Dynamic link prediction has become a trending research subject because of its wide applications in the web, sociology, transportation, and bioinformatics. Currently, the prevailing approach for dynamic link prediction is based on graph neural networks, in which graph representation learning is the key to perform dynamic link prediction tasks. However, there are still great challenges because the structure of graphs evolves over time. A common approach is to represent a dynamic graph as a collection of discrete snapshots, in which information over a period is aggregated through summation or averaging. This way results in some fine-grained time-related information loss, which further leads to a certain degree of performance degradation. We conjecture that such fine-grained information is vital because it implies specific behavior patterns of nodes and edges in a snapshot. To verify this conjecture, we propose a novel fine-grained behavior-aware network (BehaviorNet) for dynamic network link prediction. Specifically, BehaviorNet adapts a transformer-based graph convolution network to capture the latent structural representations of nodes by adding edge behaviors as an additional attribute of edges. GRU is applied to learn the temporal features of given snapshots of a dynamic network by utilizing node behaviors as auxiliary information. Extensive experiments are conducted on several real-world dynamic graph datasets, and the results show significant performance gains for BehaviorNet over several state-of-the-art (SOTA) discrete dynamic link prediction baselines. Ablation study validates the effectiveness of modeling fine-grained edge and node behaviors.

Current state of handwritten Chinese character recognition (HCCR) conducted on well-confined character set, far from meeting industrial requirements. The paper describes the creation of a large-scale handwritten Chinese character database. Constructing the database is an effort to scale up Chinese handwritten character classification task to cover the full list of GBK character set specification. It consists of 21-thousand Chinese character categories and 20-million character images, larger than previous databases both in scale and diversity. We present solutions to the challenges of collecting and annotating such large-scale handwritten character samples. We elaborately design the sampling strategy, extract salient signals in a systematic way, annotate the tremendous characters through three distinct stages. Experiments are conducted the generalization to other handwritten character databases and our database demonstrates great values. Surely, its scale opens unprecedented opportunities both in evaluation of character recognition algorithms and in developing new techniques.

The prediction phase is used to interact with end users, so its response speed is critical for a good user experience to large category recognition tasks. This paper presents a novel and fast algorithm for prototype prediction which may solve the current computing challenges in character input applications on smart terminals. We construct a social network for prototypes and their pair-wise connections. Such "prototype network" falls into "scale-free network", emerging "small-world effect". Under reasonable conditions, there exists a small geodesic path between each node pair. This feature guarantees us to search "better" nodes following the directed edges. Unfortunately, the naive search strategy results in a computing complexity of exponential order. To convert the problem to a manageable scale, we enhance the generic breadth-first search with greedy selectivity. As a result, our method just consider a small candidate set and further propagate from those seeds. Thorough analysis both on network structure and algorithmic propagation patterns is conducted and advantages in efficiency and practicality are revealed. Finally, we evaluated the proposed algorithm on a large-scale, large-category handwritten Chinese character recognition task. Experimental results show that the proposed algorithm can be tuned with either faster prediction speed or higher prediction accuracy.

Recurrent neural networks (RNNs) have been widely used in many sequential labeling fields. Decades of research fruits show that artificial neuron as the building blocks plays great role in its success. Different RNN neurons are proposed, such as long-short term memory (LSTM) and gated recurrent unit (GRU), and used in most applications let alone character recognition, to encode the long-term contextual dependencies. Inspired by both LSTM and GRU, a new structure named gated memory unit (GMU) is presented which carries forward their merits. GMU preserves the constant error carousels (CEC) which is devoted to enhance a smooth information flow. GMU also lends both the cell structure of LSTM and the interpolation gates of GRU. The proposed neuron is evaluated on both online English handwriting recognition and online Chinese handwriting recognition tasks in terms of parameter volumes, convergence and accuracy. The results show that GMU is of potential choice in handwriting recognition tasks.

The HMM-based segmentation-free strategy for Chinese handwriting recognition has the merit that the model parameters can be trained with text line samples without annotation of character boundaries. However, the recognition performance has been limited to the general maximum likelihood estimation framework. In this paper, we investigate the discriminative training framework based on MPE/MWE criteria in the context of Chinese handwriting recognition for the first time. It optimizes a objective function that is a smooth measure of recognition error. Then EBW procedure is used to solve such criteria. Some key issues for robust MPE/MWE training are explored. We reveal that MPE/MWE requires more training samples, however, Chinese handwriting recognition poses severe data sparsity problem. We explore the sample synthesizing to help the training process. Experiments are conducted on Chinese handwriting database and the effectiveness of MPE/MWE training is manifested. In particular, at least 28% error reduction of recognition rates is observed in MPE/MWE training with 50 copies of synthetic sample when big ram is used to approximate the language model.