Aspect-Based Sentiment Analysis (ABSA) is a fundamental Natural Language Processing (NLP) task that aims to determine fine-grained sentiment polarity toward specific aspects mentioned in text. With the emergence of Large Language Models (LLMs) and transformer-based architectures, significant improvements have been achieved in contextual representation learning for sentiment analysis. However, existing LLM-inspired and transformer-based ABSA frameworks often suffer from inadequate aspect-context alignment, redundant feature integration, limited interpretability, and insufficient coordination between contextual and sequential modeling components. To address these challenges, this paper proposes two hybrid architectures, namely HABSA and AAC-HABSA, centered on a novel Adaptive Aspect Conditioning Layer (AACL) that injects aspect information prior to transformer-based contextual encoding. The proposed framework follows a structured pipeline comprising tokenization, token and positional embeddings, AACL, transformer encoding, BiLSTM refinement, aspect-guided attention, fully connected projection, and focal loss optimization. By conditioning token representations before contextual encoding, the framework enables aspect-aware contextual learning that better captures sentiment-relevant semantic dependencies. Subsequent sequential refinement and attention-based reasoning further enhance sentiment polarity alignment while improving model interpretability. To evaluate the proposed approach, a robust ABSA dataset containing approximately 10,000 recent reviews annotated across five sentiment intensity levels was developed. Extensive experiments demonstrate that HABSA and AAC-HABSA consistently outperform transformer-only and conventional hybrid baselines in terms of accuracy, macro-F1 score, robustness, and attention-based interpretability. The proposed framework provides a computationally efficient, mathematically coherent, and interpretable solution for fine-grained sentiment analysis. By strengthening aspect-aware representation learning within transformer and LLM-oriented sentiment analysis pipelines, this work contributes to the development of scalable and deployable AI-driven opinion analytics systems across real-world domains.

Intrusion detection in large-scale IoT deployments becomes particularly challenging during ongoing attack scenarios, where malicious traffic may temporarily dominate benign traffic. In such conditions, streaming network data exhibits severe class imbalance in favor of attack traffic, while device behavior remains heterogeneous, non-identically distributed (non-IID), and temporally evolving. Within federated learning environments, this imbalance can destabilize early aggregation rounds, dominant attack gradients bias the global model, distort decision boundaries, and degrade reliable discrimination of residual benign behavior. Since the server has no access to raw data, these effects can persist across communication rounds if not addressed at initialization. This article addresses the problem of distributed intrusion detection in streaming IoT networks under severe class imbalance and partial adversarial behavior. In ongoing attack scenarios, malicious traffic may dominate streaming observations, creating an inverted imbalance where benign behavior becomes underrepresented. To mitigate these issues, we propose a resilient federated ensemble framework with three key components: a similarity-guided balancing phase that selects a structurally diverse subset of majority-class samples to form a balanced initialization dataset, an incremental ensemble composed of Bernoulli Naïve Bayes, Passive–Aggressive, and SGD classifiers for pre-training over this data to produce a warm-start global model, and an accuracy-gating mechanism that accepts only performance-preserving local updates during online training. The proposed approach is evaluated on flow-level data from the CICIoT2023 benchmark that demonstrated stable client-wise convergence with high accuracy and consistent minority-class discrimination, even under skewed attack-dominant distributions. Under complete label-flipping poisoning, corrupted updates are systematically rejected, preventing global degradation. Ablation analysis confirms that removing structured initialization significantly increases early instability. The results indicate that balanced global conditioning and selective federation are critical for maintaining detection reliability in streaming IoT systems operating under sustained attack conditions.