DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
Towards an effective recommendation algorithm for e-commerce
No full textE-commerce has revolutionized the way people shop, offering convenience and a vast array of products. However, with the sheer volume of items available, users often face difficulty in finding products that match their preferences.
Recommendation algorithms play a crucial role in enhancing user experience by suggesting products that align with individual tastes and needs. Over the years, various recommendation systems have been developed, ranging from collaborative filtering to content-based and hybrid approaches.
This project focuses on evaluating and comparing the effectiveness of four state-of-the-art recommendation algorithms: Matrix Factorization (MF), Bayesian Personalized Ranking (BPR), Neural Collaborative Filtering (NCF), and Neural Graph Collaborative Filtering (NGCF). Using the widely recognized MovieLens dataset, these algorithms are assessed through a comprehensive set of evaluation metrics, including Mean Average Precision (MAP), Normalized Discounted Cumulative Gain (NDCG), Precision, and Recall, to determine their performance in an e-commerce context.
The project first compares the four algorithms to identify their relative strengths and weaknesses. Following this, a series of parameter tuning experiments will be conducted to examine the impact of various hyperparameters on performance. By systematically adjusting key parameters, the study aims to optimize each algorithm and assess how these modifications influence the evaluation metrics. Based on the results, the project recommends optimal parameter configurations to achieve the best performance, providing actionable insights for the implementation of effective recommendation systems in e-commerce platforms.Bachelor's degre
Transfer vision-language-action models for general mobile manipulation
No full textMobile manipulation is a critical challenge in robotics, aiming to assist humans in performing diverse tasks across varying environments. Traditional approaches often struggle with generalization due to the limited scope of task-specific training. While recent advancements in vision-language-action (VLA) models exhibit remarkable generalization, they primarily focus on fixed-base manipulation — mainly because fixed-base robots provide a stable reference frame, simplifying kinematic modeling and control. Additionally, most benchmark datasets and training scenarios are designed for fixed-base systems, making their development and optimization more straightforward. To bridge this gap, we introduce a novel framework for adapting pre-trained VLA models to mobile manipulation, enabling robust performance across tasks and environments. Our approach utilizes pre-trained VLA models to predict generalized end-effector waypoints and incorporates motion planning objectives for both the robot's base and arm to ensure physically feasible trajectories. Additionally, we propose a bi-level optimization framework, where the high-level optimization generates base movement waypoints to expand the arm’s operational range, and the low-level optimization refines end-effector trajectories to complete tasks effectively. Offline experimental results demonstrate that our method outperforms existing mobile manipulation techniques and showcases the efficiency of leveraging pre-trained VLA models for policy adaptation.Master's degre
Reading the ridges: Holocene sea level change in Bintan, Indonesia
No full textHolocene sea-level history in Southeast Asia remains poorly constrained, and beach ridges are generally unexplored as paleo sea-level proxies in the region. This study evaluates the applicability of beach ridge stratigraphy for reconstructing sea-level change using Ground Penetrating Radar (GPR) surveys at two sites in northern Bintan, Indonesia. Internal ridge structures were imaged to identify the downlap point separating foreshore and shoreface deposits, interpreted as a marker of paleo low-tide levels. These stratigraphic indicators were corrected for tidal range, and associated sediment samples were dated using optically stimulated luminescence (OSL) and radiocarbon techniques. This study reveals a distinct sea-level fall of up to 0.38 ± 0.09 mm/yr between 5.3 and 3.64 ka BP, followed by a more gradual decline of 0.05 ± 0.04 mm/yr between 1.8 and 1.6 ka BP. Local relative sea-level trends display a stepwise pattern characterized by phases of rapid fall and extended stability, consistent with the steep–stable–steep pattern reported elsewhere on the central Sunda Shelf. Comparison with existing records from Singapore and Peninsular Malaysia shows close agreement with data from northwestern Bintan, whereas the northeastern site exhibits significantly higher relative sea levels. This spatial divergence may reflect localized tectonic influences and warrants further investigation. Overall, the study demonstrates that beach ridge GPR stratigraphy is an effective method for reconstructing Holocene sea-level history in Southeast Asia.Bachelor's degre
Hardware/software test and validation system
No full textThe MDP is an essential module for students in the College of Computing and Data Science
emphasizing collaboration between teams managing various components to build a
functional robot car. The utilization of diverse sensors and hardware components particularly
the STM32, plays a critical role in the project. However, challenges often arise in testing and
validating the functionality of these components, hindering students' progress and resulting
in inefficiencies during the development of the robot car.
Currently, no standardized system exists to address the unique challenges of the MDP. This
forces students to write separate code to test individual hardware components of the robot,
leading to wasted time and effort.
To address this, the objective of this project is to develop a comprehensive
Hardware/Software Test and Validation System tailored for MDP students. A desktop app
named MDPHelper was developed to assist students in testing and validating their
embedded systems. The app provides tools for hardware testing, communication with
STM32, flashing custom firmware to the STM32, and visualizing real-time data from
connected hardware components.
The system was developed using .NET MAUI for the frontend, with a .NET Core and
MongoDB backend to manage firmware files and app updates. This solution simplifies the
testing and validation process, enabling students to focus on the design and integration
aspects of their projects while improving overall efficiency and collaboration.Bachelor's degre
Deep learning based prediction and planning for autonomous driving
No full textWe propose TrajDiffusion, a diffusion-based decoder model designed for trajectory prediction tasks involving single or multiple agents. Our decoder module has several distinctive features. First, as the input to the diffusion process relies on 20 anchor points sampled from real-world data, our model effectively captures diverse multimodal predictions. Second, by incorporating the DPM-Solver method into the diffusion model, we achieve high-quality predictions within just four inference steps. This significantly alleviates real-time constraints commonly faced when applying diffusion models to trajectory prediction tasks.
Third, our approach is flexible, being applicable to both single-agent and multi-agent prediction scenarios. Additionally, we introduce a compact trajectory representation via Whitened Principal Component Analysis (PCA). Unlike standard PCA, which only removes correlations among components, Whitened PCA also normalizes each component to unit variance. This adjustment makes the trajectory representation better aligned with the isotropic Gaussian noise used in diffusion models, reducing the mismatch between the noise distribution and the data distribution. Consequently, it enhances the effectiveness of the denoising process and improves overall prediction performance.
Finally, we present an optional constrained sampling function, enabling controlled trajectory generation. By computing differentiable cost functions based on generated trajectories and incorporating rules or physical priors, we guide trajectories toward more physically plausible outcomes. Our proposed TrajDiffusion framework can be readily integrated with existing encoder architectures, resulting in diverse outputs, fast inference speed, comprehensive utilization of encoder features, and excellent performance on the Argoverse 2 dataset.Master's degre
BLE-based indoor asset-tracking system
No full textIndoor asset tracking plays a vital role in industries such as healthcare and logistics, where real-time location monitoring enhances operational efficiency and asset management. This project focuses on developing a Bluetooth Low Energy (BLE)-based asset tracking system, integrating BLE-enabled asset tags and anchors with a web application for asset visualization and device management.
The system is designed for indoor localization using the fingerprinting technique on Received Signal Strength Indicator (RSSI) data, coupled with a web-based interface which allows users to monitor and configure devices easily. Due to the absence of a trained Convolutional Neural Network (CNN) model, localization was simulated within a grid-based environment to assess feasibility.
Results show that the system successfully utilizes BLE communication for real-time tracking, with a localization precision of up to 1m-by-1m grid cells.
Nevertheless, further enhancements to scalability, security, and tracking scope can be made to improve the system’s overall performance and applicability.Bachelor's degre
Reinforcement learning with possibility theory
No full textDeep Reinforcement Learning (RL) agents often struggle to balance exploration and exploitation due to their value estimates not accounting for epistemic uncertainty. Possibility theory, with its maxitive calculus and less restrictive normalization constraints, offers a method for tracking and using epistemic uncertainty. This thesis investigates whether possibilistic modeling of uncertainty can drive principled optimistic exploration. Three algorithmic variants are proposed: (i) Possibilistic DQN, which models Q-values via a Gaussian and acts via a parameter-free closed-form maximum expected value; (ii) Possibilistic Q-Ensembles, which maintain possibility weighting over multiple Q-networks and update them using a possibilistic Bayes rule; and (iii) Possibilistic Model-Based Learning, a pair of zero- and one-step planning algorithms that propagate optimistic targets from possibilistic models. The models are tested on benchmark environments along with sparse and stochastic variants. The study confirms the efficacy of maximum-expected-value optimism in the first method, particularly in deterministic and sparse environments. The ensemble possibilistic approach outperforms standard non-weighted ensembles and DQN. The third model underperformed relative to baselines; further work is needed to tune hyperparameters and improve model accuracy. Overall, we demonstrate the utility of possibility theory to capture uncertainty and make exploration more efficient.Bachelor's degre
Memory-augmented personalized agents for long-term dialogue
No full textIn the field of open-domain dialogue systems, with the advent of large language models (LLMs), significant progress has been witnessed. However, most existing systems are designed for short, single-session interactions, falling short of meeting real-world needs for long-term companionship and personalized exchanges with chatbots. Addressing these demands requires robust event summarization and persona management, which enable reasoning for contextually appropriate responses in extended dialogues.
Recent advancements in the cognitive and reasoning capabilities of LLMs highlight their potential to transform automated perception, decision-making, and problem-solving. This dissertation implements a memory-augmented personalized agent framework which integrates three independently modules focused on event memory, persona extraction, and response generation.
The event memory module employs both long-term and short-term memory banks to handle historical and ongoing session data, respectively. A topic-based retrieval mechanism further improves the accuracy of memory retrieval. Meanwhile, the persona module dynamically models the personas of both users and agents, enabling nuanced interactions. Retrieved memories and extracted personas are then combined within the response generator to produce relevant and coherent outputs.
Empirical evaluations demonstrate the effectiveness of this memory-augmented personalized agent framework, establishing it as a versatile solution for long-term dialogue scenarios.Master's degre
Stable Zn-metal anode enabled by solvation structure modulation and in-situ SEI layer construction
No full textAqueous zinc-ion batteries encounter impediments on their trajectory towards commercialization, primarily due to challenges such as dendritic growth, hydrogen evolution reaction. Throughout recent decades of investigation, electrolyte modulation by using function additives is widely considered as a facile and efficient way to prolong the Zn anode lifespan. Herein, N-(2-hydroxypropyl)ethylenediamine is employed as an additive to attach onto the Zn surface with a substantial adsorption energy with (002) facet. The as-formed in-situ solid-electrolyte interphase layer effectively mitigates hydrogen evolution reaction by constructing a lean-water internal Helmholtz layer. Additionally, N-(2-hydroxypropyl)ethylenediamine establishes a coordination complex with Zn2+, thereby modulating the solvation structure and enhancing the mobility of Zn2+. As expected, the Zn-symmetrical cell with N-(2-hydroxypropyl)ethylenediamine additive demonstrated successful cycling exceeding 1500 h under 1 mA cm−2 for 0.5 mAh cm−2. Furthermore, the Zn//δ-MnO2 battery maintains a capacity of approximately 130 mAh g−1 after 800 cycles at 1 A g−1, with a Coulombic efficiency surpassing 98%. This work presents a streamlined approach for realizing aqueous zinc-ion batteries with extended service life.Published versionThis work is supported by the National Natural Science Foundation of China (52272258 and 52411530056), the Beijing Nova Program (20220484214), Key R&D and Transformation Projects in Qinghai Province (2023-HZ-801). J-LY acknowledges the financial support from the China Scholarship Council (No. 202006210070)
Exploiting diffusion prior for image super-resolution
No full textImage super-resolution refers to the fundamental task in computer vision that produces high-resolution images from low-resolution inputs. This process has diverse applications, including medical imaging, surveillance and security, satellite imagery and scientific research. Traditional approaches for super-resolution are often constrained by predefined assumptions about the image, whereas recently developed deep generative models have overcome these limitations, demonstrating significant improvements in realism and versatility in handling different scenarios.
In recent years, diffusion models have shown a remarkable ability to generate highly detailed and realistic images, often outperforming the GANs-based generative models. This motivates the incorporation of diffusion models into the deep image priors, aiming to further improve the accuracy of the results.
This paper presents a new approach based on Deep Generative Prior (DGP) for super-resolution tasks by incorporating DGP with diffusion models. More precisely, we integrate the open-source text-to-image generation model Stable Diffusion 3 into DGP models to capture richer image statistics and improve super-resolution outputs. We compare the performance of the integrated solution with a GANs-based approach and demonstrate that diffusion models can effectively generate sharp high-frequency details and solve the one-to-many problem for super-resolution.Bachelor's degre