Search

The recent advances usually mine market information from the chaotic data to conduct a stock movement prediction task. However, the current stock price movement prediction approaches mainly compute attention weighted sum of the global contextual semantic embeddings, which fails to combine local word-level or char-level ones to jointly learn news-level representation. Moreover, for Chinese stock price movement prediction task, some collected news texts are chaotic even irrelevant to the target stock. It suggests that the models need filter some news-level representations (viewed as noises) to enhance the performance. To that aim, we develop a novel stock price movement prediction network via bidirectional gated recurrent unit (GRU) network based on reinforcement learning (RL) with incorporated attention mechanism. In specific, to reduce the noise of news texts and learn news-level representation with more abundant semantics, two novel attention mechanisms respectively based on add and dot operation were first proposed in this work. We then design a novel GRU structure based on RL to filter some irrelated news-level representations (i.e., news-level noises) and capture abundant long-term dependencies. Finally, the experimental results show that the proposed model far outperforms the recent advances and achieves state-of-the-art performances.

Fine-Grained Visual Categorization (FGVC) is a challenging task due to the large intra-subcategory and small inter-subcategory variances. Recent studies tackle this task through a weakly supervised manner without using the part annotation from the experts. Of those, methods based on bilinear pooling are one of the main categories for computing the interaction between deep features and have shown high effectiveness. However, these methods mainly focus on the correlation within one specific layer but largely ignore the high interactions between multiple layers. In this study, we argue that considering the high interaction between the features from multiple layers can help to learn more distinguishing fine-grained features. To this end, we propose a High-Order-Interaction (HOI) method for FGVC. In our HOI, an efficient cross-layer trilinear pooling is introduced to calculate the third-order interaction between three different layers. Third-order interactions of different combinations are then fused to form the final representation. HOI can produce more discriminative representations and be readily integrated with the two popular techniques, attention mechanism and triplet loss, to obtain superposed improvement. Extensive experiments conducted on four FGVC datasets show the great superiority of our method over bilinear-based methods and demonstrate that the proposed method achieves the state of the art.

This study considers the problem of text-based person search, which aims to find the corresponding person of a given text description in an image gallery. Existing methods usually learn a similarity mapping of local parts between image and text, or embed the whole image and text into a unified embedding space. However, the relevance between local and the whole is largely underexplored. In this paper, we design a Divide-and-Merge Embedding (DME) learning framework for text-based person search. DME explicitly 1) models the relations between local parts and global embedding, 2) incorporates local details into global embedding. Specifically, we design a Feature Dividing Network (FDN) to embed the input into K locally guided semantic representations by self-attentive embedding, each representation depicts a local part of the person. Then, we propose a Relevance based Subspace Projection (RSP) method for merging diverse local representations to a compact global embedding. RSP helps the model to obtain discriminative embedding by jointly minimizing the redundancy of local parts and maximizing the relevance between local parts and global embedding. Extensive experimental results on three challenging benchmarks, i.e., CUHK-PEDES, CUB and Flowers datasets, have demonstrated the effectiveness of the proposed method.

Predicting 30-day hospital readmission is a core research task in the development of personalized healthcare. However, the imbalanced class distribution and the heterogeneity of electronic health records are the major challenges to establish an effective machine learning model for this task. To overcome these issues, we propose a new 30-day readmission prediction algorithm to improve the performance. First, we solve the problem of class-imbalance readmission prediction by learning sample weights based on hypothesis margin loss. At the same time, we consider the character of data heterogeneity, and learn the weights of heterogeneous data sources to improve the generalization ability. Based on this, we construct an optimization framework, which involves two variables, i.e., sample weights and source weights. By iterative optimization, we obtain the prediction result for readmission. Finally, we conduct experiments on three real-world readmission datasets to verify the effectiveness of the proposed method. The experimental results show that the proposed algorithm has the advantages to deal with the task of 30-day hospital readmission prediction.

The recent advances usually attempt to mine the effective market information from the chaotic data and learn multilevel representations by using attention mechanisms to conduct a stock prediction task. However, such methods usually lack the full utilization of local semantic embedding which contains the abundant textual semantics information. Moreover, these models suffer from the severe noise diffusion in contextual embeddings from a sequence after passing through the RNN. The noises diffusion constrains the performance of the proposed methods. In this work, we propose a stock movement predictive network via incorporative attention mechanisms. The core innovation is that the incorporative attention combines local and contextual attention mechanisms to clean the contextual embeddings by using local semantics. As a result, the attention effectively reduce the noises in the constructed higher-level representations and enhance the performance. Moreover, the local semantics and context are merged into the constructed higher-level representations which provide more abundant local semantic and contextual information. The experimental results demonstrate the state-of-the-art performance of the proposed approach on tweet and historical price dataset.

Most previous works on future market price forecasting only utilize the historical transaction data, while ignoring many other valuable factors. Recently, many research works propose multiple-source data-based predicting approaches in the stock market. Although the futures market and the stock market are very similar, the futures market still has its uniqueness. Most importantly, the subject matter of futures is usually commodity entities with prominent competing products or upstream, downstream industries, which can significantly influence the price. Therefore, it is essential to propose a future specific analysis framework by considering different factors. In this study, we constructed a Multi-source Heterogeneous Data Analysis (MHDA) method for future price prediction by integrating multiple-source information, i.e., trading data, news event data, and investor comments. Firstly, we first constructed a relation map to capture all related news events from upstream and downstream commodities and then built a future-specific sentiment dictionary to accurately quantify the sentiment impact of related news and investor comments during the feature extraction. Finally, we model the quantified multi-source heterogeneous information by an extended Hidden Markov Model to capture the underlying temporal dependency in the data. Evaluations on the data of palm oil futures from 2016.9 to 2017.9 show the effectiveness of our proposed framework.

Multi-label learning has been extensively studied in many areas such as information retrieval, bioinformatics, and multimedia annotation. However, multi-label datasets often have noisy, irrelevant and redundant features with high dimensionality. Accompanying with these issues, a critical challenge is known as the curse of dimensionality. As an effective data preprocessing method, feature selection has received much attention for that it can provide a way in reducing computation time, improving prediction performance and enhancing understanding of the data. Based on this, a large number of information-theoretical-based feature selection methods are developed to solve the learning problem, i.e. multi-label classification. Unfortunately, most of existing information-theoretical-based feature selection methods are either directly transformed from single-label feature selection methods or insufficient in light of using heuristic algorithms as the search component. Motivated by this, in this paper, we propose a novel mutual-information-based feature selection method, which obtains the optimal solution via constrained convex optimization with less time. Specifically, by incorporating the label information into the feature selection process, label-correlation is taken into consideration to generate the generalized model. Finally, the experimental results on various multi-label datasets demonstrate the effectiveness and efficiency of the proposed framework.

Human action recognition from RGBD videos has attracted much attention recently in the area of computer vision. Mainstream methods focus on designing highly discriminative features, which suffer from high dimension. As for human experience, discriminative parts, such as hands or legs, play an important role for identifying human actions. Motivated by this phenomenon, we propose a Random Forest (RF) Out-of-Bag (OB) estimation based approach to extract discriminative parts for each action. First, all the features of joint-based parts are separately fed into the RF Classifier. The OB estimation of each part is used to evaluate the discrimination of the joints in the part. Second, joints with high discrimination for the whole dataset are selected to design feature. Therefore, feature dimension is reduced efficiently. Experiments conducted on MSR Action 3D and MSR Daily Activity3D dataset show that our proposed approach outperforms state-of-the-art methods in accuracy with lower feature dimensions.

Object proposal generation is an important step in object detection, obtaining high-quality proposals can effectively improve the performance of detection. In this paper, we propose a semantic, class-specific approach to re-rank object proposals, which can consistently improve the recall performance even with fewer proposals. Specifically, we first extract features for each proposal including semantic segmentation, stereo information, contextual information, CNN-based objectness and low-level cue, and then score them using class-specific weights learned by Structured SVM. The advantages of the proposed model are two-fold: 1) it can be easily merged to existing generators with few computational costs, and 2) it can achieve high recall rate under strict critical even using fewer proposals. Experimental evaluation on the KITTI benchmark demonstrates that our approach significantly improves existing popular generators on recall performance. Moreover, in the experiment conducted for object detection, even with 1500 proposals, our approach can still have higher average precision (AP) than baselines with 5000 proposals.

Product quantization and its variances have emerged in approximate nearest neighbor search, with a wide range of applications. However, the optimized division of product subspaces retains as an open problem that largely degenerates the retrieval accuracy. In the paper, an extremely optimized product quantization scheme is introduced, which ensures, both theoretically and experimentally, a much better subspace partition comparing to the existing state-of-the-arts PQ and OPQ. The key innovation is to formulate subspace partition as a graph-based optimization problem, by which dynamic programming is leveraged to pursuit optimal quantizer learning. Another advantage is that the proposed scheme is very easily integrated with the cutting-edge multi-indexing structure, with a nearly eligible overhead in addition. We have conducted a serial of large-scale quantitative evaluations, with comparisons to a group of recent works including PQ, OPQ, and multi-Index. We have shown superior performance gain in the widely used SIFT1B benchmark, which validates the merits of the proposed algorithm.

Keypoint descriptors play an indisputable role in modern vision, whether it is for image registration, content-based retrieval, 3D reconstruction, or object recognition. In that perspective, SIFT descriptor is the most widely used due to its robustness to scale, rotation, and perspective distortions as well as illumination changes. However, the processing effort and the amount of memory required by SIFT make it ill-suited for applications with limited memory and processing power. In this paper, we propose a novel keypoint descriptor called CBDF (Compressed Binary Discriminative Feature) designed to be accurate, memory efficient and fast at the same time. High speed and memory efficiency are made possible due to a binary descriptor obtained by aggregating gradient features. Local discriminative analysis (LDA) is used to make sure the aggregated feature leads to a discriminative descriptor. Accuracy is also enforced by the use of scale, rotation, and illumination invariant features. The CBDF descriptor is stored on a 32-byte vector (which is one order of magnitude smaller than the usual 512-byte vector for SIFT) while matching is done with the Hamming distance. Extensive evaluations on various benchmark datasets show that CBDF is not only faster and generally more accurate than SIFT and SURF, but also more accurate than recent descriptors such as ORB, BRISK and BRIEF.

While much exciting progress is being made in 3D reconstruction of scenes, object labeling of 3D point cloud for indoor scenes has been left as a challenge issue. How should we explore the reference images of 3D scene, in aid of scene parsing? In this paper, we propose a framework for 3D indoor scenes labeling, based upon object detection on the RGB-D frames of 3D scene. First, the point cloud is segmented into homogeneous segments. Then, we utilize object detectors to assign class probabilities to pixels in every RGB-D frame. After that, the class probabilities are projected into the segments. Finally, we perform accurate inference on a MRF model over the homogeneous segments, in combination with geometry cues to output the labels. Experiment on the challenging RGB-D Object Dataset demonstrates that our detection based approach produces accurate labeling and improves the robustness of small object detection for indoor scenes.

Coming with the emerging of depth sensors link Microsoft Kinect, human hand gesture recognition has received ever increasing research interests recently. A successful gesture recognition system has usually heavily relied on having a good feature representation of data, which is expected to be task-dependent as well as coping with the challenges and opportunities induced by depth sensor. In this paper, a feature learning approach based on sparse auto-encoder (SAE) and principle component analysis is proposed for recognizing human actions, i.e. finger-spelling or sign language, for RGB-D inputs. The proposed model of feature learning is consisted of two components: First, features are learned respectively from the RGB and depth channels, using sparse auto-encoder with convolutional neural networks. Second, the learned features from both channels is concatenated and fed into a multiple layer PCA to get the final feature. Experimental results on American sign language (ASL) dataset demonstrate that the proposed feature learning model is significantly effective, which improves the recognition rate from 75% to 99.05% and outperforms the state-of-the-art.

In this paper we propose an online multiple instance boosting algorithm with instance-level semi-supervised learning, termed SemiMILBoost, to achieve robust object tracking. Our work revisits the multiple instance learning (MIL) formulation to alleviate the drifting problem in tracking, which addresses two key issues in the existing MIL based tracking-by-detection methods, i.e., the unselective treatment of instances in the positive bag during weak classifier updating and the lack of object prior knowledge in instance modeling. We tackle both issues in a principled way by using a robust SemiMILBoost algorithm, which treats instances in the positive bag as unlabeled while the ones in the negative bag as negative. To improve the discriminability of weak classifiers online, we iteratively update them with the pseudo-labels and importance of all instances in the positive bag, which are predicted by employing the instance-level semi-supervised learning technique with object prior knowledge during boosting. Experimental results demonstrate that our proposed algorithm outperforms the state-of-the-art tracking methods on several challenging video sequences.