Personalized hashtag recommendation aims to au-tomatically recommend user-specific hashtags to annotate the posts (e.g., tweets). It is actually an unwieldy procedure because only a small number of posts contain hashtags, which can hinder the quality of recommendation results. Moreover, users always have their own preferences while choosing the hashtag to describe the post contents. Thus, solving label scarcity and user bias both have raised researchers' attention toward hashtag recommendation. The existing methods can be generally divided into two groups: The first group models the relationship between hashtags and posts taking only post contents into account, without considering the user preferences. The second group models the relationship between users and posts assuming that each user has adequate posts tapped with hashtags. In this paper, we propose a Meta-learning based Personalized Hashtag Recommendation (MetaTag) framework to address above mentioned challenges at once. Our method can leverage the prior experience learned from other users and quickly adapt to a new user. In this framework, model is trained in an episodic manner with user-specific historical posts to learn semantic embeddings that can distinguish different hashtags even they have not been seen before. The results of experiments on the data collected from Twitter demonstrated that our method outperforms state-of-the-art approaches by a large margin.

A social media user's geographical location is vital to many applications like local search and event detection. The scarcity of publicly available location information motivates researchers to predict user geolocation based on information such as tweet text and social interaction data. In this paper, we investigate and improve on the task of predicting a Twitter user's city-level location based on the content of the user's historical tweets. In order to train a reliable location classifier, previous studies on this topic have typically assumed that there are sufficient amount of users living in each cities. However, they simply ignore the fact that different demographic groups may participate in social media platforms, which results in a highly imbalanced data distribution. Being aware of this population imbalance issue, we propose an episodic learning based framework to extract a single representative for each class (location), so that classifiers can later be trained on a balanced class distribution. To examine the effectiveness of our method, we design experiments which involve two kinds of baselines, the state-of-the-art geolocation detection methods and the well-known approaches handling imbalanced data in classification. The results of experiments on the data collected from Twitter demonstrated the superiority of our method when compared with baselines.

Non-stationary data stream mining aims to classify large scale online instances that emerge continuously. The most apparent challenge compared with the offline learning manner is the issue of consecutive emergence of new categories, when tackling non-static categorical distribution. Non-stationary stream settings often appear in real-world applications, e.g., online classification in E-commerce systems that involves the incoming productions, or the summary of news topics on social networks (Twitter). Ideally, a learning model should be able to learn novel concepts from labeled data (in new tasks) and reduce the abrupt degradation of model performance on the old concept (also named catastrophic forgetting problem). In this work, we focus on improving the performance of the stream mining approach under the constrained resources, where both the memory resource of old data and labeled new instances are limited/scarce. We propose a simple yet efficient resource-constrained framework CLEAR to facilitate previous challenges during the one-pass stream mining. Specifically, CLEAR focuses on creating and calibrating the class representation (the prototype) in the embedding space. We first apply the contrastive-prototype learning on large amount of unlabeled data, and generate the discriminative prototype for each class in the embedding space. Next, for updating on new tasks/categories, we propose a drift estimation strategy to calibrate/compensate for the drift of each class representation, which could reduce the knowledge forgetting without storing any previous data. We perform experiments on public datasets (e.g., CUB200, CIFAR100) under stream setting, our approach is consistently and clearly better than many state-of-the-art methods, along with both the memory and annotation restriction.

Continuously mining complexity data stream has recently been attracting an increasing amount of attention, due to the rapid growth of real-world vision/signal applications such as self-driving cars and online social media messages. In this paper, we aim to address two significant problems in the lifelong/incremental stream mining scenario: first, how to make the learning algorithms generalize to the unseen classes only from a few labeled samples; second, is it possible to avoid storing instances from previously seen classes to solve the catastrophic forgetting problem? We introduce a novelty stream mining framework to classify the infinite stream of data with different categories that occurred during different times. We apply a few-sample learning strategy to make the model recognize the novel class with limited samples; at the same time, we implement an incremental generative model to maintain old knowledge when learning new coming categories, and also avoid the violation of data privacy and memory restrictions simultaneously. We evaluate our approach in the continual class-incremental setup on the classification tasks and ensure the sufficient model capacity to accommodate for learning the new incoming categories.

Existing solutions to multi-task learning typically rely on manually enumerating multiple network architectures to find the optimal structure, which incurs a heavy design workload. In addition, extending these models to new tasks is difficult in many cases, since it often requires a network redesign for achieving the best performance. To overcome these limitations, in this paper, we propose a novel principle for multitask learning, which focuses on the learning process itself. For each task, we consider its feature as a state, and treat the learning problem as a state transformation process, which is driven by a task-specific gradient. Specifically, we introduce a Gradient Modification Unit (GMU), which consists of a Gradient Estimation Network (GEN) (shared among all tasks) and multiple tiny task-specific gradient correctors (one for each task). At each iteration, for any task T, the corresponding gradient corrector modifies the gradient estimated by the GEN to produce the desired task-specific gradient tensor that is applied to update the state of task T. Our solution has several benefits including automatic end-to-end learning of the optimal transformation process, simplification of network design and high expansibility to new tasks. We demonstrate the superiority of our approach over existing solutions on a variety of datasets, across both image classification and image retrieval tasks.

Automated diagnosis of significant abnormalities (or lesions) from radiology images has been well exploited in Deep Learning (DL) because of the ability to model sophisticated features. However, a deep neural network should be trained on a huge amount of data to infer the parameter values. Unfortunately, for the problems in lesion diagnosis, there is only a limited amount of data annotated in a manner that is suitable to learn powerful deep models. Moreover, the lesion in the radiology image is often vague and hard to identify without expert knowledge. In this paper, we focus on previous challenges in the automated diagnosis and propose the approach named Adaptive Multi-region Network (AdapNet). The key idea is that we adaptively encode the similarity of lesions in different context regions through margin-max learning strategy, which incorporates the metrics learned on those regions to enhance the effectiveness of the model. Our experiments show that the proposed method can effectively obtain superior performance compared to the existing methods, on the DeepLesion data sets.

A primary challenge in label prediction over a data stream is the emergence of instances belonging to unknown or novel class over time. Traditionally, studies addressing this problem aim to detect such instances using cluster-based mechanisms. They typically assume that instances from the same class are closer to each other than those belonging to different classes in observed feature space. Unfortunately, this may not hold true in higher-dimensional feature space such as images. In recent years, Convolutional neural network (CNN) have emerged as a leading system to be employed in many real-world application. Yet, based on the assumption of closed world dataset with a fixed number of categories, CNN lacks robustness for novel class detection, so it is unclear on how such models can be used to deal with novel class instances along a high-dimensional image stream. In this paper, we focus on addressing this challenge by proposing an effective learning framework called CNN-based Prototype Ensemble (CPE) for novel class detection and correction. Our framework includes a prototype ensemble loss (PE) to improve the intra-class compactness and expand inter-class separateness in the output feature representation, thereby enabling the robustness of novel class detection. Moreover, we provide an incremental learning strategy which maintains a constant amount of exemplars to update the network, making it more practical for real-world application. We empirically demonstrate the effectiveness of our framework by comparing its performance over multiple realworld image benchmark data streams with existing state-of-theart data stream detection techniques. The implementation of CPE is on: https://github.com/Vitvicky/Convolutional-Net-PrototypeEnsemble

With the tremendous increase of the online data, training a single classifier may suffer because of the large variety of data domains. One solution could be to learn separate classifiers for each domain. However, this would arise a huge cost to gather annotated training data for a large number of domains and ignore similarity shared across domains. Hence, it leads to our problem setting: can labeled data from a related source domain help predict the unlabeled data in the target domain? In this paper, we consider two independent simultaneous data streams, which are referred to as the source and target streams. The target stream continuously generates data instances from one domain where the label is unknown, while the source stream continuously generates labeled data instances from another domain. Most likely, the two data streams would have different but related feature spaces and different data distributions. Moreover, these streams may have asynchronous concept drifts between them. Our problem setting, which is called Cross-domain Multistream Classification, is to predict the class labels of data instances in the target stream using a classifier trained on the labeled source stream. In this paper, we propose an efficient solution for cross-domain multistream classification by integrating change detection into online data stream adaptation. The class labels of data instances in the target stream are predicted using the sufficient amount of label information in the related source stream. And the concept drifts along the two independent streams are continuously being addressed at the same time. Experimental results on real-world data sets indicate significantly improved performance over baseline methods.

A primary challenge in label prediction over a stream of continuously occurring data instances is the emergence of instances belonging to unknown or novel classes. It is imperative to detect such novel-class instances quickly along the stream for a superior prediction performance. Existing techniques that perform novel class detection typically employ a clustering-based mechanism by observing that instances belonging to the same class (intra-class) are closer to each other (cohesion) than inter-class samples (separation). While this is generally true in low dimensional feature spaces, we observe that such a property is not intrinsic among instances in complex real-world high-dimensional feature space such as images and text. In this paper, we focus on addressing this key challenge that negatively affects prediction performance of a data stream classifier. Concretely, we develop a metric learning mechanism that transforms high-dimensional features into a latent feature space to make above property holds true. Unlike existing metric learning method which only focus on classification task, our approach address the novel class detection and stream classification simultaneously. We showcase a framework along the stream to achieve larger prediction performance compared to existing state-of-the-art detection techniques while using the least amount of labeled data during detection. Extensive experimental results on simulated and real-world stream demonstrate the effectiveness of our approach.

Under a newly introduced setting of multistream classification, two data streams are involved, which are referred to as source and target streams. The source stream continuously generates data instances from a certain domain with labels, while the target stream does the same task without labels from another domain. Existing approaches assume that domains for both data streams are identical, which is not quite true in real world scenario, since data streams from different sources may contain distinct features. Furthermore, obtaining labels for every instance in a data stream is often expensive and time-consuming. Therefore, it has become an important topic to explore whether labeled instances from other related streams can be helpful to predict those unlabeled instances in a given stream. Note that domains of source and target streams may have distinct features spaces and data distributions. Our objective is to predict class labels of data instances in the target stream by using the classifiers trained by the source stream.

Regression over a stream of data is challenging due to unbounded data size and non-stationary distribution over time. Typically, a traditional supervised regression model over a data stream is trained on data instances occurring within a short time period by assuming a stationary distribution. This model is later used to predict value of response-variable in future instances. Over time, the model may degrade in performance due to changes in data distribution among incoming data instances. Updating the model for change adaptation requires true value for every recent data instances, which is scarce in practice. To overcome this issue, recent studies have employed techniques that sample fewer instances to be used for model retraining. Yet, this may introduce sampling bias that adversely affects the model performance. In this paper, we study the regression problem over data streams in a novel setting. We consider two independent, yet related, non-stationary data streams, which are referred to as the source and the target stream. The target stream continuously generates data instances whose value of response variable is unknown. The source stream, however, continuously generates data instances along with corresponding value for the response-variable, and has a biased data distribution with respect to the target stream. We refer to the problem of using a model trained on the biased source stream to predict the response-variable’s value in data instances occurring on the target stream as Multistream Regression. In this paper, we describe a framework for multistream regression that simultaneously overcomes distribution bias and detects change in data distribution represented by the two streams over time using a Gaussian kernel model. We analyze the theoretical properties of the proposed approach and empirically evaluate it on both real-world and synthetic data sets. Importantly, our results indicate superior performance by the framework compared to other baseline regression methods.

In traditional machine learning, it is assumed that training data conforming to the stationary distribution of test data is readily available. Yet, such an assumption is not valid in practice due to a high cost of obtaining the truth value of data instances. This is particularly true when computing over non-stationary data streams. Recent studies in the multistream setting aim to address this issue by leveraging a stream of data with biased labeled instances (called the source stream) to train a suitable model for prediction over unlabeled instances (called the target stream). They use sampling bias correction techniques as a preprocessing step for estimating source instance weights for training a bias-corrected classifier useful to predict label of target data instances. In this regard, a recent framework proposes to utilize a Gaussian kernel model to estimate source instance weights. Unfortunately, it suffers from large computational time complexity and consequently deteriorates the rate at which streaming data is processed. In this paper, we address this issue by proposing a divide-and-conquer method suitable for simultaneously evaluating source instance weights and detecting changes in distribution over time. Our empirical results demonstrate a significant gain in execution time, compared to the previous approach, while achieving similar or better classification accuracy on real-world datasets.