Academic exams

Leveraging Existing Technologies to Improve Large-Scale Recommender Systems


Recommender Systems have as goal providing valuable recommendations to its users. Most research on Recommender Systems aims to improve the recommendation quality exclusively, overlooking the computational efficiency of such solutions. Although Recommender Systems, based on collaborative filtering, do not have many ratings available in this work by strategically removing redundant ratings is able to offer a similarity metric with an improved the computational efficiency, for Recommender Systems. This work focus on improving the computational efficiency of similarity metrics and enhance quality, using two different approaches. The first relies on Collaborative Filtering i.e. exclusively on ratings, to produce a computationally efficient similarity metric for Recommender Systems. The second approach uses contextual information regarding users and items to further improve recommendation quality, while still maintaining the same computational efficiency. The solutions here proposed can be readily deployed on real Recommender Systems. The first approach methodology intends to improve similarity metrics for Memory-based Collaborative Filtering using Fuzzy Models. Memory-based Collaborative Filtering relies heavily on similarity metrics to produce recommendations. Fuzzy Fingerprints are used as a framework to create a novel similarity metric, providing a fast and effective solution. The second approach also uses Fuzzy Fingerprints, and it combines contextual information with ratings into a single Fuzzy Fingerprint, or create a multi-context Fuzzy Fingerprint where each contextual information has its own Fuzzy Fingerprint. Each contextual Fuzzy Fingerprint allows a ranking fusion algorithm to produce similarity values. This work is validated using four well-known datasets which are ML-1M, HetRec 2011, Netflix and Jester. The application of the Fuzzy Fingerprint similarity improves recommendation quality but, more importantly, requires four times less computational resources than current solutions on large datasets. When using contextual information the recommendation quality improves either by combining contextual information and ratings into a single Fuzzy Fingerprint or by using multi-context Fuzzy Fingerprints. The solutions using contextual information achieve further recommendation quality improvements while maintaining a comparable computational efficiency in comparison well-known similarity metrics.

access_time September 05, 2019 at 02:00PM
place Anfiteatro PA-3 (Piso -1, Pavilhão de Matemática), IST, Alameda
local_offer Doctoral exam
person Candidate: André Filipe Caldaça da Silva Carvalho
supervisor_account Advisor: Prof. Pável Pereira Calado / Prof. João Paulo Baptista de Carvalho

PerGUID: Personality-Based Graphical User Interface Design Guidelines


Individual differences play a major role in human-computer interaction. In particular, personality shapes how we process and act on the world, and how users perceive and accept technology. Nevertheless, there is limited evidence on the impact of different personality types in graphical user interface (GUI). Moreover, there is limited work on implicit personality assessment from biofeedback. To approach these issues, we propose the study and inclusion of psychological variables from the Five-Factor Model and the Locus of Control in GUI design to provide a better user experience (UX) with a personality-based adaptive GUI which detects psychological traits from biofeedback. Participants (N=100) will use scales NEO Personality Inventory Revised (NEO PI-R) and Levenson Multidimensional Locus of Control (LMLoC) for psychological evaluation, System Usability Scale (SUS), Technology Acceptance Model 3 (TAM3), and NASA Task Load Index (NASA-TLX) for UX assessment, and a Bitalino for biofeedback acquisition. A personality-based adaptive brain-computer interface carries the promise to improve UX design techniques and personality classification based on biofeedback by allowing designers better understand their audience while taking advantage of physiological computing to implicitly collect users’ psychological constructs.

access_time July 18, 2019 at 02:00PM
place Sala 0.19, Pavilhão Informática II, IST, Alameda
local_offer CAT exam
person Candidate: Tomás Almeida e Silva Martins Alves
supervisor_account Advisor: Prof.ª Sandra Gama / Prof. Daniel Gonçalves / Prof.ª Joana Calado

Complex networks analysis from an edge perspective


If we observe our daily lives and the systems in which we participate carefully, it is easy to see that everything is somehow connected. From species evolution to social relations, passing through all the supply chain systems we know, networks portray the simplest representation of these systems. Notwithstanding this simplicity, these networks often underlie complex dynamics. Species and populations evolution are subject to many complex interactions, and individuals states -- from individual choices, epidemic states, strategic behaviors, opinions, among others -- are influenced by social ties and by the overall topology of interaction. These networks, called complex networks, show a prevalence of certain features, which are shared between completely different systems, thus defying the limits of the traditional techniques of analysis and intriguing the research community. In this thesis we aim to contribute to the study of the relationship between structure and dynamics of these complex networks. Usually, the main approach to study complex networks is centered on the importance of nodes. However, it is our understanding that the edge-perspective analysis also provides fundamental and complementary information on the structure and behavior of complex networks. Given this, throughout this dissertation we approach complex networks under an edge perspective, centering our attention in the properties of the edges. In our contributions we provide new metrics, models and computational tools. We start by contributing with a new edge centrality measure. Next, we focus on analyzing local patterns/subgraphs whose edges contain informative labels, highlighting that sometimes observing only nodes and edges, individually, is not enough to fully understand the dynamics and/or the structure of a system. Finally, we observe that often representing a system with a single network is insufficient to reproduce its behavior, being necessary to consider networks at multiple scales, i.e. networks of networks. Our contribution in this subject is a new computational tool that allows us to model and simulate a system represented as a network of networks.

access_time July 02, 2019 at 02:30PM
place Sala 4.41, 2.º Piso do Pavilhão de Civil, IST, Alameda
local_offer Doctoral exam
person Candidate: Andreia Sofia Monteiro Teixeira
supervisor_account Advisor: Prof. Alexandre Paulo Lourenço Francisco / Prof. Francisco João Duarte Cordeiro

Software-Defined Systems for Network-Aware Service Composition and Workflow Placement


Composing and scheduling workflows at Internet scale require communication and coordination across various services in heterogeneous execution environments - from data centers and clouds to the edge environments operated by multiple service providers. Services are diverse and inclusive of several variants such as web services, network services, and data services. Service description standards and protocols focus on interoperability across service interfaces, to enable workflows spanning various service providers. Nevertheless, in practice, standardization of the interfaces remains mostly limited. Furthermore, efficient resource provisioning for workflows of several users from multiple infrastructures requires collaboration and cooperation of the infrastructure providers. The current approaches are limited in scalability and optimality in efficiently provisioning resources for user workflows spanning numerous infrastructure providers. Network Softwarization revolutionizes the network landscape in various stages, from building, incrementally deploying, and maintaining the environment. Software-Defined Networking (SDN) and Network Functions Virtualization (NFV) are two core tenets of network softwarization. SDN offers a logically centralized control plane by abstracting away the control of the network devices in the data plane. NFV virtualizes dedicated hardware middleboxes and deploys them on top of servers and data centers as network functions. Thus, network softwarization enables efficient management of the system, by enhancing its control and improving the reusability of the network services. In this work, we aim at exploiting network softwarization to compose workflows of distinct services, in heterogeneous infrastructures, ranging from data centers to the edge. We thus intend to mitigate the challenges concerning resource management and interoperability of heterogeneous infrastructures, to efficiently compose service workflows, while sharing the network and the computing resources across several users. To this end, we propose three significant contributions. First, we extend SDN in cloud and data center environments to unify various phases of development and deploy the workloads seamlessly, from simulations and emulations to physical deployment environments. We further extend this work to support multiple Service Level Agreements (SLAs) across diverse network flows in the data centers, by selectively enforcing redundancy on the network flows. Thus, we aim for Quality of Service (QoS) and efficient resource provisioning, while adhering to the policies of the users. Finally, we design a cloud-assisted overlay network, as a latency-aware virtual connectivity provider. Consequently, we propose cost-efficient data transfers at Internet scale, by separating the network from the infrastructure. Second, we propose a scalable architecture to compose service chains in wide area networks efficiently. We extend SDN and Message-Oriented Middleware (MOM), for a logically centralized composition and execution of service workflows. We thus propose a Software-Defined Service Composition (SDSC) framework for web service compositions, Network Service Chains (NSCs), and a network-aware execution of data services. We further present Software-Defined Systems (SDS) consisting of virtual network allocation strategies for multi-tenant service executions in large-scale networks comprised of multiple domains. Third, we investigate how our proposed SDS can operate efficiently for real-world application scenarios of heterogeneous infrastructures. While traditionally web services are built following standards and best practices such as Web Services Description Language (WSDL), network services and data services offered by different service providers often fall short in providing common Application Programming Interfaces (APIs), often resulting in vendor lock-in. We look into facilitating interoperability across service implementations and deployments, to enable seamless migrations. We propose big data applications and smart environments such as Cyber-Physical Systems (CPS) and the Internet of Things (IoT) as our two application scenarios. We thus build CPS and big data applications as composable service chains, offering them an interoperable execution.

access_time July 01, 2019 at 03:00PM
place Sala 0.17, Pavilhão de Informática II do IST, Alameda
local_offer Doctoral exam
person Candidate: Pradeeban Kathiravelu
supervisor_account Advisor: Prof. Luís Manuel Antunes Veiga

Intrusion Recovery in Cloud Computing


Intrusion prevention mechanisms aim to reduce the probability of intrusions to occur. However, sooner or later an attacker may succeed in exploiting an unknown vulnerability or by stealing a user's access credentials, leading to the execution of undesired operations. These intrusions may corrupt the state of the application requiring intrusion recovery mechanisms to revert the effect of these actions. Simple solutions such as, the use of backups, allow reverting the effects of an intrusion, however, by restoring a previous backup of the system, every legitimate data that is not present in that backup is lost. Intrusion recovery mechanisms aim to revert only the damage caused by intrusions without affecting the legitimate data that was created by authorized users. This thesis explores the problem of intrusion recovery for distributed systems running in the cloud. The presented mechanisms take into account the distribution of the systems, the limitations of the cloud services and the development paradigms for this kind of systems. Given the heterogeneity of the different cloud computing models, there were designed different intrusions recovery mechanisms for the different models. For the infrastructure level of the cloud, namely, for file storage we propose RockFS, an intrusion recovery system designed for cloud-backed file systems. This kind of systems is accessed remotely allowing attackers to illegally modify files by accessing a legitimate user's account. RockFS protects the access credentials of the user through secret sharing mechanisms. which allow distributing fragments of the access credentials through several storage devices. For recovery, RockFS allows reverting unintended actions through the combination of operation logs and a multiversioned of file system. RockFS runs on the client-side and can be used in single-cloud and cloud-of-clouds file systems. At the infrastructure level of the cloud, but for databases, we present NoSQL Undo, an intrusion recovery system for NoSQL databases. This system does not require modifications to the source code of the Database Management System (DBMS), making it possible to be adopted by DBMS that do not provide the source code. NoSQL Undo takes advantage of the logs used by the database for replication, reducing the performance overhead. NoSQL Undo runs on the client-side and can be used to recover from intrusions without a previous installation or configuration. NoSQL Undo provides two algorithms for recovery: focused recovery and full recovery. Focused recovery only fixes the database records that were affected by the attack, while the full recovery fixes the entire database. The use of one algorithm as opposed to the other depends on the amount of affected database records by the attack. At the application level, of the cloud computing model, we propose an intrusion recover system called Rectify that allows reverting the effects of the attack in web applications. It is possible to use Rectify in any web application that uses a SQL database to store its state. Rectify identifies malicious operations in the database that were generated by malicious requests performed by the application. This association, of database operations with application level requests, is done through machine learning algorithms. The main advantage of this technique is that it does not require modifying the source code of the application or the source coed of the database. Rectify allows recovering from intrusions while keeping the application available for users. For modern distributed applications running in the cloud developed in a microservices paradigm, we propose µVerum, an intrusion recovery system that adopts the microservices architecture. In this kind of applications each component of the system is distributed in independent services that interact with each other through the network. µVerum was designed taking into account the distribution and self-contained characteristics of each microservice. µVerum allows propagating the compensating operations, the revert the effects of the intrusion, on the affected services. µVerum presents a modular architecture that allows to replicate the components with higher traffic demands in order to maintain the performance level of the application. µVerum allows the developers of the application to define invariants in order to fulfill the consistency requirements of the application. These invariants can be of two types: atomicity, in which several microservices should be executed together; and ordering, in which a group of microservices should be executed in a specific order. The work presented in this thesis allows applications deployed in the cloud to overcome intrusions. The proposed systems can be deployed in existing cloud services and interfere as less as possible with the user experience.

access_time June 27, 2019 at 02:00PM
place Anfiteatro PA-3 (Piso -1 do Pavilhão de Matemática) do IST, Alameda
local_offer Doctoral exam
person Candidate: David Rogério Póvoa de Matos
supervisor_account Advisor: Prof. Miguel Nuno Dias Alves Pupo Correia / Prof. Miguel Filipe Leitão Pardal

Filter by period  filter_list