Deep Sequence Sentiment Modeling via Graph-Theoretic Negation Boundary Resolution and Multi-Classifier Ensemble for E-Commerce Opinion Spam Detection

Traditional lexicon-based and frequency-domain sentiment analysis paradigms exhibit systemic vulnerabilities when processing syntactic negation windows, resulting in severe polarity inversion and degraded classification boundaries. While standard Bag-of-Words (BoW) frameworks dominate baseline Natural Language Processing (NLP), they inherently decouple sequence order, neutralizing contextual dependencies. This investigation introduces an Enhanced Negation Detection and Scope Resolution Framework tailored for unstructured electronic and social commerce streaming data. By formulating negation propagation as a deterministic parse-tree traversal algorithm, the system calculates shift coefficients over long-range dependencies, resolving semantic polarity inversions before feature extraction. The methodology evaluates a 62,000-record dataset spanning multi-channel mobile hardware reviews. A robust Multi-Classifier Ensemble—incorporating Multinomial Naïve Bayes, Bernoulli Naïve Bayes, and Logistic Regression with Sigmoid mapping, and Linear Support Vector Classifiers with hinge-loss optimization, and Stochastic Gradient Descent—is deployed over parameterized validation distributions. Empirical metrics demonstrate that a prioritized 80:20 training-to-testing tensor partition combined with our negation handler yields optimal convergence, with Logistic Regression achieving an absolute peak accuracy of 82.74%. These results provide an empirically verifiable blueprint for protecting e-commerce platforms from malicious opinion spamming and manipulative market dynamics.