain to predict interactions between cold-start users and items in the target domain. Since reviews contain users' preferences and items' attributes, many review-based cross-domain recommendation methods are proposed. However, existing methods cannot either 1) select important words and reviews from multiple reviews of users/items, or 2) learn a unified representation space for different domains without enough overlapping users. To address these problems, we propose a Hierarchical Attention model trained with Random Back-Transfer for cross-domain recommendation (HARBT). Specifically, the hierarchical attention extracts text information related to a given user or item which leads to an accurate interaction prediction. The random back-transfer works as a data augmentation algorithm to utilize data of users and items which are in the same domain for better matching of representations in different domains. Extensive experiments on real-world datasets show that our approach outperforms state-of-the-art methods significantly.