Tentative Program

Kickoff for ORAIN 2023 - OR and AI talks

Introduction to School of Information and Communication Technology (SoICT) 

Assoc. Prof. Ta Hai Tung, Dean of SoICT, HUST

Operations management and research at LIMOS -

An illustration on a Pick-up and delivery problem with cooperative robots

Prof. Jean-Philippe Gayon (LIMOS)

Abstract. In this talk, we will present the LIMOS research team "Decision-support Tools for Production and Services". We will illustrate its activities through several research examples, with a focus on a Pick-up and delivery problem with cooperative robots. In this problem, a set of robots has to perform a set of pick-up and delivery tasks. Each transportation task requires several robots and the objective is to achieve all tasks in a minimum time. 

Vehicle routing problem with drones 

Dr. Nguyen Khanh Phuong (SoICT - HUST)

Abstract. In the past decade, technological advances in robotics have significantly promoted the application of drones in many fields of life, such as logistics, healthcare, disaster management,

agriculture, and surveillance. In particular, the use of drones in last-mile pickup and delivery has witnessed a remarkable growth of interest in research due to the tremendous popularity of e-commerce. In this talk, we focus on vehicle routing problems with drones, and survey different types of integration of drones into last-mile delivery that have been investiated by academics. Finally,  we give a brief introduction of our research on the vehicle routing problems with drones.

Learning Times Series Data under Uncertainty

Prof. Engelbert Mephu Nguifo (LIMOS)

Abstract. The field of time series analysis has been very active during the last decade. This task that aims at analyzing chronological data has been used in a diverse range of application domains including meteorology, medicine, physics and also computing wrt the Internet of Things. Recently, lots of accurate methods have been developed to perform time series analysis. However, applications in which the time series data have uncertainty are still challenging and under-explored. All measurements performed by a mechanical system contain uncertainty, and ignoring the uncertainty of the data during their analysis can lead to inaccurate conclusions, hence the need to implement uncertain data analysis techniques is of great importance. This talk will first introduce basic concepts around time series data, then review state-of-the art on uncertainty in data mining tasks, and finally present our contributions in clustering as well as classification of time series data under uncertainty.

Operations Research in Fintech: some applications 

Dr. Bui Quoc Trung (SoICT - HUST)

Abstract. In the literature, there are many research works that apply operations research techniques to efficiently tackle fintech problems. In this presentation, we would like to discuss two ongoing research problems. Firstly, we will discuss the optimal binning approach for credit scoring models, which involves discretizing a variable or feature into bins that satisfy a set of predefined constraints and maximize the information value of the feature after binning. Secondly, we will address a variant of subset selection problem, in which a subset of a given set of passbooks must be selected to achieve a desired amount of money while minimizing the penalty for early termination before the due date 

Closing of Day 1

Machine Learning, Networks and Graphs talks

Computer Vision for Healthcare and Remote Sensing

Dr. Dinh Viet Sang (SoICT - HUST)

Abstract. Computer vision has numerous applications in healthcare and remote sensing. It enables doctors and researchers to gather and analyze vast amounts of visual data, improving diagnoses and environmental monitoring. This talk will present our recent research on semantic segmentation and some applications in healthcare and remote sensing.

Study and improvement of LoRa throughput

Prof. Alexandre Guitton (LIMOS)

Abstract. In this talk, I focus on the improvement of the throughput of the LoRa (Long Range) wireless communication technology. I present three types of monitoring applications where LoRa is suitable: wide areas (e.g., forests or volcanoes), smart cities, and isolated areas (e.g., deserts or oceans). For each application, I describe the research questions I focus on, as well as some solutions.

Quality assessment of multimedia data over 5G network

Dr. Trinh Van Chien (SoICT - HUST)

Abstract. The subjective acceptability of a network service's perceived multimedia data quality is known as quality of experience (QoE). This talk discusses the effects of the fifth generation (5G) problems on QoE and broadens the view to new QoE acceptability for future 5G networks. Based on the neural network method, an effective QoE calculation method specifically designed for 5G systems is also put forth. We visualize that neural networks are suited to achieve QoE self-optimization for 5G due to their capacity to thoroughly learn the causal relationship between network parameters of quality of services (QoS) and the resulting QoE.  Neural network designs for QoE in 5G networks open new research issues and challenges.

Machine Learning for designing Next-Generation Networks

Prof. Oussama Habachi (LIMOS)

Abstract. Nowadays, we should tackle several interesting, and sometimes contradictory, challenges when dealing with wireless communication networks, such as quality of service (QoS)-provisioning in terms of data rate, latency and quality of experience (QoE), reliability and massive access. In this presentation, designing Next-Generation Networks (NGNs) in order to handle these challenges will be investigated using Artificial Intelligence (AI), edge intelligence, network softwarization, and data-plane programmability. Given the expected rise in system complexity and the exponential increase in the amount of data exchanged through networks, smart and autonomous services and applications should be implemented at the users’ side through AI and Machine Learning (ML). For example, the integration of Federated Learning (FL) and Deep Reinforcement Learning (DRL) can support scalable, secure and diversified services and applications. Indeed, network intelligence could be implemented directly on programmable devices enabling a faster reaction to network events without depending on a time-consuming exchange. Moreover, NGNs are expected to be self-learning, self-reconfigurable, self-optimized, self-healing, self-organized, self-aggregated, and self-protected. In this vein, in order to be more flexible and robust, human control is no more possible for the management of NGNs. Sophisticated ML techniques seem to be very helpful to support network autonomy as well as to capture insights and comprehension of the surrounding environment in which they operate.

Robust Federated Learning for non-IID data 

Dr. Nguyen Phi Le  (SoICT-HUST)

Abstract. Federated learning enables edge devices to traina global model collaboratively without exposing their data. Despite achieving outstanding advantages in computing efficiencyand privacy protection, federated learning faces a significantchallenge when dealing with non IID data, i.e., data generatedby clients that are typically not independent and identicallydistributed. In this paper, we tackle a new type of Non-IID data,called cluster skewed non-IID, discovered in actual data sets.The cluster-skewed non-IID is a phenomenon in which clientscan be grouped into clusters with similar data distributions. Byperforming an in-depth analysis of the behavior of a classificationmodel’s penultimate layer, we introduce a metric that quantifiesthe similarity between two clients’ data distributions withoutviolating their privacy. We then propose an aggregation schemethat guarantees equality between clusters. In addition, we offera novel local training regularization based on the knowledgedistillationtechnique that reduces the overfitting problem atclients and dramatically boosts the training scheme’s performance.We theoretically prove the superiority of the proposedaggregation over the benchmark FedAvg. Extensive experimentalresults on both standard public datasets and our in-house realworlddataset demonstrate that the proposed approach improvesaccuracy by up to 16% compared to the FedAvg algorithm.

Learning, inferring and predicting on manifolds

Assoc. Prof.  Chafik Samir (LIMOS)

Abstract. In this talk, we will try to summarize some of our recent works for generalizing classification and regression models on manifolds. We will focus on novel constructions of Gaussian processes indexed by nonlinear data: Probability density functions, cumulative functions, shape of curves and surfaces, etc. Following different formulations, our main goal is to define an appropriate model with respect to the underlying geometric structure. Hence, we first give an overview on how to build such models then we show their applications on real data.

Social media network analysis: a review on two applications

Dr. Le Chi Ngoc (SAMI - HUST)

Abstract. In this talk, we will present some applications of graph and network theories in social media networks. The two selected problems are fake acounts detection and spam caller identifications. Some machine learning methods are also considered and employed in the two problems.

Graphs, algorithms, approximation, complexity

Prof. Christian Laforest (LIMOS)

Abstract. I will give a general presentation/illustration of some of my active recent research works. They concern complexity results on classical graph problems (like dominating sets, vertex cover…) with additive new constraints. In their classical versions, these optimization (mainly minimization) problems, vertices (or edges) of the input graph are allowed to be included one by one in the future solution. In our variant, an instance is composed of a graph and by a partition of its elements (edges or vertices). Each part of this partition can model a team or a set of devices that must be used collectively (if an element is used then all the other ones must also be used). We obtained a lot of hardness results and a few algorithms.

We also investigated another variant where the constraints are the opposite, namely conflicts, i.e. pairs of elements that cannot be simultaneously in a solution.

Community detection in Directed Graphs using Stationary Distribution and Hitting Times Method 

Dr. Do Duy Hieu (Hanoi Institute of Mathematics - Vietnam Academy of Science)

Abstract. Community detection has been extensively developed using various algorithms. One of the most powerful algorithms for undirected graphs is Walktrap, which determines the distance between vertices by employing random walks and evaluates clusters using modularity based on vertex degrees. Although several directions have been explored to extend this method to directed graphs, they have yet to be effective. This paper investigates the Walktrap algorithm and extends it to directed graphs. We propose a novel approach in which the distance between vertices is defined using hitting time. This definition is highly effective, as algorithms for hitting time have been developed, allowing for good computational complexity. Our proposed method is particularly useful for directed graphs, with the well-known results for undirected graphs being special cases. Additionally, we have implemented our algorithms to demonstrate their plausibility and effectivene.