MaRENet achieves F1 scores of 0.9694 and 0.9883 on two large public transaction datasets, outperforming all tested baseline models. The model combines masked attention, recurrent encoding, and fixed-length history windows, and its ability to detect patterns in future transaction sequences is validated through time-ordered splits. In GPU testing, MaRENet-GRU has lower latency, higher throughput, and fewer trainable parameters than TH-LSTM.
Source:BIOENGINEER.ORG · Read Original
Leave a Reply