The rise of manipulative reviews on the web threats consumer confidence and platform honesty. This paper introduces an improved fake review detection model that combines deep contextual text embeddings from transformer models with behavioral features. Two publicly available datasets, the Fake Reviews Dataset of Joni Salminen and the Yelp Reviews Dataset, were used to test the performance of the proposed method. Textual features were mined with BERT and some of its variations, such as RoBERTa, ALBERT, DistilBERT, DeBERTa, and ELECTRA, whereas behavior features were obtained from reviewer metadata, such as rating deviation, review frequency, and content readability. A neural network classifier was trained on the combined feature vectors, yielding superior performance over traditional machine learning classifiers. Moreover, explainability using SHAP was added to explain the impact of behavioral attributes on the prediction made by the proposed model. Results from experiments show that fusing rich text context and user behavior signals greatly improves detection accuracy and domain generalizability.