Current ICT Research Projects

Supervisors: Dr. Mohammad Awrangjeb and Professor Bela Stantic

Description: The speedy development in electricity infrastructure due to urge in domestic and business usage as well as its importance in national economy requires a safe and secure maintenance of power line corridors (PLC) to ensure the efficient and uninterrupted power supply of electricity to consumers. The monitoring of PLC primarily includes two of the following aspects: electrical components such as power lines and pylons and surrounding objects, such as vegetation. For reliable transmission, the stability of power lines and pylons and monitoring of vegetation near PLC is important.

As power lines are comprised of very thin conductors, thus detailed information is required for accurate mapping. Airborne light detection and ranging (LiDAR) has been proven a powerful tool to overcome these challenges to enable more efficient inspection in recent years. Active airborne LiDAR systems directly capture the 3D information of power infrastructure and surrounding objects. Nevertheless,

PLCs are built with multi-loop, multi-phase structures (bundle conductors) and exists in intricate environments (e.g., mountains and forests), thus raises challenges to process airborne point cloud data for extraction and modelling of individual PLC objects.

This study aims to overcome these challenges by providing an automated and more robust solutions for PLC mapping. This research incorporates three main objectives; (i) power lines extraction, pylons and vegetation extraction, (ii) reconstruction of power lines and pylons using for 3D modelling, (iii) vegetation monitoring from airborne LiDAR data.

Related publications

ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci. (Google Scholar Metrics (GSM) -Rank: 10 in Remote Sensing, GSM makes only 20 top cited in each area combining both conference and journal articles.)

DICTA 2019 (Australian in Core 2018)

Supervisors: Dr. Mohammad Awrangjeb and Professor Bela Stantic

Description: Building extraction with individual roof parts and other components such as chimneys and dormers is important for building reconstruction and 3D modelling. Using Light Detection and Ranging (LiDAR) point-cloud data the task is more complex and difficult because of the unknown semantic characteristics and inharmonious behaviour of the LiDAR input data. Most of the existing state-of-the-art methods fail to detect small true roof planes with exact boundaries due to outliers, occlusions, complex building structures, and other inconsistent nature of LiDAR data thus, accurate building detection, reconstruction, and 3D modelling a challenging and complex task. Studies have been conducted over the last two decades on individual building extraction and reconstruction using LiDAR data. The main objective of this PhD thesis is to extract buildings and individual roof parts effectively using LiDAR data for the purpose of 3D reconstruction and modelling of buildings.

Related publications

Dey, E. K., Awrangjeb, M., & Stantic, B. (2019, July). An Unsupervised Outlier Detection Method For 3D Point Cloud Data. In IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium (pp. 2495-2498). IEEE.

Dey, E. K., Awrangjeb, M., & Stantic, B. (2020). Outlier detection and robust plane fitting for building roof extraction from LiDAR data. International Journal of Remote Sensing, 41(16), 6325-6354.

Dey, E. K. and Awrangjeb, M., "A Robust Performance Evaluation Metric for Extracted Building Boundaries From Remote Sensing Data," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 13, pp. 4030-4043, 2020, doi: 10.1109/JSTARS.2020.3006258.

Supervisors: A/Prof. Alan Wee-Chung Liew

Description: Continual learning (CL) or lifelong learning is the ability of a model to learn continually from a stream of data. The idea of CL is to mimic human’s ability to continually acquire, fine-tune, and transfer knowledge and skills throughout their lifespan. With CL, we want to use the data that is coming to update the model autonomously based on the new activity. Data are typically discarded after use, and there is no opportunity to re-use the data for model retraining. Continual learning is a challenge for deep neural network models since the continual acquisition of incrementally available information from non-stationary data distributions generally leads to catastrophic forgetting or interference. Other challenges in CL includes adapting to emerging and disappearing concepts, adapting to concept drift, adapting to nonstationary noise, dealing with highly imbalance classes, etc. This project aims to develop novel (supervised and unsupervised) machine learning algorithms that overcome these challenges.

Related publications

T.T. Nguyen, M.T. Dang, V.A. Luong, A.W.C. Liew, T.C. Liang, J. McCall, “Multi-Label Classification via Incremental Clustering on Evolving Data Stream”, Pattern Recognition, Vol. 95: 96-113, 2019.

T.T. Nguyen, T.T.T. Nguyen, V.A. Luong, N.Q.V. Hung, A.W.C. Liew, B, Stantic, “Multi-label classification via labels correlation and first order feature dependence on data stream”, Pattern Recognition, Vol. 90: 35-51, 2019.

T.T.T. Nguyen, T.T. Nguyen, A.W.C. Liew, S.L. Wang, “Variational Inference based Bayes Online Classifiers with Concept Drift Adaptation”, Pattern Recognition, Vol. 81: 280-293, 2018.

Supervisors: Dr. Gervase Tuxworth

Description: Recognising objects in images is an important task for many applications including security, autonomous navigation and image tagging and markup. Recently the field has been dominated by convolutional neural networks, with some networks reaching sizes of over 100 million parameters. These networks are typically run on specialised hardware that consumes a high amount of power, but when considering applications running on light-weight low-cost hardware, these solutions may not be suitable. This project seeks to find solutions to allow for accurate object detection on low powered devices.

Related publications

Shaikh D, Manoonpong P, Tuxworth G, Bodenhagen L. Multi-sensory guidance of goal-oriented behaviour of legged robots. Proceedings of the 20th International Conference on Climbing and Walking Robots and the Support Technologies for Mobile Machines, CLAWAR 2017.

Supervisors: Dr. Gervase Tuxworth

Description: Fine-grained image classification is a challenge in computer vision, which aims at identifying the correct object in a dataset where there is both low between-class variance (different objects appear visually similar) and high intra-class variance (objects of the same class appear different). This work looks at implementing new models and techniques within convolutional neural networks to improve performance in these challenging datasets.

Related publications

Park YJ, Tuxworth G, Zhou J. Insect Classification Using Squeeze-and-Excitation and Attention Modules - a Benchmark Study. IEEE International Conference on Image Processing, 2019.

Supervisors: A/Prof. Jun Zhou

Description: Hyperspectral videos contains rich spectral, spatial, and temporal information. Traditional methods treat these domains separately to undertake video analysis tasks, ignoring the intrinsic relationship embedded in the cross-modal data space. In this project, we propose to develop joint spectral-spatial-temporal processing methods to fully explore the abundant information embedded in hyperspectral videos. Fundamental theories and methods will be developed based on physics and statistical models and will be powered by the latest deep learning approaches. A number of applications in environment, agriculture, and medicine will be used to showcase the usefulness of the methods.

Related publications

Fengchao Xiong, Jun Zhou, and Yuntao Qian. Material based object tracking in hyperspectral videos, IEEE Transactions on Image Processing, Vol 29, No. 1, pages 3719-3733, 2020.

Suhad Lateef Al-khafaji, Jun Zhou, Ali Zia and Alan Wee-Chung Liew. Spectral-spatial scale invariant feature transform for hyperspectral images. IEEE Transactions on Image Processing, Vol. 27, No. 2, pages 837-850, 2018.

Supervisors: A/Prof. Jun Zhou

Description: Object detection and recognition is a fundamental task for microscopic imaging. It’s applications range from disease detection, cell recognition to microplastic classification. Traditional detection and recognition techniques are based on images captured in the visible light wavelength, limits the discrimination capability of systems deployed for complex microscopic imaging environment. Hyperspectral images contain light wavelength indexed reflectance from objects, therefore, enable the capability of material detection that is essential for many real-world tasks. This project provides unique opportunities to work with cross-disciplinary researchers in medical and environmental areas. The goal is to develop innovative technologies that can revolutionise the current microscopic imaging practice.

Related publications

Chee Meng Ho, Qi Sun, Adrian Teo; David Wibowo, Yongsheng Gao, Jun Zhou, Yanyi Huang, Say Hwa Tan, and Chun-Xia Zhao. Development of a microfluidic droplet-based microbioreactor for microbial cultivation. ACS Biomaterials Science & Engineering, Vol. 6, No. 6, pages 3630-3637, 2020.

Yanyang Gu, Zongyuan Ge, Paul Bonnington, and Jun Zhou. Progressive transfer learning and adversarial domain adaptation for cross-domain skin disease classification. IEEE Journal of Biomedical and Health Informatics, Vol. 24, No. 5, pages 1379-1393, 2020.

Jie Liang, Jun Zhou, Lei Tong, Xiao Bai and Bin Wang. Material based salient object detection from hyperspectral images. Pattern Recognition, Vol. 76, Pages 476-490, 2018.

Supervisors: A/Prof. Ernest Foo

Description: Industrial Control systems use SCADA protocols to control the electricity grid or water treatment plants or other critical infrastructure.  Many of these systems are being connected to the Internet and are vulnerable to cyber attacks.  This project will employ machine learning and artificial intelligence to automatically detect attacks against these systems and automate the best response for defense.

Related publications

IEEE Transactions on Industrial Informatics, IEEE Transactions on Information Forensics and Security, Computers & Security

Supervisors: A/Prof. Ernest Foo

Description: Next generation manufacturing systems use advanced robotic technologies and complex processes to function.  However many of these systems are connected to the Internet and are vulnerable to cyber attacks.  Stealthy cyber attacks are often difficult to detect.  This project will develop algorithms to monitor system processes for anomalies to automatically detect faults and cyber attacks.

Related publications

IEEE Transactions on Industrial Informatics, IEEE Transactions on Information Forensics and Security, Computers & Security, IEEE Access

Supervisors: A/Prof. Ernest Foo

Description: Driver-less vehicles and Intelligent Transport Systems need to use wireless communications to function with safety.  However these communications may be vulnerable to cyber attacks that allow attackers to manipulate traffic and cause accidents. This project will explore new ways to ensure efficient authentication to detect and prevent attacks against vehicle communication systems.

Related publications

IEEE Transactions on Industrial Informatics, Vehicular Communications, IEEE Transactions on Vehicular Technology

Supervisors: Dr. Qinyi Li

Description: Our daily digital life is protected by public-key cryptosystems like public-key encryption and digital signature systems. The security of most public-key cryptosystems have been deployed is ultimately based on the difficulties of solving number-theoretic problems (e.g., integer factoring problem and discrete logarithm problem) using classic computers. It turns out these number-theoretic problems can be efficiently solved by large-scale quantum computers which have been theorised about for decades. There has been substantial progress towards making quantum computing practical. To protect our communication in the long-term, we need a new generation of cryptosystems to defeat quantum computers. Cryptography based on decoding problems (e.g., decoding random linear codes) is a very promising candidate. In this project, you will explore the field of post-quantum cryptography and conduct research on one the two directions: 1) designing advanced post-quantum cryptosystems e.g., attributed-based encryption, functional encryption, fully homomorphic encryption, ring/group signatures and apply them to the real-world problems, e.g., fine-grained access control on encrypted data for cloud computing, efficient search and query on the encrypted database, smart contract and cryptocurrency 2) designing and implementing (in software or hardware) practical public-key encryption and digital signature systems with strong practical security (i.e., secure against various side-channel attacks) and high practicality (i.e., can be used for the Internet security protocols or computing-resource-restricted devices like IoT devices).

Related publications

Xavier Boyen, Malika Izabachene, Qinyi Li (Corresponding Author): An Efficient Lattice CCA-Secure KEM in the Standard Model. The 12th International Conference on Security and Cryptography for Networks (SCN 2020). Accepted on 14 June, 2020.

Xavier Boyen, Qinyi Li (Corresponding Author): Direct CCA-Secure KEM and Deterministic PKE from Plain LWE. The 10th International Conference on Post-Quantum Cryptography (PQCrypto 2019). LNCS 11505, pp.116-130. Springer 2019.

Xavier Boyen, Qinyi Li (Corresponding Author): All-but-Many Lossy Trapdoor Functions from Lattices and Applications. The 37th International Cryptology Conference (Crypto 2017). LNCS 10403, pp. 298-331, Springer 2017.

Xavier Boyen, Qinyi Li (Corresponding Author): Towards Tightly Secure Lattice Short Signature and Id-Based Encryption. The 22nd International Conference on Theory and Applications of Cryptography and Information Security (AsiaCrypt 2016). LNCS 10032, pp. 404-434. Springer 2016.

Supervisors: Dr. Wee Lum Tan

Description: There is great potential in applying machine learning techniques to design self-organising, self-aware, intelligent wireless networks. Machine learning enables network nodes to actively learn the state of the wireless environment, detect correlations in the data, and take actions to optimise network operations and make efficient use of the limited wireless spectrum resources.

The first project will develop methods to parse the massive amount of wireless network statistics/data (e.g. channel state information, signal strength, interference, noise, traffic load/patterns, etc.) in order to analyse and predict the context of the wireless environment. Using these data, we will develop machine learning-guided techniques to address a variety of challenges in wireless networks such as power control, user traffic scheduling, spectrum management, rate selection, etc.

A major challenge of machine learning is its vulnerability to adversarial attacks. Adversarial machine learning attacks in wireless networks can cause network nodes to make incorrect decisions or interfere with data transmissions. For example, network nodes can train a classifier on various wireless statistics and use it to predict future channel availability status and adapt their transmission decisions to the spectrum dynamics. An adversary can train its classifier to be functionally equivalent to the one at the transmitter, and launch attacks (e.g. sends jamming signals) when it predicts that the transmitter will transmit data to the receiver. These attacks can significantly affect network performance, e.g. reduced spectral efficiency and increased node energy consumption.

Therefore, a second project is to investigate the impact of different machine learning vulnerabilities in wireless networks and develop techniques to detect and mitigate these attacks in highly dynamic wireless networks.

Supervisors: Professor Vladimir Estivill-Castro

Related publications

David F. Nettleton, Vladimir Estivill-Castro, Julián Salas. Privacy in Multiple On-line Social Networks - Re-identification and Predictability. Trans. Data Priv. 12(1): 29-56 (2019)

Supervisors: Dr. Zhe Hou

Description: Machine learning is a subset of artificial intelligence that is focused on building mathematical models based on sample data, and making predictions without explicitly being programmed to perform the task. Machine learning has been used in data analytics for insurance, sports, tourism, marketing and many other areas. However, most existing machine learning algorithms often give excellent prediction results without telling the user how the decisions are made. This weakness results in trust issues from the user and limitations for adopting machine learning in some applications. To realise white-box machine learning, we propose to develop a new prediction model analysis method based on automated reasoning that systematically extracts logical explanations from prediction models and presents them in a way that users can easily understand. We will then leverage my previous experience in formal verification to convert prediction model into logical model and verify it against user specifications. Finally, we will develop new learning algorithms that can train correct-by-construction prediction models with respect to user specifications.

Supervisors: Dr. Zhe Hou

Description: There is a tremendous gap between today’s AI systems and the requirements in mission-critical applications. Improving reliability, safety, and security of AI decision-making is of paramount importance. These challenges drove us to develop new AI decision-making techniques which are safer and more secure. Particularly, we propose to integrate formal verification and bio-inspired optimisation techniques into (deep) reinforcement learning (RL) in order to provide a higher level of safety and security guarantees. There are three main modules for the proposed work. The first module concerns the development of new reliable optimisation algorithms that are suitable to be used as the core for reinforcement learning. The second module is about designing an efficient safe reinforcement learning algorithm using PAT and reliability-based optimisation. The third module is an application of the previous two in the scenario of cyber-physical attacks. We propose to extend the previous two modules with an adversarial deep reinforcement learning approach to train a more secure system. Finally we will apply the developed techniques in medical decision-making case studies such as the usage of respirator for COVID-19 treatments and drug dosage analysis.

Supervisors: Dr. Saiful Islam

Description: The recent growth of social media data opens-up a potentiality for automated systems to collect, process and analyse user generated data on causality. Automated discovery of causality detects the relationship between a cause and the corresponding effect. The discovered causality related information can be directly applied to several applications including automatic question answering, security and prescriptive event analysis. For instance, can we conclude that “lack of communication” caused “a disruption in bus service in Gold Coast” from the tweet “A disruption in bus service in Gold Coast due to lack of communication between translink and event organizers” posted by a user in twitter? Automated discovery of causality in social media data is not straightforward. Rather, it is a very challenging problem due to the unstructured, informal, and diverse nature of social media data. In this project, we aim to tackle this issue by developing an autonomous intelligent system that will collect and process social media data, develop transfer-learning based artificial intelligence (AI) models and algorithms to detect text causality in social media data. Some of our preliminary works have already been accepted by the community and published in the top venues of data mining and AI fields.

Supervisors: Dr. Qinyi Li

Description: Machine learning (ML) allows computer systems to train themselves to improve their performance. It is pervasive and plays a key role in a wide range of applications. At a high level, ML consists of two phases. In the first phase, it applies a learning algorithm to a set of training data drawn from some unknown distribution to generate a model (hypothesis). In the second phase, the model can be used to explain new data (e.g., classify new data from the unknown distribution, or generate new data from a distribution that is close to the unknown distribution). In many applications of ML, sensitive data is needed and therefore data privacy becomes a concern. For example, when comes to Machine Learning As a Service, remote entities (usually untrusted) provide access to machine learning algorithms using the Internet to user’s data and return the results. User’s data might be completely exposed to the remote entities if security/privacy mechanisms are not imposed. Also, even with the best privacy on the training data, output (in cleartext form) of the second phase of ML may reveal information on training data. Therefore, with ML is being applied ubiquitously, a set of techniques that protect data privacy in ML is desirable and important. In this project, you will closely analyse the data privacy issues in the context of ML and explore advanced cryptographic and privacy techniques (e.g., fully homomorphic encryption, secure multi-party computation and differential privacy) to provide innovative and practical solutions.

Supervisors: A/Prof. Alan Wee-Chung Liew

Description: This project aims to develop novel stream learning algorithms for continuous patient outcome monitoring and prognosis by taking into account patient's data collected during hospital admission. The algorithms are expected to integrate high frequency time series data with patient's demographic data, lab test data, diagnosis data, prescription data, etc. as exemplified in MIMIC-III, for accurate patient outcome monitoring and prognosis. This will in turn used to inform hospital resource planning and allocation using for example, our highly efficient binary QP solver [1]. Practical issues such as data sparsity, noisy and missing data, data non-stationarity, data leakage, prediction bias, model explainability, etc. will be investigated.

Related publications

B.S.Y. Lam, A.W.C. Liew, “A Fast Binary Quadratic Programming Solver based on Stochastic Neighborhood Search”, IEEE Trans on Pattern Analysis and Machine Intelligence, 2020. DOI: 10.1109/TPAMI.2020.3010811

Supervisors: Dr. Hui Tian

Description: Along with the advances of computing and network technologies, applying AI and machine learning techniques to analyse various types of big data from heterogeneous sources has become a major form of data processing and analysis. However, privacy leakage in accessing, processing and analysing shared (published) data is a major concern that obstacles the development of big data analytics. There have been numerous example of shocking damages and losses - both political and financial - caused by privacy breaches in different scales.

In order to safeguard data sharing for the purpose of big data analytics required by our industry and business, in the project we will investigate effective models, methods and techniques for privacy protection in data publishing, processing and analysis. For data publishing, we will study both cryptographic and non-cryptographic techniques including block cypher, randomization and anonymization to achieve effective protection of different type of data. For data processing, we will study effective privacy-preserving computing techniques including secure multi-party computation (SMC) and differential privacy. We will apply them in a cloud environment on virtualized network and computing resources. For data analysis, we will embed privacy-preserving techniques into machine learning models to achieve secure machine learning on big data.

Project outcomes will benefit both researchers and practitioners in big data analytics, machine learning, cloud computing and social network analysis, and potentially result significant economic gain for Australia's network-centric industry and business.

Related publications

Hui Tian, Wenwen Sheng, Hong Shen, Can Wang. Truth Finding by Reliability Estimation on Inconsistent Entities for Heterogeneous Data Sets. Knowledge-Based Systems, Jul. 2019. (CORE B, IF 5.921)

Hui Tian, Jingtian Liu and Hong Shen. Diffusion Wavelet-based Privacy Preserving in Social Networks. Computers & Electrical Engineering, Feb. 2018. (CORE B, IF 2.663)

Ruoxuan Wei, Hui Tian and Hong Shen. Improving k-Anonymity Based Privacy Preservation for Collaborative Filtering. Computers & Electrical Engineering, Mar. 2018. (CORE B, IF 2.663)

Supervisors: Dr. Can Wang

Description: This project aims at building a series of efficient recommender systems with high accuracy from social networks, such as Twitter, Facebook, Instagram, Netflix, and so on. The research questions may include how to quantify the coupling relationships in recommender systems from different levels, how to enhance the interpretability of recommender systems, how to involve the trend information and how to model trust in various recommendation problems, how to speed up the recommendation process but with acceptable accuracy, and etc.

Related publications

Can Wang, Chi-Hung Chi, Zhong She, Longbing Cao, Bela Stantic. Coupled Clustering Ensemble by Exploring Data Interdependence. ACM Transactions on Knowledge Discovery from Data, Vol. 12, No. 6, Article 63, pages 1-38, 2018. [Impact Factor 2.538, Q1]

Can Wang, Xiangjun Dong, Fei Zhou, Longbing Cao, Chi-Hung Chi. Coupled Attribute Similarity Learning on Categorical Data. IEEE Transactions on Neural Networks and Learning Systems, Vol. 26, No. 4, pages 781-797, 2015. [Impact Factor: 11.683, Q1]

Zhe Liu, Lina Yao, Lei Bai, Xianzhi Wang, Can Wang. Spectrum-Guided Adversarial Disparity Learning. The 2020 ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Accepted by KDD 2020). [CORE Ranking: A*]

Ye Tao, Can Wang, Lina Yao, Weimin Li, Yonghong Yu. TRec: Sequential Recommender Based On Latent Item Trend Information. International Joint Conference on Neural Networks (IJCNN 2020), pp. 1-8, 2020. [CORE Ranking: A]

Yunwei Zhao, Can Wang, Chi-Hung Chi, Kwok-Yan Lam, Sen Wang. A Comparative Study of Transactional and Semantic Approaches for Predicting Cascades on Twitter, The 27th International Joint Conference on Artificial Intelligence (IJCAI 2018), pages 1212-1218, 2018 [CORE Ranking: A*]

Supervisors: A/Prof. Junhu Wang

Description: Graphs are increasingly being used to model complex data, and collections of graphs are getting very large, which brings big challenges to query processing. On one hand, many queries in graph databases are expensive by nature, and computing their exact answers can be infeasible when the graph size is large. On the other, in many applications an error-bounded estimate will suffice.  These motivates the work on approximate query answering in large graphs.

This project will investigate approximate query answering in large dynamic graphs where nodes and edges can be frequently updated. We will focus on property graphs where the nodes (and/or edges) are associated with key-value pairs, and queries that may involve simple aggregation (e.g., counting the number of occurrences of substructures), and develop novel techniques to efficiently find high-quality approximate answers.

The approaches will generally involve offline pre-processing (e.g., summarization, smart indexing), algorithm design, and experimental evaluation.  Due to the dynamic nature of the graphs, any auxiliary data structures need to be efficiently maintainable, and ideally incremental computation of query answers will be explored.  We are particularly interested in summarization-based techniques and applying machine-learning in auxiliary structure construction.

Related publications

Xuguang Ren and Junhu Wang: Exploiting Vertex Relationships in Speeding up Subgraph Isomorphism over Large Graphs. VLDB 2015.

Xuguang Ren and Junhu Wang: Multi-Query Optimization for Subgraph Isomorphism Search. VLDB 2017.

Supervisors: A/Prof. Junhu Wang

Description: Knowledge graphs are tremendously popular nowadays because its ability to model diverse information. A knowledge graph can be regarded as a repository of facts about objects and their relationships, represented as labelled edges of a directed graph.  Over the last few years there have been growing interest in industry and academia to develop natural language question-answering (NLQL) systems over large knowledge graphs.  Such systems typically consists of two parts: question understanding and answer searching. Question understanding is to figure out the precise intention of the question, and answer searching is to actually find the answers based on the search intention.  Both tasks are challenging because of the ambiguity of natural language sentences and the fact that the same question an be raised in multiple ways in natural languages, and large size of knowledge graphs.

Existing approaches, whether based on question templates, machine-learning and graph embedding, or subgraph matching, suffer from limited capability in terms of the question types they can handle (i.e., they are limited to simple questions), accuracy, and efficiency. This PhD project will investigate NLQA over large knowledge graphs, with the aim of developing novel techniques to address the above limitations.

Related publications

Xiangnan Ren, Neha Sengupta, Xuguang Ren, Junhu Wang, Olivier Cur. Finding Structurally Compact Subgraphs with Ontology Exploration in Large RDF data (under review by PVLDB).