MMFFM: Fake-Review Detection Based on Multi-Level and Multi-Dimensional Feature Fusion

Fake-review detection aims to distinguish deceptive reviews from genuine ones by analyzing review content and related behavioral information. Existing methods have two main limitations. First, many methods represent a review at a single semantic level and therefore cannot fully model the relations among words, sentences, and the complete review. Second, textual representations are often not adequately integrated with reviewer behavior and product information. To address these limitations, we propose MMFFM, a multi-level and multi-dimensional feature fusion model for fake-review detection. First, a pretrained BERT encoder, bidirectional recurrent networks, and attention mechanisms are used to learn textual representations at the word, sentence, and review levels. Second, 40 linguistic and statistical features are extracted from review text, reviewer behavior, and product information. These features are integrated with the textual representations through hierarchical gating. Third, a multiscale convolutional module captures local patterns that may be missed by sequential encoding. Finally, the hierarchical text representation, local convolutional features, user-level and product-level features are combined for binary classification. Experiments on YelpChi, YelpNYC, and YelpZip show that MMFFM achieves F1-scores of 83.03, 89.09, and 86.84, respectively. These results exceed the strongest compared baseline by 2.56, 0.93, and 2.16 percentage points on the three datasets. MMFFM therefore achieves the best overall F1 performance among the baseline methods.