The　pre-trained　models　such　as　BERT　for　fake　review　detection　have　received　more　attention.　Most　of　research　has　overlooked　the　role　of　behavioral　features.　Additionally,　the　improvements　of　the　pre-trained　models　have　increased　the　computational　overhead.　To　enhance　BERT’s　ability　to　extract　contextual　features　while　minimizing　computational　overhead,　we　propose　an　end-to-end　model　for　fake　review　detection　that　combines　textual　and　behavioral　feature　information.　Firstly,　the　review　average　cosine　similarity　and　the　review　corpus　are　jointly　fed　into　BERT　to　obtain　textual　feature　vectors.　Then,　the　underlying　patterns　in　the　behavioral　information　are　extracted　by　CNN　to　construct　behavioral　feature　vectors.　Finally,　the　two　feature　vectors　are　concatenated　for　fake　review　detection.　The　entire　model　and　each　component　were　evaluated　on　YELP　dataset.　Compared　with　the　original　BERT　model,　the　F1　score　and　AUC　score　of　the　BERT　model　fused　cosine　similarity　are　improved　by　6.30%　and　6.37%,　respectively.　Our　model　achieves　a　further　improvement　of　12.47%　and　12.48%　in　F1　score　and　AUC　score,　and　shows　good　performance　in　precision　and　recall.　The　experimental　results　show　that　the　proposed　model　improves　the　effectiveness　of　BERT　for　detecting　fake　reviews,　and　is　particularly　suitable　for　scenarios　where　can　capture　behavioral　features.　©　2023,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.