Due to the heterogeneity of nodes and edges, heterogeneous network embedding is a very challenging task to embed highly coupled networks into a set of low-dimensional vectors. Existing models either only learn embedding vectors for nodes or only for edges. These two methods of embedding learning are rarely performed in the same model, and they both overlook the internal correlation between nodes and edges. To solve these problems, a node and edge joint embedding model is proposed for Heterogeneous Information Networks (HINs), called NEJE. The NEJE model can better capture the latent structural and semantic information from an HIN through two joint learning strategies: type-level joint learning and element-level joint learning. Firstly, node-type-aware structure learning and edge-type-aware semantic learning are sequentially performed on the original network and its line graph to get the initial embedding of nodes and the embedding of edges. Then, to optimize performance, type-level joint learning is performed through the alternating training of node embedding on the original network and edge embedding on the line graph. Finally, a new homogeneous network is constructed from the original heterogeneous network, and the graph attention model is further used on the new network to perform element-level joint learning. Experiments on three tasks and five public datasets show that our NEJE model performance improves by about 2.83% over other models, and even improves by 6.42% on average for the node clustering task on Digital Bibliography & Library Project (DBLP) dataset.
- Article type
- Year
- Co-author
![](/assets/img/article/article-lock.png)
![](/assets/img/article/article-lock.png)
Metapaths with specific complex semantics are critical to learning diverse semantic and structural information of heterogeneous networks (HNs) for most of the existing representation learning models. However, any metapaths consisting of multiple, simple metarelations must be driven by domain experts. These sensitive, expensive, and limited metapaths severely reduce the flexibility and scalability of the existing models. A metapath-free, scalable representation learning model, called Metarelation2vec, is proposed for HNs with biased joint learning of all metarelations in a bid to address this problem. Specifically, a metarelation-aware, biased walk strategy is first designed to obtain better training samples by using autogenerating cooperation probabilities for all metarelations rather than using expert-given metapaths. Thereafter, grouped nodes by the type, a common and shallow skip-gram model is used to separately learn structural proximity for each node type. Next, grouped links by the type, a novel and shallow model is used to separately learn the semantic proximity for each link type. Finally, supervised by the cooperation probabilities of all meta-words, the biased training samples are thrown into the shallow models to jointly learn the structural and semantic information in the HNs, ensuring the accuracy and scalability of the models. Extensive experimental results on three tasks and four open datasets demonstrate the advantages of our proposed model.
![](/assets/img/article/article-lock.png)
An integrated observer framework based mechanical parameters identification approach for adaptive control of permanent magnet synchronous motors is proposed in this paper. Firstly, an integrated observer framework is established for mechanical parameters’ estimation, which consists of an extended sliding mode observer (ESMO) and a Luenberger observer. Aiming at minimizing the influence of parameters coupling, the viscous friction and the moment of inertia are obtained by ESMO and the load torque is identified by Luenberger observer separately. After obtaining estimates of the mechanical parameters, the optimal proportional integral (PI) parameters of the speed-loop are determined according to third-order best design method. As a result, the controller can adjust the PI parameters in real time according to the parameter changes to realize the adaptive control of the system. Meanwhile, the disturbance is compensated according to the estimates. Finally, the experiments were carried out on simulation platform, and the experimental results validated the reliability of parameter identification and the efficiency of the adaptive control strategy presented in this paper.
![](/assets/img/article/article-lock.png)
Accurate wind speed prediction has been becoming an indispensable technology in system security, wind energy utilization, and power grid dispatching in recent years. However, it is an arduous task to predict wind speed due to its variable and random characteristics. For the objective to enhance the performance of forecasting short-term wind speed, this work puts forward a hybrid deep learning model mixing time series decomposition algorithm and gated recurrent unit (GRU). The time series decomposition algorithm combines the following two parts: (1) the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and (2) wavelet packet decomposition (WPD). Firstly, the normalized wind speed time series (WSTS) are handled by CEEMDAN to gain pure fixed-frequency components and a residual signal. The WPD algorithm conducts the second-order decomposition to the first component that contains complex and high frequency signal of raw WSTS. Finally, GRU networks are established for all the relevant components of the signals, and the predicted wind speeds are obtained by superimposing the prediction of each component. Results from two case studies, adopting wind data from laboratory and wind farm, respectively, suggest that the related trend of the WSTS can be separated effectively by the proposed time series decomposition algorithm, and the accuracy of short-time wind speed prediction can be heightened significantly mixing the time series decomposition algorithm and GRU networks.