On behalf of the ICTAI-2025 chairs and organizers, we would like to welcome you to the 37th International IEEE Conference on Tools with Artificial Intelligence (ICTAI-2025). For 37 years ICTAI has served as the first of its kind major international forum where the creation and exchange of ideas related to artificial intelligence are fostered among academia, industry, and government agencies. The conference facilitates the cross-fertilization of these ideas and promotes their transfer into practical tools, for developing intelligent systems and pursuing artificial intelligence applications. ICTAI encompasses all technical aspects of specifying, developing and evaluating the theoretical underpinnings and applied mechanisms of the AI-based components of computer tools such as algorithms, architectures and languages.
Recent advancements in Neural Machine Translation (NMT) have significantly improved translation quality. However, the increasing size and complexity of state-of-the-art models present significant challenges for deployment on resource-limited devices. Knowledge distillation (KD) is a promising approach for compressing models, but its effectiveness diminishes when there is a large capacity gap between teacher and student models. To address this issue, we propose Evolving Knowledge Distillation (EKD), a progressive training framework in which the student model learns from a sequence of teachers with gradually increasing capacities. Experiments on IWSLT14, WMT-17, and WMT-23 benchmarks show that EKD leads to consistent improvements at each stage. On IWSLT-14, the final student achieves a BLEU score of 34.24, narrowing the gap to the strongest teacher (34.32 BLEU) to just 0.08 BLEU. Similar trends are observed on other datasets. These results demonstrate that EKD effectively bridges the capacity gap, enabling compact models to achieve performance close to that of much larger teacher models.Code and models are available at https://github.com/agi-content-generation/EKD.
In recent years, implicit neural representation of 3D scenes has evolved significantly and has been rapidly extended to multiple application scenarios. However, using this type of approach for highly dynamic urban scenes remains a challenging problem. Typical images of objects captured by cameras on board autonomous vehicles from one trajectory contains limited number of views, leading unsatisfactory reconstruction quality. To address this problem, we propose a novel hybrid network structure that decomposes the urban scene into a static background and multiple dynamic objects. The background model outputs a signed distance field with its accuracy enhanced by surface constraints. To deal with the issues of insufficient camera view we use a pre-training strategy to leverage shape and appearance prior to the object model from external datasets. We design an autoencoder architecture to encode a category of objects and employ an attention-based fusion module to better extract features from multiple object images. Furthermore, we introduce a “symmetric completion” approach to leverage the inherent symmetry property of normal cars. During experiments with data from Carla Platform, we find that our model can reconstruct scenes with high-fidelity, generate novel-view successfully and edit 3D scene freely.
This work advances existing research on converting OpenAPI descriptions into ontology-based semantic descriptions by implementing an improved ontology and translation method. This updated approach efficiently maps complex Schema objects and newly introduced asynchronous features proposed in the most recent OpenAPI version like Webhooks and Callbacks. The ontology has been evaluated through a dual assessment approach: a qualitative analysis ensuring structural and semantic integrity, and a quantitative test validating real-time efficiency in ontology mapping and query searches on a triple-store repository of approximately 10,000 OpenAPI descriptions sourced from the Swagger (API) Hub.
For over a decade, the iconic Super Mario Bros. game has been used as a benchmark for research in artificial intelligence (AI) and procedural content generation (PCG). Every PCG technique relies on some sort of level validation to ensure that the generated levels are playable. For this, an artificial agent is used to substitute for a human player. The quality of such an agent directly influences the quality of the work built upon it, as it limits the complexity of levels that can be validated, and its performance affects the size of the generative space that can be explored. In this paper, we present a new Super Mario Bros. agent, which first finds a coarse path over the grid-based abstraction of a game level, and then uses the information to guide an A* search through the simulated states of the game. The proposed agent is the first agent that is able to solve all levels with standard Super Mario Bros. features, while also being the most performant (10x fewer node evaluations needed and 8x faster on the most complex level pack), thus it constitutes a new state-of-the-art for the game and should be used as the new level validation standard. Furthermore, we empirically show that the grid-based search alone can be used for the level validation task with almost the same accuracy as the agent while being faster by two orders of magnitude.
Multiple problems exist in which a Machine Learning (ML) model would benefit the outcome, but it is difficult to obtain labeled data to train said model. Thus, few-shot learning has emerged as an idea to train ML models with only a few labeled data while still generalizing effectively. In this study, we use our previous work with the Constrained State-Preserved Extreme Learning Machine (CSPELM), an extension of the standard Extreme Learning Machine (ELM), which has shown fast training times and consistently high generalization. The CSPELM uses randomization to train the weights and hidden biases of a Single Layer Feedforward Network (SLFN), then continues with another iteration using a $\varepsilon$ constraint to further perturb the randomization and improve results. Lastly, it calculates the output weights using the Moore-Penrose (MP) generalized inverse of a matrix. We then explore few-shot learning by using 7 publicly available datasets and 8 training splits of $80 \%, 60 \%, 40 \%, 20 \%$, $10 \%, 5 \%, 2 \%$, and 1 % training set sizes to find where the CSPELM generalization degrades to an unacceptable level, if at all. Our results show that 5 of the datasets had over 80 % mean accuracy at the 1 % training set size. Furthermore, it performed well on datasets with more than 2 classes averaging over 95 % mean accuracy on 2 of the 3 of those datasets. These results show the CSPELM is a quality candidate for further study with few-shot learning and its applications.
Prompt learning has demonstrated remarkable performance in tuning Vision-Language Models (VLMs) for various downstream tasks. Recent studies have shown the effectiveness of prompt distillation in transferring distribution knowledge between VLM teachers and students. However, existing prompt knowledge distillation methods are limited in diversity, focusing solely on positive probabilities. In this paper, we propose a dual prompt distillation (DPD) method, which teaches the student from both positive and negative aspects. Specifically, during the first phase of teacher training, the positive and negative prompts are both optimized by constructing complementary probability distribution signals. In the second distillation phase, the teacher guides the student with dual prompts - positive prompts to select the correct category and negative prompts to exclude incorrect ones. Extensive experimental results across 11 datasets demonstrate that the proposed DPD method either surpasses or matches the performance of existing state-of-the-art (SOTA) methods in both few-shot learning and domain generalization tasks while maintaining competitive computational efficiency. The corresponding code is available at https://github.com/wdinancy/DPD.
Running a numerical weather model such as WRF at kilometre or sub-kilometre grid spacing over a regional domain is computationally expensive. We present physics-informed deeplearning models that ingest a single 9km WRF wind field and simultaneously predict two finer-scale wind fields at 3 km and 1 km resolution via dual decoder heads. Four representative architectures are benchmarked-Deep Residual U-Net (DeepRU), DEVINE, a bespoke 3-D Transformer, and a Fourier Neural Operator (FNO)-each trained with divergence-free, vorticity, and Navier-Stokes residual constraints plus Charbonnier and gradient perceptual losses. We train and validate our models on the city of Oslo for the year 2018. DeepRU achieves $R^{2}=0.94$ (RMSE $=0.050$) at $\mathbf{3 k m}$ and $R^{2}=0.89(\mathbf{R M S E}=0.065)$ at 1 km. DEVINE, Transformer 3-D, and FNO yield 3 km scores of $0.91-0.93$, with $\mathbf{1} \mathbf{ k m}$ scores lower by $0.02-0.08$, illustrating the increased difficulty of finer-scale reconstruction. Physicsinformed losses improve all models compared to MSE-only baselines, and the residual architecture (DeepRU) remains most effective for this dual-scale task.
Accurate classification of rare and undesirable events is crucial to maximizing the safety and efficiency of the oil and gas industry. While deep learning techniques have shown promising results in this domain, they often struggle when training data is scarce, a common issue with rare events. Domain ontologies can formally represent domain knowledge about these events. In this work, we hypothesize that it is possible to use domain knowledge represented in ontologies to improve the performance of neural networks when data is limited. Based on that, we propose a novel method by integrating domain knowledge from an ontology with deep learning through a hybrid loss function. Our hybrid loss function combines crossentropy with cosine distance between the ontological embeddings of the predicted and true classes to enforce semantic accuracy. Experiments conducted on the 3W Dataset, which includes nine rare anomalous events, demonstrate that our approach can improve classification accuracy compared to standard methods on smaller datasets.
Fake news detection presents a significant challenge in multilingual digital platforms, particularly for low-resource languages such as the Tunisian dialect. In this paper, we propose an approach that uses a creative Heterogeneous Spatio-Temporal Graph Neural Network (H-STGNN). In contrast to traditional semantic representations, it integrates static and behavioral information extracted from publications (author credibility rate, number of shares, reactions, and so on). We also introduce a new annotated dataset of Tunisian Facebook posts that are fake or real. Our results show that static and behavioral features, which are often overlooked in modern approaches, are successful in detecting fake news in this context, with a precision of 0.933 and a recall of 0.937. The use of semantic representations extracted from the pre-trained AraBERT model does not generate better outcomes, illustrating the models' limitations for local dialects. These results suggest that the behavioral and static factors in fake news detection systems need to be reevaluated.
Based on the analysis of user behavior on microvideo platforms, we find that the more times users have explicit interactions with the videos posted by vloggers, the higher the possibility of explicit interactions between users and vloggers. However, the existing recommendation models mainly rely on the direct interests between users and items to model user preferences, and fail to explore the potential interests between users and items from multiple perspectives, resulting in incomplete modeling of user preferences. In response to the above problems, this paper proposes a multi-view augmented graph neural network model for micro-video vlogger recommendation (MAGNN). Specifically, we simultaneously construct bipartite graphs of the user-vlogger interaction relationship and the uservideo interaction relationship to fully capture the direct and potential vlogger preferences of users from different views. Furthermore, we have designed a bi-directional cross-attention with cross-dot-product fusion module. It adaptively learns the correlation between the preference features of different views through the dual paths of forward propagation and backward propagation, and optimizes the attention mechanism by using the cross-dot-product mechanism to enhance the discriminative ability of the attention mechanism. We conducted a large number of experiments on two public datasets. The experimental results fully verify the effectiveness of the method we propose in the recommendation task of micro-video vloggers.
The rapid expansion of personalized healthcare has positioned wearable electrocardiogram (ECG) monitoring systems as a pivotal solution for continuous cardiac assessment. However, achieving real-time signal classification under strict power and resource constraints remains a core challenge for edge-based deployment. While convolutional neural networks (CNNs) offer excellent feature extraction capabilities, their intensive computational and memory demands impede efficient implementation on lightweight hardware platforms. To overcome these limitations, we propose a hybrid hardware accelerator integrating a binarized ReActNet backbone for low-overhead feature extraction with an XGBoost decision tree classifier that provides interpretable and hardware-efficient inference. The ReActNet module leverages binary convolution operations to reduce arithmetic complexity, while the XGBoost component replaces dense layers, enhancing model transparency and reducing logic resource usage. To further streamline deployment, we introduce a generative hardware synthesis methodology based on large language models (LLMs), enabling automatic translation of trained XGBoost models into synthesizable VHDL code from their JSON representation. This LLM-assisted workflow significantly reduces development time, eliminates manual RTL coding, and promotes modular, reusable classifier generation adaptable to evolving clinical models. The proposed system is evaluated through RTL-level simulation and synthesis targeting low-power, resource-constrained FPGA platforms, demonstrating its architectural feasibility for real-time ECG classification at the edge. These findings highlight a unified and scalable framework that fuses binary neural networks, interpretable classifiers, and generative AI techniques to accelerate hardware-software co-design and broaden the deployment of edge AI in wearable health applications.
Hierarchical task network (HTN) planning extends classical planning by incorporating a hierarchy of tasks that gives plans additional structure and speeds up planning. However, it requires that each action be part of some task in the domain model, which makes it less flexible when the task hierarchy does not capture all the possibilities to achieve every task. HTN planning with task insertion (TIHTN planning) extends HTN planning by allowing the insertion of actions outside the hierarchy, thus giving more flexibility to constructing hierarchical plans. TIHTN planning has been proposed as a theoretical concept to show some decidability and complexity results. This paper describes an implemented TIHTN planner for totally ordered domains utilizing top-down grammar parsing.
Group discussions are common in both private and professional meetings. In such discussions, participants often aim to prioritize options, such as activities or features, based on the group's collective preferences. However, aligning individual preferences within a group can be challenging and may result in dissatisfaction. Group recommender systems address this by aggregating diverse and potentially conflicting preferences. This paper investigates the use of large language models (LLMs) for prioritization in meeting-based group recommendations. We show that LLMs can extract individual preferences from meeting transcripts and generate group-level recommendations without requiring additional manual input. Our user study confirms the feasibility and effectiveness of this approach, highlighting the potential of LLMs to enhance group recommendation workflows within AI-driven decision support systems.
Cloud computing enables efficient digital transformations for organizations but also raises significant challenges for cost management due to its variability and complexity. Rapid advancements in Artificial Intelligence (AI) bring promising opportunities to address these challenges, particularly in cloud cost forecasting. However, implementing AI-based models for cloud cost forecasting remains novel and challenging, as the financial domain requires high trustworthiness in AI solutions. Explainable AI (XAI) addresses this issue by developing techniques that clarify AI decisions, making models more transparent and reliable. Moreover, XAI explanations can help identify redundant features, leading to improved model performance. This paper introduces a cloud cost forecasting approach using forecasting models for time series data. The predictions are explained using the Kernel SHAP method, which highlights the impact of different features on the model's output. The forecasting model is then refined by removing low-impact features. The results demonstrate that the refined models enhanced by XAI outperform the original models due to an efficient feature selection process. Our study highlights the capability of AI and XAI to address cloud cost forecasting challenges by providing accurate predictions and clear explanations.
In most time-dependent real world phenomena, the distribution of data changes through time. This effect poses challenges to the training of time series Machine Learning models and their prediction results. Algorithms that address this problem, train the machine learning model on a subsequence of the input time series containing recent samples of the distribution, in order to capture the latest state of the distribution. However, estimating the optimal size for the training subsequence is not straightforward. State-of-the-art methods use change point detection algorithms to detect the change points of the distribution and set the most recent change point as the start of the training subsequence. In this paper, we show that this approach is not optimal. Instead, we present an algorithm that retrieves the training subsequence which minimizes the prediction error of the machine learning model. We evaluate our algorithm on two synthetic and four real-world datasets and show that it significantly increases the accuracy of the models.
This study addresses the Electric Vehicle Charging Scheduling Problem (EVCSP) with the objective of maximizing the number of scheduled charging requests while satisfying grid and charger constraints. To solve this problem, a hybrid bi-level optimization framework is developed, where a Graph Convolutional Network (GCN) guides the upper-level assignment of EVs to chargers, and exact methods ensure feasible energy allocation at the lower level. The proposed framework is evaluated on diverse and large-scale instances, with comparisons to a simulated annealing algorithm introduced in the literature as an effective method for solving the studied problem. Results demonstrate that the hybrid GCN-based approach achieves high-quality solutions with improved computational efficiency. These findings underscore the effectiveness of integrating learned graph representations with mathematical programming to solve complex and large-scale EV charging scheduling problems.
Change detection in videos is a crucial task for identifying alterations in an environment, contingent on knowledge of the base environment. In this work, we assume goal-oriented problems and want to develop an autonomous agent that uses video data from mobile cameras to detect environmental changes. The agent then autonomously adjusts its path in response to these detected changes to reach the target. We introduce an attention-based model by combining UNet and Convolution Blocks Attention Module (CBAM) for change detection, named CUCD, and use a Deep Q-Network (DQN) reinforcement learning model to build a system for autonomous path planning. To examine the effectiveness of the system, we design a pipeline connecting a drone, our change detection model, and a path planning model. The experiments conducted in both simulation and the real world show that changes in the environment can be effectively detected and the navigation path appropriately replanned using our method.
With the rapid growth of the internet and digital economy, fraudulent activities have grown more sophisticated, leading to substantial social and financial costs. Graph-based fraud detection provides an effective framework by modeling entities as nodes and interactions as edges. However, the prevalent heterophily (target nodes exhibit distinct labels from neighbors) and heterogeneity (multi-typed nodes/edges) in fraud graphs cause traditional graph neural networks to lose critical anomaly signals during message passing. Existing methods handle heterophily and heterogeneity in isolation: heterophily mitigation relies on static grouping strategies, while heterogeneity handling predominantly adopts adaptive weighting mechanisms. Yet neither systematically models their synergistic effects, limiting hierarchical pattern capturing capabilities. This paper proposes HHGFD, a synergistic framework integrating reinforcement learning-driven dynamic grouping with hierarchical attention aggregation. We formulate grouping decisions as a Markov process through reinforcement learning, optimizing grouping boundaries via real-time feedback. Additionally, we introduce a hierarchical attention aggregation framework: node-level self-attention learns cross-type feature importance weights to emphasize fraud-relevant attributes, while relation-level attention dynamically allocates semantic weights for heterogeneous edges. Extensive experiments on four datasets demonstrate the effectiveness of our proposed method.
In sports research, highly optimised machine learning (ML) models often fail to be applied “in the field“. Although useful and with strong performances, they offer limited flexibility to experts to “interact” with the model. This study proposes TREEFARMS+, a Rashomon set-based framework (a collection of prediction models that exhibit near-optimal accuracy) that addresses the aforementioned limitation of traditional ML approaches, using reasonable resources. TREEFARMS+ is jointly presented with an interactive tool which facilitates the user's trade-offs between performance, complexity, interpretation and expertise-based knowledge. Although presented here in a sports science context, showing beneficial effects to the prediction of future injuries in mid/long-range runners, this method can be easily integrated into different fields.
