Search

We are living in the age of big data, advanced analytics, and data science. The trend of “big data growth” [29, 106, 266, 288, 413] (data deluge [210]) has not only triggered tremendous hype and buzz, but more importantly presents enormous challenges, which in turn have brought incredible innovation and economic opportunities.

© 2018, Springer Nature Switzerland AG. Training a binary classifier only on positive and unlabeled examples (i.e., the PU learning) is an important yet challenging issue, widely seen in many problems in which it is difficult to obtain negative examples. Existing methods for handling this challenge often perform unsatisfactorily, since they often ignore the relations between positive and unlabeled examples and are also limited to the traditional shallow learning frameworks. Therefore, this work proposes a new approach: Deep Positive and Unlabeled learning with Manifold Regularization (Deep-PUMR), which integrates the manifold regularization with deep neural networks to address the above issues with classic PU learning. Deep-PUMR holds two major advantages: (i) Our method exploits the manifold properties of data distribution to capture the relationship of positive and unlabeled examples; (ii) The adopted deep network enables Deep-PUMR with strong learning ability, especially on large-scale datasets. Extensive experiments on five diverse datasets demonstrate that Deep-PUMR achieves the state-of-the-art performance in comparison with classic PU learning algorithms and risk estimators.

In the age of analytics, data analytics and learning form a comprehensive spectrum and evolutionary map that cover.

This chapter addresses the fundamental question: what lays the foundations for data science as a new science? Several relevant disciplines and areas are included: cognitive science, statistics, information science, intelligence science, computing, social science, management, and communication studies.

An increasing number of data science courses are available from research institutions and professional course providers. However, most such courses may look like “old wine in new bottles”, i.e., they are a re-labeling and combination of existing subjects in statistics, business and IT.

The art of data science [197] has increasingly attracted interest from a wide range of domains and disciplines. Communities or proposers from diverse backgrounds have often had contrasting aspirations, and have accordingly presented very different views or demonstrated contrasting foci.

In principle, data science can be applied to any application area or business domain. However, data science projects are intrinsically different from ordinary software development and IT projects.

What makes data science essential and different from existing developments in data mining, machine learning, statistics, and information science?

There is continuing debate about how data science will evolve in the next 50 years and what it will ultimately look like.

What are the greatest challenges of big data and data science? This question itself is problematic as data science is at a very early stage and has been built on existing disciplines. This chapter explores this important issue.

© Springer International Publishing AG, part of Springer Nature 2018. Financial variables such as asset returns in the massive market contain various hierarchical and horizontal relationships that form complicated dependence structures. Modeling these structures is challenging due to the stylized facts of market data. Many research works in recent decades showed that copula is an effective method to describe relations among variables. Vine structures were introduced to represent the decomposition of multivariate copula functions. However, the model construction of vine structures is still a tough problem owing to the geometrical data, conditional independent assumptions and the stylized facts. In this paper, we introduce a new bottom-to-up method to construct regular vine structures and applies the model to 12 currencies over 16 years as a case study to analyze the asymmetric and fat tail features. The out-of-sample performance of our model is evaluated by Value at Risk, a widely used industrial benchmark. The experimental results show that our model and its intrinsic design significantly outperform industry baselines, and provide financially interpretable knowledge and profound insights into the dependence structures of multi-variables with complex dependencies and characteristics.

Data science and big data analytics have led to next-generation economy innovation, competition and productivity [288], as typically shown by the rapidly updated big data landscape [26].

© 2017, Springer International Publishing AG. To predict customer’s next choice in the context of what he/she has bought in a session is interesting and critical in the transaction domain especially for online shopping. Precise prediction leads to high quality recommendations and thus high benefit. Such kind of recommendation is usually formalized as transaction-based recommender systems (TBRS). Existing TBRS either tend to recommend popular items while ignore infrequent and newly-released ones (e.g., pattern-based RS) or assume a rigid order between items within a transaction (e.g., Markov Chain-based RS) which does not satisfy real-world cases in most time. In this paper, we propose a neural network-based comprehensive transaction embedding model (NTEM) which can effectively perceive the next choice in a transaction context. Specifically, we learn these comprehensive embeddings of both items and their features from relaxed ordered transactions. The relevance between items revealed by the transactions is encoded into such embeddings. With rich information embedded, such embeddings are powerful to predict the next choices given those already bought items. NTEM is a shallow wide-in-wide-out network, which is more efficient than deep networks considering large numbers of items and transactions. Experimental results on real-world datasets show that NTEM outperforms three typical TBRS models FPMC, PRME and GRU4Rec in terms of recommendation accuracy and novelty. Our implementation is available at https://github.com/shoujin88/NTEM-model.

ML-\(k\)NN is a well-known algorithm for multi-label classification. Although effective in some cases, ML-\(k\)NN has some defect due to the fact that it is a binary relevance classifier which only considers one label every time. In this paper, we present a new method for multi-label classification, which is based on lazy learning approaches to classify an unseen instance on the basis of its \(k\) nearest neighbors. By introducing the coupled similarity between class labels, the proposed method exploits the correlations between class labels, which overcomes the shortcoming of ML-\(k\)NN. Experiments on benchmark data sets show that our proposed Coupled Multi-Label \(k\) Nearest Neighbor algorithm (CML-\(k\)NN) achieves superior performance than some existing multi-label classification algorithms.

Support vector machine is a supervised learning model with associated learning algorithms that analyzes data and recognizes patterns. In various applications, the SVM shows its advantage of the classification performance, however, the original SVM was designed for the numerical data. For using the SVM on the nominal data, most previous research used a certain number to replace each nominal value or transformed the nominal value into the one hot vector. Both methods could not present the original nominal data’s structure and the similarity between them, which leads to information loss from the data and reduce the classification performance. In this work, we design a novel coupled similarity metric between nominally attributed data. This metric is pairwise, we also propose an adapted SVM which can handle this. The experiment result shows the proposed method outperforms the traditional SVM and other popular classification methods on various public data sets.

© Springer International Publishing Switzerland 2015. Learning user/item relation is a key issue in recommender system, and existing methods mostly measure the user/item relation from one particular aspect, e.g., historical ratings, etc. However, the relations between users/items could be influenced by multifaceted factors, so any single type of measure could get only a partial view of them. Thus it is more advisable to integrate measures from different aspects to estimate the underlying user/item relation. Furthermore, the estimation of underlying user/item relation should be optimal for current task. To this end, we propose a novel model to couple multiple relations measured on different aspects, and determine the optimal user/item relations via learning the optimal way of integrating these relation measures. Specifically, matrix factorization model is extended in this paper by considering the relations between latent factors of different users/items. Experiments are conducted and our method shows good performance and outperforms other baseline methods.