The increasing number of fraudulent reviews erodes consumer confidence and skews product assessments on e-commerce sites. Traditional classifiers like logistic regression and Support Vector Machine are less effective against dynamic spamming strategies because current detection techniques frequently rely on single features or require large, rapidly outdated manually labelled datasets. To address this, We suggest lightweight framework based on Long Short-Term Memory that models sequential word patterns and incorporates contextual cues by merging review text, star ratings and category labels into single cleaned sequence. This integrated sequence is processed by compact neural network comprising an embedding layer, a Long Short Term Memory encoder to learn intricate dependencies and simple classification layer. This proposed study achieved a peak accuracy of 95.54%, demonstrating the efficacy of sequence sensitive modelling for the detection of fraudulent reviews and outperforming conventional techniques.