Abstract Terrain generation is a challenging problem. Procedural modelling methods lack control, while machine learning methods often need large training datasets and struggle to preserve the topology information. We propose a method that generates a new terrain from a single image for training and a simple user sketch. Our single‐shot method preserves the sketch topology while generating diversified results. Our method is based on a graph neural network (GNN) and builds a detailed relation among the sketch‐extracted features, that is, ridges and valleys and their neighbouring area. By disentangling the influence from different sketches, our model generates visually realistic terrains following the user sketch while preserving the features from the real terrains. Experiments are conducted to show both qualitative and quantitative comparisons. The structural similarity index measure of our generated and real terrains is around 0.8 on average.

Human motion segmentation (HMS) aims to segment a long human action video into a bunch of short and meaningful action clips. Existing supervised learning approaches need a large amount of training data which may be costly in real-world scenario, while most unsupervised clustering methods cannot fully explore the temporal correlations among human motions and hard to achieve promising performances. In our paper, we design a novel unsupervised framework, called Velocity-Sensitive Dual-Side Auto-Encoder (VSDA), for HMS task. Specifically, a multi-neighbor auto-encoder (MNA) is proposed to extract informative temporal features, which fully explores the local temporal patterns of human motions. In addition, a long-short distance encoding (LSE) strategy is designed. It constrains the encoded representations of close (short-distance) frames becoming similar while the representations of far-away (long-distance) frames becoming distinctive. Similarly, this strategy is also deployed on the decoded outputs as the long-short distance decoding (LSD) module. The LSE/LSD guides the learning process explicitly and implicitly to achieve the dual-side structure. Moreover, we consider the energy variations during the human motion to propose the velocity-sensitive (VS) guidance mechanism for further model improvement. VSDA leverages the temporal characteristics of human motion and derives promising HMS performance. Comprehensive experiments on six real-world human motion datasets illustrate the effectiveness of our proposed model.

Dynamic systems that consist of a set of interacting elements can be abstracted as temporal networks. Recently, higher-order patterns that involve multiple interacting nodes have been found crucial to indicate domain-specific laws of different temporal networks. This posts us the challenge of designing more sophisticated hypergraph models for these higher-order patterns and the associated new learning algorithms. Here, we propose the first model, named HIT, for full-spectrum higher-order pattern prediction in temporal hypergraphs. Particularly, we focus on predicting three types of common but important interaction patterns involving three interacting elements in temporal networks, which could be extended to even higher-order patterns. HIT extracts the structural representation of a node triplet of interest on the temporal hypergraph and uses it to tell what type of, when, and why the interaction expansion could happen in this triplet. HIT could achieve significant improvement (averaged 20% AUC gain to identify the interaction type, uniformly more accurate time estimation) compared to both heuristic and other neural-network-based baselines on 5 real-world large temporal hypergraphs. Moreover, HIT provides a certain degree of interpretability by identifying the most discriminatory structural features on the temporal hypergraphs for predicting different higher-order patterns.

Aspect-based sentiment classification aims to predict sentimental polarities of one or multiple aspects in texts. As texts always contain a large proportion of task-irrelevant words, accurate alignment between aspects and their sentimental descriptions is the most crucial and challenging step. State-of-the-art approaches are mainly based on word-level attention learned from recurrent neural network variants (e.g., LSTM) or graph neural networks. From another view, these methods essentially weight and aggregate all possible alignments. However, this mechanism heavily relies on large-scale supervision training: without enough labels, it could easily overfit with difficulty in generalization. To address this challenge, we propose SentRL, a reinforcement learning-based framework for aspect-based sentiment classification. In this framework, input texts are transformed into their dependency graphs. Then, an agent is deployed to walk on the graphs, explores paths from target aspect nodes to their potential sentimental regions, and differentiates the effectiveness of different paths. By limiting the agent’s exploration budget, our method encourages the agent to skip task-irrelevant information and focus on the most effective paths for alignment purpose. Our method considerably reduces the impact of task-irrelevant words and improves generalization performance. Compared with competitive baseline methods, our approach achieves the highest performance on public benchmark datasets with up to 3.7% improvement.

In a real-world scenario, an object could contain multiple tags instead of a single categorical label. To this end, multi-label learning (MLL) emerged. In MLL, the feature distributions are long-tailed and the complex semantic label relation and the long-tailed training samples are the main challenges. Semi-supervised learning is a potential solution. While, existing methods are mainly designed for single class scenario while ignoring the latent label relations. In addition, they cannot well handle the distribution shift commonly existing across source and target domains. To this end, a Semi-supervised Dual Relation Learning (SDRL) framework for multi-label classification is proposed. SDRL utilizes a few labeled samples as well as large scale unlabeled samples in the training stage. It jointly explores the inter-instance feature-level relation and the intra-instance label-level relation even from the unlabeled samples. In our model, a dual-classifier structure is deployed to obtain domain invariant representations. The prediction results from the classifiers are further compared and the most confident predictions are extracted as pseudo labels. A trainable label relation tensor is designed to explicitly explore the pairwise latent label relations and refine the predicted labels. SDRL is able to effectively and efficiently explore the feature-label relation as well as the label-label relation knowledge without any extra semantic knowledge. We evaluated SDRL in general and zero-shot multi-label classification tasks and we concluded that SDRL is superior to other SOTA baselines. Furthermore, extensive ablation studies have been done which reveal the effectiveness of each component in our framework.

Multi-label learning (MLL) solves the problem that one single sample corresponds to multiple labels. It is a challenging task due to the long-tail label distribution and the sophisticated label relations. Semi-supervised MLL methods utilize a small-scale labeled samples and large-scale unlabeled samples to enhance the performance. However, these approaches mainly focus on exploring the data distribution in feature space while ignoring mining the label relation inside of each instance. To this end, we proposed a Dual Relation Semi-supervised Multi-label Learning (DRML) approach which jointly explores the feature distribution and the label relation simultaneously. A dual-classifier domain adaptation strategy is proposed to align features while generating pseudo labels to improve learning performance. A relation network is proposed to explore the relation knowledge. As a result, DRML effectively explores the feature-label and label-label relations in both labeled and unlabeled samples. It is an end-to-end model without any extra knowledge. Extensive experiments illustrate the effectiveness and efficiency of our method1.

High-dimensional time series segmentation aims to segment a long temporal sequence into several short and meaningful subsequences. The high-dimensionality makes it challenging due to the complicated correlations among the sequential features. A large number of labeled data is required in existing supervised methods, and unsupervised methods mainly deploy clustering approaches, which are sensitive to outliers and hard to guarantee high performance. Also, most existing methods mainly rely on hand-craft features to deal with regular time series segmentation and achieve promising results. However, these approaches cannot effectively handle high-dimensional time series and will result in a high computational cost. In our work, we propose a novel unsupervised representation learning framework called Dual-Side Auto-Encoder (DSAE). It mainly focuses on high-dimensional time series segmentation by effectively capturing the temporal correlative patterns. Specifically, a single-to-multiple auto-encoder is designed to capture local sequential information. Besides, a long-shot distance encoding strategy is proposed. It aims to explicitly guide the learning process to obtain distinctive representations for segmentation. Furthermore, the long-short distance strategy is also executed in the decoded feature space, which implicitly directs the representation learning. Substantial experiments on six datasets illustrate the model effectiveness.

The password has become today's dominant method of authentication. While brute-force attack methods such as HashCat and John the Ripper have proven unpractical, the research then switches to password guessing. State-of-the-art approaches such as the Markov Model and probabilistic context-free grammar (PCFG) are all based on statistical probability. These approaches require a large amount of calculation, which is time-consuming. Neural networks have proven more accurate and practical in password guessing than traditional methods. However, a raw neural network model is not qualified for cross-site attacks because each dataset has its own features. Our work aims to generalize those leaked passwords and improves the performance in cross-site attacks. In this paper, we propose GENPass, a multi-source deep learning model for generating “general” password. GENPass learns from several datasets and ensures the output wordlist can maintain high accuracy for different datasets using adversarial generation. The password generator of GENPass is PCFG+LSTM (PL). We are the first to combine a neural network with PCFG. Compared with Long short-term memory (LSTM), PL increases the matching rate by 16%-30% in cross-site tests when learning from a single dataset. GENPass uses several PL models to learn datasets and generate passwords. The results demonstrate that the matching rate of GENPass is 20% higher than by simply mixing datasets in the cross-site test. Furthermore, we propose GENPass with probability (GENPass-pro), the updated version of GENPass, which can further increase the matching rate of GENPass.

Multi-view action recognition targets to integrate complementary information from different views to improve classification performance. It is a challenging task due to the distinct gap between heterogeneous feature domains. Moreover, most existing methods neglect to consider the incomplete multi-view data, which limits their potential compatibility in real-world applications. In this work, we propose a Generative Multi-View Action Recognition (GMVAR) framework to address the challenges above. The adversarial generative network is leveraged to generate one view conditioning on the other view, which fully explores the latent connections in both intra-view and cross-view aspects. Our approach enhances the model robustness by employing adversarial training, and naturally handles the incomplete view case by imputing the missing data. Moreover, an effective View Correlation Discovery Network (VCDN) is proposed to further fuse the multi-view information in a higher-level label space. Extensive experiments demonstrate the effectiveness of our proposed approach by comparing with state-of-the-art algorithms.

Password has become today's dominant method of authentication in social network. While the brute-force attack methods, such as HashCat and John the Ripper, are unpractical, the research then switches to the password guess. The state-of-the-art approaches, such as Markov Model and probabilistic context-free grammars(PCFG), are all based on statistical probability. These approaches have a low matching rate. The methods on neural network have been proved more accurate and practical for password guessing than traditional methods. However, a raw neural network model is not qualified for cross-sites attack since each data set has its own features. This paper proposes a general deep learning model for password guessing, called GENPass. GENPass can learn features from several data sets and ensure the output wordlist high accuracy in different data sets by using adversarial generation. The password generator of GENPass is PCFG+LSTM(PL), where LSTM is a kind of Recurrent Neural Network. We combine neural network with PCFG because we found people were used to set their passwords with meaningful strings. Compared with LSTM, PL increased the matching rate by 16%-30% in the cross-sites tests when learning from a single dataset. GENPass uses several PL models to learn datasets and generate passwords. The result shows that the matching rate of GENPass is 20% higher than that of simply mixing those datasets in the cross-sites test.