Muhammad Syafiq Mohd Pozi

Machine learning models deployed under distribution shift suffer performance degradation whose root cause is rarely diagnosed before adaptation is attempted. Existing frameworks either lack a closed loop from shift diagnosis to method selection, require a fully specified causal graph, or depend on multiple heterogeneous training environments—conditions rarely met in practice. We propose CausalShift, a modular, plugin-based framework for end-to-end dataset shift handling. A universal statistical core remains fully functional without any causal assumption; four optional causal plugins are activated selectively by a four-level knowledge grading system that scales from a fully specified structural causal model (Level A) down to a single observational dataset (Level D). A three-path causal attribution module estimates the relative contributions of covariate and concept shift to the performance gap, and attribution-driven routing maps the diagnosis to an adaptation strategy grounded in minimax optimality. Proof-of-concept experiments on four benchmarks—a synthetic structural causal model (SCM), Colored Modified National Institute of Standards and Technology (MNIST), Benchmarking In-the-Wild Distribution Shifts (WILDS) Camelyon17, and five American Community Survey (ACS) geographic shift tasks—demonstrate that CausalShift reduces the in-distribution to out-of-distribution (ID–OOD) accuracy gap by 19.9 percentage points on the synthetic benchmark and 72.5 percentage points on the spurious correlation benchmark, while remaining competitive on real-world image shift and achieving performance parity with ERM on mild-shift tasks. A tiered causal evaluation suite, including Λ*-robustness bounds and distribution-induced shift decomposition of error (DISDE), reveals robustness differences that scalar accuracy metrics cannot detect.

Machine learning models deployed under distribution shift suffer performance degradation whose root cause is rarely diagnosed before adaptation is attempted. Existing frameworks either lack a closed loop from shift diagnosis to method selection, require a fully specified causal graph, or depend on multiple heterogeneous training environments — conditions rarely met in practice. We propose CausalShift, a modular, plugin-based framework for end-to-end dataset shift handling. A universal statistical core remains fully functional without any causal assumption; four optional causal plugins are activated selectively by a four-level knowledge grading system that scales from a fully specified structural causal model (Level A) down to a single observational dataset (Level D). A three-path causal attribution module decomposes the performance gap into covariate and concept shift contributions, and attribution-driven routing maps the diagnosis to an adaptation strategy grounded in minimax optimality. Proof-of-concept experiments on four benchmarks — a synthetic structural causal model, Colored MNIST, WILDS Camelyon17, and five ACS geographic shift tasks — demonstrate that CausalShift reduces the ID–OOD accuracy gap by 19.9 percentage points on the synthetic benchmark and 72.5 percentage points on the spurious-correlation benchmark, while remaining competitive on real-world image shift and achieving performance parity with ERM on mild-shift tasks. A tiered causal evaluation suite, including Λ*-robustness bounds and distribution-induced shift decomposition of error (DISDE), reveals robustness differences that scalar accuracy metrics cannot detect.

Large language model (LLM)-based program generation tasks are hindered by high computational demands. These challenges, along with high deployment costs, often pose a barrier to practical applications. To address these, we propose a novel data-augmented multi-LLM model routing approach that classifies prompts based on whether they should be processed on a weak LLM engine or a strong LLM. Experimental results show up to 16 times better efficiency compared to the existing cascaded approaches, while preserving the inference accuracy. Thus, the proposed method optimally allocates prompts across multiple LLMs, reducing computational costs while maintaining inference accuracy.

Although the Delone and McLean IS Success Model (D&M) can explain the phenomenon of e-participation success (EPS), the model was initially created in an e-commerce setting and thus neglects external factors related to e-government services. To address the gap, this study revisits the D&M by extending it with four socio-cultural constructs of Trust (TR), Anonymity (AN), Nationalism (NT), and Culture (CR). Based on 428 survey data from Malaysian citizens, a hybrid methodology was employed, integrating Partial Least Squares-Structural Equation Modelling (PLS-SEM) and Multi-layer Perceptron (MLP) to capture non-linear relationships and enhance predictive accuracy. While hybrid modelling is common, past studies have often applied limited classification metrics, and the infrequent use of comprehensive metrics, such as Area Under the Receiver Operating Characteristic Curve (AUC-ROC), may potentially affect the reliability and generalizability of these models. A comparative R² analysis between the baseline and enhanced models in this study revealed significant improvements, with R² for e-participation intention (EPI) increased from 0.620 to 0.728, EPS from 0.330 to 0.345, while User Satisfaction (US) remained strong at 0.765. The analysis predicts a 94.80% success rate for e-participation, with the MLP model further demonstrating robust classification performance, achieving an accuracy of 90.1%, a precision of 0.909, a recall of 0.948, an F1 score of 0.928, and an AUC-ROC of 0.955, outperforming other benchmark classifiers. This study contributes theoretically by introducing underexplored socio-cultural variables into the D&M while methodologically extending the hybrid PLS-SEM and MLP through a robust model validation using AUC-ROC.

Small or tiny object detection has gained significant attention from researchers in recent years due to its wide range of applications across various domains. This highlights its status as a promising research area. The study employs bibliometric analysis to understand the current research status and future trends in small or tiny object detection. By analyzing documents published between 2015 and 2024 from the Scopus database, we map the quantitative structure of this research area, identify the most influential publications and key contributions, and uncover potential research directions. The contribution of this paper is to provide a comprehensive bibliometric analysis of a decade of research on small or tiny object detection, a topic that has not been previously published. A total of 1627 publications were analyzed. Our findings reveal a significant increase in the number of publications over the decade, indicating growing research interest. China leads the field, contributing the majority of publications, followed by the USA and South Korea. Foundational contributions, such as "Inside-Outside Net" and GAN-based approaches, are highlighted for their impact on detection methodologies. Key topics include deep learning, convolutional neural networks, feature extraction, and object recognition. The analysis also identifies major challenges such as data scarcity, computational efficiency, and environmental complexities like occlusion and background clutter. Emerging trends include lightweight architectures, multiscale feature fusion, and edge computing, reflecting a shift toward resource-efficient implementations for practical applications. This study provides foundational insights into the current state of the field and highlights directions for future research, aiming to guide scholars and practitioners in advancing the technology of small or tiny object detection.

Streaming data is becoming more and more common in the field of big data and incremental frameworks can address its complexity. The BDFCOM algorithm achieves good results on common form datasets by introducing the ordering mechanism of beta distribution weighting. In this paper, based on the BDFCOM algorithm, two incremental beta distribution weighted fuzzy C-ordered means clustering algorithms, SPBDFCOM and OBDFCOM, are proposed by combining the two incremental frameworks of Single-Pass and Online, respectively. In order to validate the performance of SPBDFCOM and OBDFCOM, this paper selects seven real datasets for experiments and compares their performance with six other incremental clustering algorithms using six evaluation metrics. The results show that the two proposed incremental algorithms perform significantly better compared to other algorithms.

Dataset shift will lead to uncertainty issues, and then the models will accurately decline. Using causality instead of correlation to find the invariant characteristic and solve the uncertainty issues between different dataset distributions (eg. Domain Adaptation). Summarizing datasets can be used in current domain training, building a benchmarking framework of causal learning that combines the causal inference and traditional model to detect, address, and determine the characteristic of the dataset shift.