Multivariate time series prediction task has long been a crucial problem in machine learning since it includes applications in real life such as energy consumption, traffic congestion and economic trending. Challenge is that there are many factors affecting the accuracy of prediction, such as information in time domain and frequency domain, as well as interdependence among sequences. Therefore, effectively capturing and combining features in different domains becomes the focus of research. In this paper, a deep learning framework enhanced by wavelet transform is proposed to predict multivariate time series. Our model captures the long-term and short-term characteristics of time series in the time domain respectively through convolutional network. Further more, we use LSTM to achieve trending vector of short-term features, while for long-term features, we use the multi-head attention mechanism to get its representation. For frequency-domain features, our framework extracts from the scalo-gram generated by applying wavelet transform to the time sequence. Through the design of CNN convolution kernel, our model also ensures that the dependence between time series can be extracted. Experimental results show that our proposed model outperforms other baseline approaches on three of the four realworld datasets with sophisticated patterns.

Real web datasets are often associated with multiple views such as long and short commentaries, users preference and so on. However, with the rapid growth of user generated texts, each view of the dataset has a large feature space and leads to the computational challenge during matrix decomposition process. In this paper, we propose a novel multi-view clustering algorithm based on the non-negative matrix factorization that attempts to use feature sampling strategy in order to reduce the complexity during the iteration process. In particular, our method exploits unsupervised semantic information in the learning process to capture the intrinsic similarity through a graph regularization. Moreover, we use Hilbert Schmidt Independence Criterion (HSIC) to explore the unsupervised semantic diversity information among multi-view contents of one web item. The overall objective is to minimize the loss function of multi-view non-negative matrix factorization that combines with an intra-semantic similarity graph regularizer and an inter-semantic diversity term. Compared with some state-of-the-art methods, we demonstrate the effectiveness of our proposed method on a large real-world dataset Doucom and the other three smaller datasets.

Measuring semantic relatedness between two words is a significant problem in many areas such as natural language processing. Existing approaches to the semantic relatedness problem mainly adopt the co-occurrence principle and regard two words as highly related if they appear in the same sentence frequently. However, such solutions suffer from low coverage and low precision because i) the two highly related words may not appear close to each other in the sentences, e.g., the synonyms; and ii) the co-occurrence of words may happen by chance rather than implying the closeness in their semantics. In this paper, we explore the latent semantics (i.e., concepts) of the words to identify highly related word pairs. We propose a hierarchical association network to specify the complex relationships among the words and the concepts, and quantify each relationship with appropriate measurements. Extensive experiments are conducted on real datasets and the results show that our proposed method improves correlation precision compared with the state-of-the-art approaches.

For many large-scale multimedia datasets and web contents, the nearest neighbor search methods based on the hashing strategy for cross-modal retrieval have attracted considerable attention due to its fast query speed and low storage cost. Most existing hashing methods try to map different modalities to Hamming embedding in a supervised way where the semantic information comes from a large manual label matrix and each sample in different modalities is usually encoded by a sparse label vector. However, previous studies didn't address the semantic correlation learning challenges and couldn't make the best use of the prior semantic information. Therefore, they cannot preserve the accurate semantic similarities and often degrade the performance of hashing function learning. To fill this gap, we firstly proposed a novel Deep Semantic Correlation learning based Hashing framework (DSCH) that generates unified hash codes in an end-to-end deep learning architecture for cross-modal retrieval task. The major contribution in this work is to effectively automatically construct the semantic correlation between data representation and demonstrate how to utilize correlation information to generate hash codes for new samples. In particular, DSCH integrates latent semantic embedding with a unified hash embedding to strengthen the similarity information among multiple modalities. Furthermore, additional graph regularization is employed in our framework, to capture the correspondences from the inter-modal and intra-modal. Our model simultaneously learns the semantic correlation and the unified hash codes, which enhances the effectiveness of cross-modal retrieval task. Experimental results show the superior accuracy of our proposed approach to several state-of-the-art cross-modality methods on two large datasets.

The task of automatically generating image captions draws a lot of attention in the past few years because it shows great potential in a wide range of application scenarios. The encoder-decoder structure with attention mechanism has been extensively applied to solve this task. However, most researches apply attention mechanism only to pay attention to image features but neglect the relations between image features which we think play an important role in scene understanding. To tackle this problem, we propose a novel attention mechanism named “attention to Time-Dependent Pre-Attention” (TDPA-attention) and the TDPA-attention is combined with a hierarchical LSTM decoder to compose our captioning model (TDPA-model). Within our TDPA-attention, at every time step, every image feature pays attention to all image features according to a semantic context and the attended feature is treated as an aggregated feature that contains relations between this image feature and all image features. All these aggregated features form a new feature set that the hierarchical LSTM decoder attends to. We evaluate our model on public image caption dataset Microsoft COCO and achieve state-of-the-art performance on most evaluation metrics.

For resource management purpose, administrators usually need to perform what-if analyses to predict the impact of any workload growths or planned changes on the performance of web services. A what-if analysis requires not only the design of system models, but also the workload models that represent the real-world user behavior. Existing methods of workload characterization based on probabilistic graphical models are quite complex if there are many web services provided by a system. Meanwhile, bandwidth resource is usually not taken into account in many related works, though it is a relatively expensive resource in cloud markets. In fact, it's very challenging to predict the network throughput of modern web services due to the factors of client-side caching, miscellaneous service responses and complex network transportation. In this paper we propose a methodology of what-if analysis named Log2Sim for the bandwidth management of web systems. We use a lightweight workload model to describe user behavior, an automated mining approach to obtain characteristics of workloads and responses from massive web logs, and traffic-aware simulations to predict the impact on the network throughput and the response time within changing contexts of user behavior. We also choose a real-life web system as use case to evaluate the effectiveness, accuracy and stability of this methodology.