Search

With the rapid development of transportation systems, traffic data have been largely produced in daily lives. Finding the insights of all these complex data is of great significance to vehicle dispatching and public safety. In this work, we propose a multitask deep learning model called Multitask Recurrent Graph Convolutional Network (MRGCN) for accurately predicting traffic flows in the city. Specifically, we design a multitask framework consisting of four components: a region-flow encoder for modeling region-flow dynamics, a transition-flow encoder for exploring transition-flow correlations, a context modeling component for contextualized fusion of two types of traffic flows and a task-specific decoder for predicting traffic flows. Particularly, we introduce Dual-attention Graph Convolutional Gated Recurrent Units (DGCGRU) to simultaneously capture spatial and temporal dependencies, which integrate graph convolution and recurrent model as a whole. Extensive experiments are carried out on two real-world datasets and the results demonstrate that our proposed method outperforms several existing approaches.

Destination prediction is known as an important problem for many location based services (LBSs). Existing solutions generally apply probabilistic models or neural network models to predict destinations over a subtrajectory, and adopt the standard attention mechanism to improve the prediction accuracy. However, the standard attention mechanism uses fixed feature representations, and has a limited ability to represent distinct features of locations. Besides, existing methods rarely take the impact of spatial and temporal characteristics of the trajectory into account. Their accuracies in fine-granularity prediction are always not satisfactory due to the data sparsity problem. Thus, in this paper, a carefully designed deep learning model called LATL model is presented. It not only adopts an adaptive attention network to model the distinct features of locations, but also implements time gates and distance gates into the Long Short-Term Memory (LSTM) network to capture the spatial-temporal relation between consecutive locations. Furthermore, to better understand the mobility patterns in different spatial granularities, and explore the fusion of multi-granularity learning capability, a hierarchical model that utilizes tailored combination of different neural networks under multiple spatial granularities is further proposed. Extensive empirical studies verify that the newly proposed models perform effectively and settle the problem nicely.

With great value in real applications, sequential recommendation aims to recommend users the personalized sequential actions. To achieve better performance, it is essential to consider both long-term preferences and sequential patterns ( i . e ., short-term dynamics). Compared to widely used Recurrent Neural Network (RNN) and Convolutional Neural Network (CNN), Self-Attention Network (SAN) obtains a surge of interest due to fewer parameters, highly parallelizable computation, and flexibility in modeling dependencies. However, existing SAN-based models are inadequate in characterizing and distinguishing users’ long-term preferences and short-term demands since they do not emphasize the importance of the current interest and temporal order information of sequences. In this paper, we propose a novel multi-layer long- and short-term self-attention network (LSSA) for sequential recommendation. Specifically, we first split the entire sequence of a user into multiple sub-sequences according to the timespan. Then the first self-attention layer learns the user’s short-term dynamics based on the last sub-sequence, while the second one captures the user’s long-term preferences through the previous sub-sequences and the last one. Finally, we integrate the long- and short-term representations together to form the user’s final hybrid representation. We evaluate the proposed model on three real-world datasets, and our experimental results show that LSSA outperforms state-of-the-art methods with a wide margin.

Recent years have witnessed the increasing popularity of Location-based Social Network (LBSN) services, which provides unparalleled opportunities to build personalized Point-of-Interest (POI) recommender systems. Existing POI recommendation and location prediction tasks utilize past information for future recommendation or prediction from a single direction perspective, while the missing POI category identification task needs to utilize the check-in information both before and after the missing category. Therefore, a long-standing challenge is how to effectively identify the missing POI categories at any time in the real-world check-in data of mobile users. To this end, in this paper, we propose a novel neural network approach to identify the missing POI categories by integrating both bi-directional global non-personal transition patterns and personal preferences of users. Specifically, we delicately design an attention matching cell to model how well the check-in category information matches their non-personal transition patterns and personal preferences. Finally, we evaluate our model on two real-world datasets, which clearly validate its effectiveness compared with the state-of-the-art baselines. Furthermore, our model can be naturally extended to address next POI category recommendation and prediction tasks with competitive performance., Accepted by Neural Networks. arXiv admin note: text overlap with arXiv:2112.15285

Knowledge graph (KG) reasoning is a significant method for KG completion. To enhance the explainability of KG reasoning, some studies adopt reinforcement learning (RL) to complete the multi-hop reasoning. However, RL-based reasoning methods are severely limited by few-shot relations (only contain few triplets). To tackle the problem, recent studies introduce meta-learning into RL-based methods to improve reasoning performance. However, the generalization abilities of their models are limited due to the problem of low reasoning accuracies over hard relations (e.g., language and title). To overcome this problem, we propose a novel model called THML (Two-level Hardness-aware Meta-reinforcement Learning). Specifically, the model contains the following two components: (1) A hardness-aware meta-reinforcement learning method is proposed to predict the missing element by training hardness-aware batches. (2) A two-level hardness-aware sampling is proposed to effectively generate new hardness-aware batches from relation level and relation-cluster level. The generalization ability of our model is significantly improved by repeating the process of these two components in an alternate way. The experimental results demonstrate that THML notably outperforms the state-of-the-art approaches in few-shot scenarios.

Next Point-of-Interest (POI) recommendation is of great value for location-based services. Existing solutions mainly rely on extensive observed data and are brittle to users with few interactions. Unfortunately, the problem of few-shot next POI recommendation has not been well studied yet. In this paper, we propose a novel meta-optimized model MFNP, which can rapidly adapt to users with few check-in records. Towards the cold-start problem, it seamlessly integrates carefully designed user-specific and region-specific tasks in meta-learning, such that region-aware user preferences can be captured via a rational fusion of region-independent personal preferences and region-dependent crowd preferences. In modelling region-dependent crowd preferences, a cluster-based adaptive network is adopted to capture shared preferences from similar users for knowledge transfer. Experimental results on two real-world datasets show that our model outperforms the state-of-the-art methods on next POI recommendation for cold-start users.

Named Entity Recognition (NER) is a core task of NLP. State-of-art supervised NER models rely heavily on a large amount of high-quality annotated data, which is quite expensive to obtain. Various existing ways have been proposed to reduce the heavy reliance on large training data, but only with limited effect. In this paper, we propose a crowd-efficient learning approach for supervised NER learning by making full use of the online encyclopedia pages. In our approach, we first define three criteria (representativeness, informativeness, diversity) to help select a much smaller set of samples for crowd labeling. We then propose a data augmentation method, which could generate a lot more training data with the help of the structured knowledge of online encyclopedia to greatly augment the training effect. After conducting model training on the augmented sample set, we re-select some new samples for crowd labeling for model refinement. We perform the training and selection procedure iteratively until the model could not be further improved or the performance of the model meets our requirement. Our empirical study conducted on several real data collections shows that our approach could reduce 50% manual annotations with almost the same NER performance as the fully trained model.

Next basket recommendation with consideration of user sequential shopping behaviors plays a significant role in E-commerce to improve the user experience and service quality. Recently, recurrent neural networks (RNNs), especially attention-based RNN, have been widely adopted in the next basket recommendation. However, existing fixed attention mechanisms are not designed to model the dynamic and diverse characteristics of user appetites. In this paper, we propose an inter-basket and intra-basket adaptive attention network (IIAAN) for the next basket recommendation. Specifically, the inter-basket adaptive attention acts on all historical user baskets to model user's diverse long-term preferences, while the intra-basket adaptive attention is designed to act on item-level in the most recent basket to model user's dynamic and different short-term preferences. Then, we further integrate inter-basket and intra-basket adaptive attentions together to improve recommendation effectiveness. Finally, we evaluate the proposed model IIAAN using three real-world datasets from various E-commerce platforms. Our experimental results show that our model IIAAN significantly outperforms the state-of-the-art approaches for the next basket recommendation.

User behaviour on purchasing is always driven by complex latent factors, which are highly disentangled in the real world. Learning latent factorized representation of users can uncover user intentions behind the observed data (i.e. user-item interaction) and improve the robustness and interpretability of the recommender system. However, existing collaborative filtering methods learning disentangled representation face problems of balancing the trade-off between reconstruction quality and disentanglement. In this paper, we propose a controllable variational autoencoder framework for collaborative filtering. Specifically, we adopt a modified Proportional-Integral-Derivative (PID) control to the \(\beta \)-VAE objective to automatically tune the hyperparameter \(\beta \) using the output of Kullback-Leibler divergence as feedback. We further introduce item embeddings to guide the system to learn representation related to the real-world concepts using a factorized Gaussian distribution. Experimental results show that our model can get a crucial improvement over state-of-the-art baselines. We further evaluate our model’s effectiveness to control the trade-off between reconstruction error and disentanglement quality in the recommendation.

Service robots should be able to operate autonomously in dynamic and daily changing environments over an extended period of time. While Simultaneous Localization And Mapping (SLAM) is one of the most fundamental problems for robotic autonomy, most existing SLAM works are evaluated with data sequences that are recorded in a short period of time. In real-world deployment, there can be out-of-sight scene changes caused by both natural factors and human activities. For example, in home scenarios, most objects may be movable, replaceable or deformable, and the visual features of the same place may be significantly different in some successive days. Such out-of-sight dynamics pose great challenges to the robustness of pose estimation, and hence a robot's long-term deployment and operation. To differentiate the forementioned problem from the conventional works which are usually evaluated in a static setting in a single run, the term \textit{lifelong SLAM} is used here to address SLAM problems in an ever-changing environment over a long period of time. To accelerate lifelong SLAM research, we release the OpenLORIS-Scene datasets. The data are collected in real-world indoor scenes, for multiple times in each place to include scene changes in real life. We also design benchmarking metrics for lifelong SLAM, with which the robustness and accuracy of pose estimation are evaluated separately. The datasets and benchmark are available online at https://lifelong-robotic-vision.github.io/dataset/scene., Comment: To be published on ICRA 2020; 7 pages, 3 figures; v2 fixed a number in Table III

Repeat consumption is a common phenomenon in sequential recommendation tasks, where a user revisits or repurchases items that (s)he has interacted before. Previous researches have paid attention to repeat recommendation and made great achievements in this field. However, existing studies rarely considered the phenomenon that the consumers tend to show different behavior periodicities on different items, which is important for recommendation performance. In this paper, we propose a holistic model, which integrates Graph Convolutional Network with Periodic-Attenuated Self-Attention Network (GPASAN) to model user’s different behavior patterns for a better recommendation. Specifically, we first process all the users’ action sequences to construct a graph structure, which captures the complex item connection and obtains item representations. Then, we employ a periodic channel and an attenuated channel that incorporate temporal information into the self-attention mechanism to model the user’s periodic and novel behaviors, respectively. Extensive experiments conducted on three public datasets show that our proposed model outperforms the state-of-the-art methods consistently.

Cross-domain recommendation leverages the knowledge from relevant domains to alleviate the data sparsity issue. However, we find that the state-of-the-art cross-domain models are vulnerable to adversarial examples, leading to possibly large errors in generalization. That’s because most methods rarely consider the robustness of the proposed models. In this paper, we propose a new Adversarial Cross-Domain Network (ACDN), in which adversarial examples are dynamically generated to train the cross-domain recommendation model. Specifically, we first combine two multilayer perceptrons by sharing the user embedding matrix as our base model. Then, we add small but intentionally worst-case perturbations on the model embedding representations to construct adversarial examples, which can result in the model outputting an incorrect answer with a high confidence. By training with these aggressive examples, we are able to obtain a robust cross-domain model. Finally, we evaluate the proposed model on two large real-world datasets. Our experimental results show that our model significantly outperforms the state-of-the-art methods on cross-domain recommendation.

Attention mechanisms have been successfully applied in many fields, including sequential recommendation. Existing recommendation methods often use the deterministic attention network to consider latent user preferences as fixed points in low-dimensional spaces. However, the fixed-point representation is not sufficient to characterize the uncertainty of user preferences that prevails in recommender systems. In this paper, we propose a new Hierarchical Variational Attention Model (HVAM), which employs variational inference to model the uncertainty in sequential recommendation. Specifically, the attention vector is represented as density by imposing a Gaussian distribution rather than a fixed point in the latent feature space. The variance of the attention vector measures the uncertainty associated with the user’s preference representation. Furthermore, the user’s long-term and short-term preferences are captured through a hierarchical variational attention network. Finally, we evaluate the proposed model HVAM using two public real-world datasets. The experimental results demonstrate the superior performance of our model comparing to the state-of-the-art methods for sequential recommendation.

Factorization Machines (FMs) are a series of effective solutions for sparse data prediction by considering the interactions among users, items, and auxiliary information. However, the feature representations in most state-of-the-art FMs are fixed, which reduces the prediction performance as the same feature may have unequal predictabilities under different input instances. In this paper, we propose a novel Feature-adjusted Factorization Machine (FaFM) model by adaptively adjusting the feature vector representations from both vector-level and bit-level. Specifically, we adopt a fully connected layer to adaptively learn the weight of vector-level feature adjustment. And a user-item specific gate is designed to refine the vector in bit-level and to filter noises caused by over-adaptation of the input instance. Extensive experiments on two real-world datasets demonstrate the effectiveness of FaFM. Empirical results indicate that FaFM significantly outperforms the traditional FM with a 10.89% relative improvement in terms of Root Mean Square Error (RMSE) and consistently exceeds four state-of-the-art deep learning based models.

The prediction of traffic flow is of great importance to urban planning and intelligent transportation systems. Recently, deep learning models have been applied to study this problem. However, there still exist two main limitations: (1) They do not effectively model dynamic traffic patterns in irregular regions; (2) The traffic flow of a region is strongly correlated to the transition-flow between different regions, while this issue is largely ignored by existing approaches. To address these issues, we propose a multitask deep learning model called MTGCN for a more accurate traffic flow prediction. First, to process the input traffic network data, we propose using graph convolution in place of traditional grid-based convolution to model spatial dependencies between irregular regions. Second, as original graph convolution can not well respond to traffic dynamics, we design a novel attention mechanism to capture dynamic traffic patterns. At last, to obtain a more accurate prediction result, we integrate two correlated tasks which respectively predict two types of traffic flows (region-flow and transition-flow) as a whole, by combining the representations learned from each task in a rational way. We conduct extensive experiments on two real-world datasets and the results show that our proposed method achieves better performance compared with other baseline models.

Multi-label image classification has attracted considerable attention in machine learning recently. Active learning is widely used in multi-label learning because it can effectively reduce the human annotation workload required to construct high-performance classifiers. However, annotation by experts is costly, especially when the number of labels in a dataset is large. Inspired by the idea of semi-supervised learning, in this study, the authors propose a novel, semi-supervised multi-label active learning (SSMAL) method that combines automated annotation with human annotation to reduce the annotation workload associated with the active learning process. In SSMAL, they capture three aspects of potentially useful information – classification prediction information, label correlation information, and example spatial information – and they use this information to develop an effective strategy for automated annotation of selected unlabelled example-label pairs. The experimental results obtained in this study demonstrate the effectiveness of the authors' proposed approach.

Multilabel image classification has been a hot topic in the field of computer vision and image understanding in recent years. To achieve better classification performance with fewer labeled images, multilabel active learning is used for this scenario. Several active learning methods have been proposed for multilabel image classification. However, all of them assume that either all training images have complete labels or label correlations are given at the beginning. These two assumptions are unrealistic. In fact, it is very difficult to obtain complete labels for each example, in particular when the size of labels in a multilabel dataset is very large. Typically, only partial labels are available. This is one type of “weak label” problem. To solve this weak label problem inside multilabel active learning, this paper proposes a novel solution called AE-WLMAL. AE-WLMAL explores conditional label correlations on the weak label problem with the help of input features and then utilizes label correlations to construct a unified sampling strategy and evaluate the informativeness of each example-label pair in a multilabel dataset for active sampling. In addition, a pruning strategy is adopted to further improve its computation efficiency. Moreover, AE-WLAML exploits label correlations to infer labels for unlabeled images, which further reduces human labeling cost. Our experimental results on seven real-world datasets show that AE-WLMAL consistently outperforms existing approaches.

Machine-constructed knowledge bases often contain noisy and inaccurate facts. There exists significant work in developing automated algorithms for knowledge base refinement. Automated approaches improve the quality of knowledge bases but are far from perfect. In this paper, we leverage crowdsourcing to improve the quality of automatically extracted knowledge bases. As human labelling is costly, an important research challenge is how we can use limited human resources to maximize the quality improvement for a knowledge base. To address this problem, we first introduce a concept of semantic constraints that can be used to detect potential errors and do inference among candidate facts. Then, based on semantic constraints, we propose rank-based and graph-based algorithms for crowdsourced knowledge refining, which judiciously select the most beneficial candidate facts to conduct crowdsourcing and prune unnecessary questions. Our experiments show that our method improves the quality of knowledge bases significantly and outperforms state-of-the-art automatic methods under a reasonable crowdsourcing cost.

The time signals behind a user's historical behaviors are important for better inferring what she prefers to interact with at the next time. For the attention-based recommendation methods, relative position encoding and time intervals division are two common ways to model the time signal behind each behavior. They either only consider the relative position of each behavior in the behavior sequence, or process the continuous temporal features into discrete category features for subsequent tasks, which can hardly capture the dynamic preferences of a user. In addition, although the existing recommendation methods have considered both long-term preference and short-term preference, they ignore the fact that the long-term preference of a user may be multi-faceted, and it is difficult to learn a user's fine-grained short-term preference. In this paper, we propose a Dynamic Multi-faceted Fine-grained Preference model (DMFP), where the multi-hops attention mechanism and the feature-level attention mechanism together with a vertical convolution operation are adopted to capture users' multi-faceted long-term preference and fine-grained short-term preference, respectively. Therefore, DMFP can better support the next item recommendation. Extensive experiments on three real-world datasets illustrate that our model can improve the effectiveness of the recommendation compared with the state-of-the-art methods.

Question generation (QG) aims to create a fluent question from a passage and a target answer. State-of-the-art approaches are mainly based on encoder-decoder models to generate questions from the given passage and answer, which focus on using the information contained in a particular part of the passage for QG, but unaware of the clues hidden in other parts of the passage. Besides, the existing work on QG mainly focus on generating factoid questions, which are less suitable for generating non-factoid questions such as why-questions. In this paper, we propose to augment encoder-decoder framework with a pair-wise self-matching attention mechanism to dynamically collect inter-sentential evidence from the whole passage according to the current passage word and answer information. Besides, to let the model be more suitable for why-question generation, we also involve some causal features in the encoding process. Finally, to tackle the lack of why-question generation training data problem, we adopt a distant supervised method with an initial causal knowledge base to generate a large training data for why-question generation. Extensive experiments on several data sets show that our model significantly outperforms state-of-the-art question generation models not only on why-question generation tasks, but also on other types of question generation tasks.

The sequential recommendation, which models sequential behavioral patterns among users for the recommendation, plays a critical role in recommender systems. However, the state-of-the-art Recurrent Neural Networks (RNN) solutions rarely consider the non-linear feature interactions and non-monotone short-term sequential patterns, which are essential for user behavior modeling in sparse sequence data. In this paper, we propose a novel Recurrent Convolutional Neural Network model (RCNN). It not only utilizes the recurrent architecture of RNN to capture complex long-term dependencies, but also leverages the convolutional operation of Convolutional Neural Network (CNN) model to extract short-term sequential patterns among recurrent hidden states. Specifically, we first generate a hidden state at each time step with the recurrent layer. Then the recent hidden states are regarded as an “image”, and RCNN searches non-linear feature interactions and non-monotone local patterns via intra-step horizontal and inter-step vertical convolutional filters, respectively. Moreover, the output of convolutional filters and the hidden state are concatenated and fed into a fully-connected layer to generate the recommendation. Finally, we evaluate the proposed model using four real-world datasets from various application scenarios. The experimental results show that our model RCNN significantly outperforms the state-of-the-art approaches on sequential recommendation.

Due to the dynamic and evolutionary characteristics of user interests, sequential recommendation plays a significant role in recommender systems. A fundamental problem in the sequential recommendation is modeling dynamic user preference. Recurrent Neural Networks (RNNs) are widely adopted in the sequential recommendation, especially attention-based RNN becomes the state-of-the-art solution. However the existing fixed attention mechanism is insufficient to model the dynamic and evolutionary characteristics of user sequential preferences. In this work, we propose a novel solution, Adaptive Attention-Aware Gated Recurrent Unit (3AGRU), to learn adaptive user sequential representations for sequential recommendation. Specifically, we adopt an attention mechanism to adapt the representation of user sequential preference, and learn the interaction between steps and items from data. Moreover, in the first level of 3AGRU, we construct adaptive attention network to describe the relevance between input and the candidate item. In this way, a new input based on adaptive attention can reflect users’ diverse interests. Then, the second level of 3AGRU applies adaptive attention network to hidden state level to learn a deep user representation which is able to express diverse interests of the user. Finally, we evaluate the proposed model using three real-world datasets from various application scenarios. Our experimental results show that our model significantly outperforms the state-of-the-art approaches on sequential recommendation.

User preferences are dynamic and diverse in real world, while historical preference of a user may not be equally important as current preference when predicting future interests. As a result, learning the evolving user representation effectively becomes a critical problem in personalized recommendation. However, existing recommendation solutions often use a fixed user representation, which is not capable of modeling the complex interests of users. To this end, we propose a novel metric learning approach named Adaptive Collaborative Metric Learning (AdaCML) for recommendation. AdaCML employs a memory component and an attention mechanism to learn an adaptive user representation, which dynamically adapts to locally activated items. In this way, implicit relationships of user-item pairs can be better determined in the metric space and users’ interests can be modeled more accurately. Comprehensive experimental results demonstrate the effectiveness of AdaCML on two datasets, and show that AdaCML outperforms competitive baselines in terms of Precision, Recall, and Normalized Discounted Cumulative Gain (NDCG).

Entity alignment aims to find entities referring to the same real-world object across different knowledge graphs (KGs). Most existing works utilize the relations between entities contained in the relation triples with embedding-based approaches, but require a large number of training data. Some recent attempt works on using types of their attributes in attribute triples for measuring the similarity between entities across KGs. However, due to diverse expressions of attribute names and non-standard attribute values across different KGs, the information contained in attribute triples can not be fully used. To tackle the drawbacks of the existing efforts, we novelly propose an unsupervised entity alignment approach using both attribute triples and relation triples of KGs. Initially, we propose an interactive model to use attribute triples by performing entity alignment and attribute alignment alternately, which will generate a lot of high-quality aligned entity pairs. We then use these aligned entity pairs to train a relation embedding model such that we could use relation triples to further align the remaining entities. Lastly, we utilize a bivariate regression model to learn the respective weights of similarities measuring from the two aspects for a result combination. Our empirical study performed on several real-world datasets shows that our proposed method achieves significant improvements on entity alignment compared with state-of-the-art methods.

Named Entity Recognition (NER) is a core task of NLP. State-of-art supervised NER models rely heavily on a large amount of high-quality annotated data, which is quite expensive to obtain. Various existing ways have been proposed to reduce the heavy reliance on large training data, but only with limited effect. In this paper, we propose a novel way to make full use of the weakly-annotated texts in encyclopedia pages for exactly unsupervised NER learning, which is expected to provide an opportunity to train the NER model with no manually-labeled data at all. Briefly, we roughly divide the sentences of encyclopedia pages into two parts simply according to the density of inner url links contained in each sentence. While a relatively small number of sentences with dense links are used directly for training the NER model initially, the left sentences with sparse links are then smartly selected for gradually promoting the model in several self-training iterations. Given the limited number of sentences with dense links for training, a data augmentation method is proposed, which could generate a lot more training data with the help of the structured data of encyclopedia to greatly augment the training effect. Besides, in the iterative self-training step, we propose to utilize a graph model to help estimate the labeled quality of these sentences with sparse links, among which those with the highest labeled quality would be put into our training set for updating the model in the next iteration. Our empirical study shows that the NER model trained with our unsupervised learning approach could perform even better than several state-of-art models fully trained on newswires data.

Destination prediction is known as an important problem for many location based services (LBSs). Existing solutions generally apply probabilistic models to predict destinations over a sub-trajectory, but their accuracies in fine-granularity prediction are always not satisfactory due to the data sparsity problem. This paper presents a carefully designed deep learning model called TALL model for destination prediction. It not only takes advantage of the bidirectional Long Short-Term Memory (LSTM) network for sequence modeling, but also gives more attention to meaningful locations that have strong correlations w.r.t. destination by adopting attention mechanism. Furthermore, a hierarchical model that explores the fusion of multi-granularity learning capability is further proposed to improve the accuracy of prediction. Extensive experiments on Beijing and Chengdu real datasets finally demonstrate that our proposed models outperform existing methods without considering external features.

Traffic speed prediction is known as an important but challenging problem. In this paper, we propose a novel model, called LC-RNN, to achieve more accurate traffic speed prediction than existing solutions. It takes advantage of both RNN and CNN models by a rational integration of them, so as to learn more meaningful time-series patterns that can adapt to the traffic dynamics of surrounding areas. Furthermore, since traffic evolution is restricted by the underlying road network, a network embedded convolution structure is proposed to capture topology aware features. The fusion with other information, including periodicity and context factors, is also considered to further improve accuracy. Extensive experiments on two real datasets demonstrate that our proposed LC-RNN outperforms six well-known existing methods.

Entity resolution (ER) is the problem of identifying and grouping different manifestations of the same real world object. Algorithmic approaches have been developed where most tasks offer superior performance under supervised learning. However, the prohibitive cost of labeling training data is still a huge obstacle for detecting duplicate query records from online sources. Furthermore, the unique combinations of noisy data with missing elements make ER tasks more challenging. To address this, transfer learning has been adopted to adaptively share learned common structures of similarity scoring problems between multiple sources. Although such techniques reduce the labeling cost so that it is linear with respect to the number of sources, its random sampling strategy is not successful enough to handle the ordinary sample imbalance problem. In this paper, we present a novel multi-source active transfer learning framework to jointly select fewer data instances from all sources to train classifiers with constant precision/recall. The intuition behind our approach is to actively label the most informative samples while adaptively transferring collective knowledge between sources. In this way, the classifiers that are learned can be both label-economical and flexible even for imbalanced or quality diverse sources. We compare our method with the state-of-the-art approaches on real-word datasets. Our experimental results demonstrate that our active transfer learning algorithm can achieve impressive performance with far fewer labeled samples for record matching with numerous and varied sources.

Multi-label active learning for image classification has attracted great attention over recent years and a lot of relevant works are published continuously. However, there still remain some problems that need to be solved, such as existing multi-label active learning algorithms do not reflect on the cleanness of sample data and their ways on label correlation mining are defective. For one thing, sample data is usually contaminated in reality, which disturbs the estimation of data distribution and further hinders the model training. For another, previous approaches for label relationship exploration are purely based on the observed label distribution of an incomplete training set, which cannot provide sufficiently efficient information. To address these issues, we propose a novel adaptive low-rank multi-label active learning algorithm, called LRMAL. Specifically, we first use low-rank matrix recovery to learn an effective low-rank feature representation from the noisy data. In a subsequent sampling phase, we make use of its superiorities to evaluate the general informativeness of each unlabeled example-label pair. Based on an intrinsic mapping relation between the example space and the label space of a certain multi-label dataset, we recover the incomplete labels of a training set for a more comprehensive label correlation mining. Furthermore, to reduce the redundancy among the selected example-label pairs, we use a diversity measurement to diversify the sampled data. Finally, an effective sampling strategy is developed by integrating these two aspects of potential information with uncertainty based on an adaptive integration scheme. Experimental results demonstrate the effectiveness of our approach.

In multi-label image classification, each image is always associated with multiple labels and labels are usually correlated with each other. The intrinsic relation among labels can definitely contribute to classifier training. However, most previous studies on active learning for multi-label image classification purely mine label correlation based on observed label distribution. They ignore the mapping relation between examples and their labels. This mapping relation also implicates label relationship. Ignoring the mapping relation leads to an uncomprehensive label correlation estimation and results in a bad performance for classification. In this paper, we propose a novel multi-label active learning with low-rank mapping for image classification, called LMMAL, to solve this issue. More precisely, we train a low-rank mapping matrix to signify the mapping relation between the feature space and the label space of a certain multi-label dataset. Using this low-rank mapping relation, we exploit a full label correlation. Subsequently, an effective sampling strategy is designed by integrating this potential information with uncertainty to select the most informative example-label pairs. In addition, we extend LMMAL with automatic labeling (denoted as AL-LMMAL) to further reduce the annotation workload of active learning. Empirical results demonstrate the effectiveness of our approaches.

With the popularity of OSNs, finding a set of most influential users (or nodes) so as to trigger the largest influence cascade is of significance. For example, companies may take advantage of the “word-of-mouth” effect to trigger a large cascade of purchases by offering free samples/discounts to those most influential users. This task is usually modeled as an influence maximization problem, and it has been widely studied in the past decade. However, considering that users in OSNs may participate in various kinds of online activities, e.g., giving ratings to products, joining discussion groups, etc., influence diffusion through online activities becomes even more significant.

SIFT is an image local feature description algorithm based on scale-space. Due to its strong matching ability, SIFT has many applications in different fields, such as image retrieval, image stitching, and machine vision. After SIFT was proposed, researchers have never stopped tuning it. The improved algorithms that have drawn a lot of attention are PCA-SIFT, GSIFT, CSIFT, SURF and ASIFT. In this paper, we first systematically analyze SIFT and its variants. Then, we evaluate their performance in different situations: scale change, rotation change, blur change, illumination change, and affine change. The experimental results show that each has its own advantages. SIFT and CSIFT perform the best under scale and rotation change. CSIFT improves SIFT under blur change and affine change, but not illumination change. GSIFT performs the best under blur change and illumination change. ASIFT performs the best under affine change. PCA-SIFT is always the second in different situations. SURF performs the worst in different situations, but runs the fastest.

Image classification is a hot topic of pattern recognition in computer vision. In order to achieve high accuracy of classification, a certain amount of high quality pictures are needed. As a matter of fact, high quality pictures are scarce. Active learning can solve such a problem. Label dependences play an important role in multi-label active learning for image classification. The interdependences between different labels are usually different and asymmetrical. This paper first brings the asymmetrical conditional label dependences into a novel active learning method for multi-label image classification based on the asymmetrical conditional label dependence, called ACDAL. Our extensive experimental results on three image and two non-image datasets show that our new approach ACDAL significantly outperforms existing approaches.

Entity Matching (EM) identifies records referring to the same entity within or across databases. Existing methods using structured attribute values (such as digital, date or short string values) only may fail when the structured information is not enough to reflect the matching relationships between records. Nowadays more and more databases may have some unstructured textual attribute containing extra Consolidated Textual information (CText for short) of the record, but seldom work has been done on using the CText information for EM. Conventional string similarity metrics such as edit distance or bag-of-words are unsuitable for measuring the similarities between CTexts since there are hundreds or thousands of words with each CText, while existing topic models either can not work well since there is no obvious gaps between the various sub-topics in CText. In this paper, we work on employing CText in EM. A baseline algorithm identifying important phrases with high IDF scores from CTexts and then measuring the similarity between CTexts based on these phrases does not work well since it estimates the similarity in one dimension and neglects that these phrases belong to different topics. To this end, we propose a novel cooccurrence-based topic model to identify various sub-topics from each CText, and then measure the similarity between CTexts on the multiple sub-topic dimensions. Our empirical study on two real-world data set shows that our method outperforms the state-of-the-art EM methods and Text Understanding models by reaching a higher EM precision and recall.

In order to achieve better classification performance with even fewer labeled images, active learning is suitable for these situations. Several active learning methods have been proposed for multi-label image classification, but all of them assume that all training images with complete labels. However, as a matter of fact, it is very difficult to get complete labels for each example, especially when the size of labels in a multi-label domain is huge. Usually, only partial labels are available. This is one kind of "weak label" problems. This paper proposes an ingeniously solution to this "weak label" problem on multi-label active learning for image classification (called WLMAL). It explores label correlation on the weak label problem with the help of input features, and then utilizes label correlation to evaluate the informativeness of each example-label pair in a multi-label dataset for active sampling. Our experimental results on three real-world datasets show that our proposed approach WLMAL consistently outperforms existing approaches significantly.

Label correlation analysis is very important for multi-label classification. And there is no study to measure the label correlation for example-label based active learning. In this paper, from a statistical point of view, we proposed a cosine similarity based multi-label active learning (CosMAL), which uses cosine similarity to accurately evaluate the correlations between all labels. It further uses the average correlation between the potential label and the other unlabeled labels as the label information for each sample-label pair. And then we select the most informativeness example-label pairs. Our empirical results demonstrate that our proposed method CosMAL outperforms the state-of-the-art active learning for multi-label classification. It significantly reduces the labeling workload and improves the performance of a classifier learned.

Markov logic networks which unify probabilistic graphical model and first-order logic provide an excellent framework for ontology matching. The existing approach requires a threshold to produce matching candidates and use a small set of constraints acting as filter to select the final alignments. We introduce novel match propagation strategy to model the influences between potential entity mappings across ontologies, which can help to identify the correct correspondences and produce missed correspondences. The estimation of appropriate threshold is a difficult task. We propose an interactive method for threshold selection through which we obtain an additional measurable improvement. Running experiments on a public dataset has demonstrated the effectiveness of proposed approach in terms of the quality of result alignment.

Active learning is to select the most informative examples to request their labels. Most previous studies in active learning for multi-label classification didn't pay enough attention on label correlations. This leads to a bad performance for classification. In this paper, we proposed a chi-square statistics multi-label active learning (CSMAL) algorithm, which uses chi-square statistics to accurately evaluate correlations between labels. CSMAL considers not only positive relationships but also negative ones. It uses the average correlation between a potential label and its rest unlabeled labels as the label information for each sample-label pair. CSMAL further integrates uncertainty and label information to select example-label pairs to request labels. Our empirical results demonstrate that our proposed method CSMAL outperforms the state-of-the-art active learning methods for multi-label classification. It significantly reduces the labeling workloads and improves the performance of a classifier built.

Active learning has increasingly become an important paradigm for classification of networked data, where instances are connected with a set of links to form a network. In this paper, we propose a novel batch mode active learning method for networked data (BMALNeT). Our novel active learning method selects the best subset of instances from the unlabeled set based on the correlation matrix that we construct from the dedicated informativeness evaluation of each unlabeled instance. To evaluate the informativeness of each unlabeled instance accurately, we simultaneously exploit content information and the network structure to capture the uncertainty and representativeness of each instance and the disparity between any two instances. Compared with state-of-the-art methods, our experimental results on three real-world datasets demonstrate the effectiveness of our proposed method.

For multi-label image classification, we use active learning to select example-label pairs to acquire labels from experts. The core of active learning is to select the most informative examples to request their labels. Most previous studies in active learning for multi-label classification have two shortcomings. One is that they didn’t pay enough attention on label correlations. The other shortcoming is that existing example-label selection methods predict all the rest labels of the selected example-label pair. This leads to a bad performance for classification when the number of the labels is large. In this paper, we propose a semi-automatic labeling multi-label active learning (SLMAL) algorithm. Firstly, SLMAL integrates uncertainty and label informativeness to select example-label pairs to request labels. Then we choose the most uncertain example-label pair and predict its partial labels using its nearest neighbor. Our empirical results demonstrate that our proposed method SLMAL outperforms the state-of-the-art active learning methods for multi-label classification. It significantly reduces the labeling workloads and improves the performance of a classifier built.

Multi-label active learning for image classification has been a popular research topic. It faces several challenges, even though related work has made great progress. Existing studies on multi-label active learning do not pay attention to the cleanness of sample data. In reality, data are easily polluted by external influences that are likely to disturb the exploration of data space and have a negative effect on model training. Previous methods of label correlation mining, which are purely based on observed label distribution, are defective. Apart from neglecting noise influence, they also cannot acquire sufficient relevant information. In fact, they neglect inner relation mapping from example space to label space, which is an implicit way of modeling label relationships. To solve these issues, we develop a novel multi-label active learning with low-rank application (ENMAL) algorithm in this paper. A low-rank model is constructed to quantize noise level, and the example-label pairs that contain less noise are emphasized when sampling. A low-rank mapping matrix is learned to signify the mapping relation of a multi-label domain to capture a more comprehensive and reasonable label correlation. Integrating label correlation with uncertainty and considering sample noise, an efficient sampling strategy is developed. We extend ENMAL with automatic labeling (denoted as AL-ENMAL) to further reduce the annotation workload of active learning. Empirical research demonstrates the efficacy of our approaches.

To improve the classification performance of imbalanced learning, a novel oversampling method, immune centroids oversampling technique (ICOTE) based on an immune network, is proposed. ICOTE generates a set of immune centroids to broaden the decision regions of the minority class space. The representative immune centroids are regarded as synthetic examples in order to resolve the imbalance problem. We utilize an artificial immune network to generate synthetic examples on clusters with high data densities, which can address the problem of synthetic minority oversampling technique (SMOTE), which lacks reflection on groups of training examples. Meanwhile, we further improve the performance of ICOTE via integrating ENN with ICOTE, that is, ICOTE + ENN. ENN disposes the majority class examples that invade the minority class space, so ICOTE + ENN favors the separation of both classes. Our comprehensive experimental results show that two proposed oversampling methods can achieve better performance than the renowned resampling methods.

Multi-label image data is becoming ubiquitous. Image semantic understanding is typically formulated as a classification problem. This paper focuses on multi-label active learning for image classification. It first extends a traditional example based active learning method for multilabel active learning for image classification. Since the traditional example based active method doesn't work well, we propose a novel example-label based multi-label active learning method. Our experimental results on two image datasets demonstrate that the proposed method significantly reduces the labeling workload and improves the performance of the built classifier. Additionally, we conduct experiments on two other types of multi-label datasets for validating the versatility of our proposed method, and the experimental results show the consistent effect.

In most image classification algorithms, the classifier model needs to train a large number of training samples. In practical application, labeling numbers of samples is a tedious and time-consuming task. So, how to select fewer suitable training samples from the numbers of unlabeled samples is a difficulty in the image classification algorithm. This paper proposes a trademark number recognition technique based on active learning algorithm. The method uses the human interaction to get trademark number area, and then uses the projection method to extract character characteristic which using the characteristic to split characters. Finally, use BvSB active learning algorithm to select high information samples which was used to train support vector machine classifier, and use the trained classifier to recognize trademark number. The experimental result shows that the classifier trained by the method has higher classification accuracy in the case of labeled fewer samples.

Active learning is well-motivated in many modern machine learning problems, where unlabeled data may be abundant or easily obtained, but labels are difficult, time-consuming, or expensive to obtain. The selection and use of sampling strategy is core of active learning. Most active learning methods select those uncertain or representative unlabeled samples to query their labels. And some of the active learning methods consider both in the query selection. However, the uncertainty sampling methods rely on the relative correctness or confidence of the current model and suffer from a lack of the feature space. The representative sampling methods avoid the drawbacks associated with uncertainty sampling, but tend not to improve the learning model very efficiently. The combining methods are lack of the consideration of sample redundancy. This paper proposes a multicriterion active learning technique for solving multiclass problems. First, use the Best-versus-Second-Best (BvSB) method to calculate the sample’s uncertainty and then select the most valuable component to constitute the uncertain set; further, use the kernel k-means clustering algorithm and the resulting sample set is divided into h different clusters; finally, use Gaussian process to select the most informative sample in each cluster and submit to human experts for annotation. The results show that the labeling cost can be reduced without degrading the performance.

Active Learning is a machine learning and data mining technique that selects the most informative samples for labeling and uses them as training data. It aims to obtain a high performance classifier by labeling as little data as possible from large amount of unlabeled samples, which means sampling strategy is the core issue. Existing approaches either tend to ignore information in unlabeled data and are prone to querying outliers or noise samples, or calculate large amounts of non-informative samples leading to significant computation cost. In order to solve above problems, this paper proposed a serial active learning framework. It first measures uncertainty of unlabeled samples and selects the most uncertain sample set. From which, it further generates the most representative sample set based on the mutual information criterion. Finally, the framework selects the most informative sample from the most representative sample set based on expected error reduction strategy. Experimental results on multiple datasets show that our approach outperforms Random Sampling and the state of the art adaptive active learning method.

A motion trajectory is an intuitive representation form in time-space domain for a micromotion behavior of moving target. Trajectory analysis is an important approach to recognize abnormal behaviors of moving targets. Against the complexity of vehicle trajectories, this paper first proposed a trajectory pattern learning method based on dynamic time warping (DTW) and spectral clustering. It introduced the DTW distance to measure the distances between vehicle trajectories and determined the number of clusters automatically by a spectral clustering algorithm based on the distance matrix. Then, it clusters sample data points into different clusters. After the spatial patterns and direction patterns learned from the clusters, a recognition method for detecting vehicle abnormal behaviors based on mixed pattern matching was proposed. The experimental results show that the proposed technical scheme can recognize main types of traffic abnormal behaviors effectively and has good robustness. The real-world application verified its feasibility and the validity.

Multi-label classification, where each instance is assigned with multiple labels, has been an attractive research topic in data mining. The annotations of multi-label instances are typically more difficult and time consuming, since they are simultaneously associated with multiple labels. Therefore, active learning, which reduces the labeling cost by actively querying the labels of the most valuable data, becomes particularly important for multi-label learning. Study reveals that methods querying instance-label pairs are more effective than those query instances, since for each sample, only some effective labels need to be annotated while others can be inferred by exploring the label correlations. However, with the high dimensionality of label space, the instance-label pair selective algorithm will be affected since the computational cost of training a multi-label model may be strongly affected by the number of labels. In this paper we propose an approach that combines instance sampling with optimal label subset selection, which can effectively improve the classification model performance and substantially reduce the annotation cost. Experimental results demonstrate the superiority of the proposed approach to state-of-the-art methods on three benchmark datasets.