We present a method called SketchyCoreSVD to compute the near-optimal rank r SVD of a data matrix by building random sketches only from its subsampled columns and rows. We provide theoretical guarantees under incoherence assumptions, and validate the performance of our SketchyCoreSVD method on various large static and time-varying datasets.
Federated RDF systems allow users to retrieve data from multiple independent sources without needing to have all the data in the same triple store. The performance of these systems can be poor for large and geographically distributed RDF data where network transfer costs are high. This paper introduces CBTP, a novel join algorithm that takes advantage of network topology to decrease the cost of processing SPARQL queries in a geographically distributed environment. Federation members are grouped in clusters, based on the network communication cost between the members, and the bulk of the join processing is pushed to the clusters. We use an overlap list to efficiently compute join results from triples in different clusters. We implement our algorithms in OpenRDF Sesame federated framework and use Apache Rya triple store instances as federation members. Experimental evaluation results show the advantages of our approach over existing techniques.
In recent years, the number of industry-4.0-enabled manufacturing sites has been continuously growing, and both the quantity and variety of signals and data collected in plants are increasing at an unprecedented rate. At the same time, the demand of Big Data processing platforms and analytical tools tailored to manufacturing environments has become more and more prominent. Manufacturing companies are collecting huge amounts of information during the production process through a plethora of sensors and networks. To extract value and actionable knowledge from such precious repositories, suitable data-driven approaches are required. They are expected to improve the production processes by reducing maintenance costs, reliably predicting equipment failures, and avoiding quality degradation. To this aim, Machine Learning techniques tailored for predictive maintenance analysis have been adopted in PREMISES (PREdictive Maintenance service for Industrial procesSES), an innovative framework providing a scalable Big Data service able to predict alarming conditions in slowly-degrading processes characterized by cyclic procedures. PREMISES has been experimentally tested and validated on a real industrial use case, resulting efficient and effective in predicting alarms. The framework has been designed to address the main Big Data and industrial requirements, by being developed on a solid and scalable processing framework, Apache Spark, and supporting the deployment on modularized containers, specifically upon the Docker technology stack.
Nowadays, data-centers are largely under-utilized because resource allocation is based on reservation mechanisms which ignore actual resource utilization. Indeed, it is common to reserve resources for peak demand, which may occur only for a small portion of the application life time. As a consequence, cluster resources often go under-utilized. In this work, we propose a mechanism that improves compute cluster utilization and their responsiveness, while preventing application failures due to contention in accessing finite resources such as RAM. Our method monitors resource utilization and employs a data-driven approach to resource demand forecasting, featuring quantification of uncertainty in the predictions. Using demand forecast and its confidence, our mechanism modulates cluster resources assigned to running applications, and reduces the turnaround time by more than one order of magnitude while keeping application failures under control. Thus, tenants enjoy a responsive system and providers benefit from an efficient cluster utilization.
With the rapid development of Chinese economy, it is significant to examine the economic activities in China. Each transaction behavior is recorded by the invoice. The invoice contains the transaction content, the classification of the transaction behavior (in accordance with the Tax Classification and Coding for Commodities and Services issued by the state) and transaction price, etc. Our work uses real mass invoice data collected from Zhejiang Province and conducts a multi-dimensional analysis of Chinese transaction behavior based on transaction behavior classification model. Firstly, we propose a compositional CNN-RNN model with attention mechanism to recommend the corresponding categories of transaction behavior collected from tax invoices. It maps the transaction behavior recorded in the invoice to transaction code in the Tax Classification and Coding for Commodities and Services issued by the state. Preliminary experiments show that the top-one accuracy of classifying transaction behavior achieves 75%. Then, we focus on the quantity distribution of invoice data and draw a conclusion that the major category with larger number of invoice records is more diversified in subdivided categories. After that, we studied the price distribution of various transaction behaviors to discover the difference in price distribution between different industries. Prices in the major categories of goods are more concentrated in the middle or lower prices. We can analyze the regional industrial structure through the price distribution of the industry which makes sense to study the economy of the region from the perspective of industry.
Evaluating the degradation of predictive models over time has always been a difficult task, also considering that new unseen data might not fit the training distribution. This is a well-known problem in real-world use cases, where collecting the historical training set for all possible prediction labels may be very hard, too expensive or completely unfeasible. To solve this issue, we present a new unsupervised approach to detect and evaluate the degradation of classification and prediction models, based on a scalable variant of the Silhouette index, named Descriptor Silhouette, specifically designed to advance current Big Data state-of-the-art solutions. The newly proposed strategy has been tested and validated over both synthetic and real-world industrial use cases. To this aim, it has been included in a framework named SCALE and resulted to be efficient and more effective in assessing the degradation of prediction performance than current state-of-the-art best solutions.
Rumor detection becomes a major issue concerning the public and government as the proliferation of social media in information dissemination. However, most existing methods only extract hand-crafted features, far from adequate in interpreting semantics latent in texts. For social events, there also exists rich social contextual information and highlevel interactions among significant features, which provides cues for interpreting semantics. In this paper, we propose a novel attention learning framework via deep visual perception based recurrent neural network (ViP-RNN), considering both high-level feature interactions and contextual information. In particular, the proposed model is based on RNN for capturing the long-distance temporal dependencies of contextual information of relevant posts and composing low-level lexical features into high-level semantic interactions hierarchically by visual perception of convolutional neural network (CNN). To incorporate information learned by RNN and CNN, we combine convolutional and recurrent layers into one model so that the model can capture a discriminative semantic representation of social events more efficiently by utilizing visual perception attention vector i.e. outputs of CNN to align long-distance temporal dependencies. We conduct experiments on real datasets collected from social media websites, which demonstrates the effectiveness of our approach and the merits of model integration.
The current explosion of data, which is impacting many different areas, is especially noticeable in biomedical research thanks to the development of new technologies that are able to capture high-dimensional and high-resolution data at the single-cell scale. Processing such data in an interpretable way often requires the computation of pairwise dissimilarity measures between the multiple features of the data, a task that can be very difficult to tackle when the dataset is large enough, and which is prone to numerical instability. In this paper we propose a distributed framework to efficiently compute dissimilarity matrices in arbitrarily large datasets in a numerically robust way. It implements a combination of the pairwise and two-pass algorithms for computing the variance, in order to maintain the numerical robustness of the former while reducing its overhead. The proposal is parallelizable both across multiple computers and multiple cores, maximizing the performance while maintaining the benefits of memory locality. The proposal is tested on a real use case: a dataset generated from high-content screening images composed by a billion individual cells and 786 features. The results showed linear scalability with respect to the size of the dataset and close to linear speedup.
With the growing reliance on the ubiquitous availability of IT systems and services, these systems become more global, scaled, and complex to operate. To maintain business viability, IT service providers must put in place reliable and cost efficient operations support. Artificial Intelligence for IT Operations (AIOps) is a promising technology for alleviating operational complexity of IT systems and services. AIOps platforms utilize big data, machine learning and other advanced analytics technologies to enhance IT operations with proactive actionable dynamic insight. In this paper we share our experience applying the AIOps approach to a production cloud object storage service to get actionable insights into system's behavior and health. We describe a real-life production cloud scale service and its operational data, present the AIOps platform we have created, and show how it has helped us resolving operational pain points.
Lexicon-based sentiment analysis is a popular and practical approach for sentiment analysis. However, sentiment lexicons, which may be abundant in some language such as English, are scarce in many other languages. The cross-lingual lexicon learning aims to extend lexicons for the language with less resources from those lexicons available in other languages. In this paper, we propose an approach that builds a skip-gram variant to map word spaces across languages so as to construct lexicons for the language with less resources. We show in our preliminary experiment that our approach can generate lexicons that are similar to those crafted by human experts.
Relation extraction is a critical task in building a knowledge base from unstructured text documents. Most works in automatic relation extraction have applied deep learning techniques such as Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) in large text corpora. However, they require a large amount of human labelling data, which is labour intensive and is hardly applied in a new domain of document without human supervision. This paper proposes a novel framework to extract relations in multi-domain texts effectively. In particular, we construct the framework in three phases including preprocessing, feature embedding and relation extraction. We show that a small proportion of training data is sufficient to train our relation extraction framework and achieve a good accuracy in relation extraction works.
We introduce Big Data Analytics Modeling Languages (BiDaML), a novel integrated suite of visual languages aimed at supporting end users during the process of designing big data analytics solutions. BiDaML comprises five diagrammatic notations: a data analytics brainstorming diagram; a process diagram; technique diagrams; data diagrams; and output diagrams. Tool support in the form of an integrated design environment for creating BiDaML diagrams has also been developed. To demonstrate the utility of BiDaML, we illustrate our approach with a real-world example of property price prediction and evaluate BiDaML using the physics of notations and cognitive walkthroughs with target end users - data scientists and software engineers.
This paper describes recent results achieved in the analysis of geo-distributed sensor data generated in the context of the energy sector. The approaches described have roots in the Big Data Analytics and Predictive Modeling research fields and are based on distributed architectures. They tackle the energy forecasting task for a network of energy production plants, by also taking into consideration the detection and treatment of anomalies in the data. This research is motivated by and consistent with the objectives of research projects funded by the European Commission and by many national governments.
Insights generated from Big Data through analytics processes are often unstable over time and thus lose their value, as the analysis typically depends on elements that change and evolve dynamically. However, the cost of having to periodically "redo" computationally expensive data analytics is not normally taken into account when assessing the benefits of the outcomes. The ReComp project addresses the problem of efficiently re-computing, all or in part, outcomes from complex analytical processes in response to some of the changes that occur to process dependencies. While such dependencies may include application and system libraries, as well as the deployment environment, ReComp is focused exclusively on changes to reference datasets as well as to the original inputs. Our hypothesis is that an efficient re-computation strategy requires the ability to (i) observe and quantify data changes, (ii) estimate the impact of those changes on a population of prior outcomes, (iii) identify the minimal process fragments that can restore the currency of the impacted outcomes, and (iv) selectively drive their refresh. In this paper we present a generic framework that addresses these requirements, and show how it can be customised to operate on two case studies of very diverse domains, namely genomics and geosciences. We discuss lessons learnt and outline the next steps towards the ReComp vision.
Digital Repository Systems have been used in most modern digital library platforms. Even so, Digital Repository Systems often suffer from problems such as low discoverability, poor usability, and high drop-off visit rates. With these problems, the majority of the content in the digital library platforms may not be exposed to end users, while at the same time, users are desperately looking for something which may not be returned from the platforms. The recommendation systems for digital libraries were proposed to solve these problems. However, most recommendation systems have been implemented by directly adopting one specific type of recommender like Collaborative-Filtering (CF), Content-Based Filtering (CBF), Stereotyping, or hybrid recommenders. As such, they are either (1) not able to accommodate the variation of the user groups, (2) require too much labor, or (3) require intensive computational complexity. In this paper, we design and implement a new recommendation system framework for Digital Repository Systems, named dpSmart, which allows multiple recommenders to work collaboratively on the same platform. In the proposed system, a user-group based recommendation strategy is applied to accommodate the requirements from the different types of users. A user recognition model is built, which can avoid the intensive labor of the stereotyping recommender. We implement the system prototype as a sub-system of the FIU library site (http://dpanther.fiu.edu) and evaluate it on January 2019 and February 2019. During this time, the Page Views have increased from 8,502 to 10,916 and 10,942 to 12,314 respectively, compared to 2018, demonstrating the effectiveness of our proposed system.
Privacy is a fundamental concern that confronts systems dealing with sensitive data. The lack of robust solutions for defining and enforcing privacy measures continues to hinder the general acceptance and adoption of these systems. Edge computing has been recognized as a key enabler for privacy enhanced applications, and has opened new opportunities. In this paper, we propose a novel privacy model based on context-aware edge computing. Our model leverages the context of data to make decisions about how these data need to be processed and managed to achieve privacy. Based on a scenario from the eHealth domain, we show how our generalized model can be used to implement and enact complex domain-specific privacy policies. We illustrate our approach by constructing real world use cases involving a mobile Electronic Health Record that interacts with, and in different environments.
Recently, deep learning has become one of the most disruptive trends in the technology world. Deep learning techniques are increasingly achieving significant results in different domains such as speech recognition, image recognition and natural language processing. In general, there are various reasons behind the increasing popularity of deep learning techniques. These reasons include increasing data availability, the increasing availability of powerful hardware and computing resources in addition to the increasing availability of deep learning frameworks. In practice, the increasing popularity of deep learning frameworks calls for benchmarking studies that can effectively evaluate the performance characteristics of these systems. In this paper, we present an extensive experimental study of six popular deep learning frameworks, namely TensorFlow, MXNet, PyTorch, Theano, Chainer, and Keras. Our experimental evaluation covers different aspects for its comparison including accuracy, speed and resource consumption. Our experiments have been conducted on both CPU and GPU environments and using different datasets. We report and analyze the performance characteristics of the studied frameworks. In addition, we report a set of insights and important lessons that we have learned from conducting our experiments.
Recommendation algorithms have been investigated and employed by many important companies in the past years: some scenarios, such as the one where a system suggests the points of interest to tourists, well adapt to sequence of recommendations to (groups of) users. We envision that sequence recommendations can be useful whenever the group of users has a limited time interval to spend together, since they reduce the time wasted in selecting the best next activity. In this paper, we investigate the role played by the context, i.e. the situation the group is currently experiencing, in the design of a system that recommends sequences of activities. We model the problem as a multi-objective optimization, where the satisfaction of the group and the available time interval are two of the functions to be optimized. In particular, the dynamic evolution of the group can be considered as the key contextual feature to produce better suggestions.
As the number and variety of services increase, it is becoming difficult and time-consuming to locate services that satisfy users' need. Service clustering is efficacious method to prune the query space, to narrow the searching space, and improve the accuracy of locating services that satisfied users' needs. At present, clustering method of web services adopted single or traditional clustering algorithms. However, accuracy and stability of single or traditional clustering algorithms is poor. In the paper, we proposed SWOC a service clustering method based on wisdom of crowd. Firstly, by using SWOC we calculated document similarity. Secondly, we implemented a mapping algorithm that reduces the correlation of web services and improve accuracy of method. And then, we applyed different number of clusters using different individual clustering methods that increase the number of partitions so as to enhance the robustness of SWOC. Lastly, the diversity algorithm evaluates and selects the partitions to extract interesting information for the final aggregation with the weight of each individual result. Experiments were performed on the real web service dataset crawled from ProgrammableWeb which prove the accuracy, recall, F-value and stability of proposed method.
Distributed hierarchical file systems typically decouple the storage of the file system's metadata from the data (file system blocks) to enable the scalability of the file system. This decoupling, however, requires the introduction of a periodic synchronization protocol to ensure the consistency of the file system's metadata and its blocks. Apache HDFS and HopsFS implement a protocol, called block reporting, where each data server periodically sends ground truth information about all its file system blocks to the metadata servers, allowing the metadata to be synchronized with the actual state of the data blocks in the file system. The network and processing overhead of the existing block reporting protocol, however, increases with cluster size, ultimately limiting cluster scalability. In this paper, we introduce a new block reporting protocol for HopsFS that reduces the protocol bandwidth and processing overhead by up to three orders of magnitude, compared to HDFS/HopsFS' existing protocol. Our new protocol removes a major bottleneck that prevented HopsFS clusters scaling to tens of thousands of servers.
