Unified Virtual Memory (UVM) relieves the developers from the onus of maintaining complex data structures and explicit data migration by enabling on-demand data movement between CPU memory and GPU memory. However, on-demand paging soon becomes a performance bottleneck of UVM due to the high latency caused by page table walks and data migration over interconnect. Prefetching is considered a promising solution to this problem given its ability to leverage the locality of program memory access patterns. However, existing locality-based prefetching schemes can not handle all the situations. %Data structures like arrays tend to be stored in contiguous blocks, and accessed repeatedly. An ideal prefetcher should not only look at narrow regions of the requested address space but also capture global context to deliver a good prediction of the memory access pattern. This paper proposes a novel approach for page prefetching for UVM through deep learning. We first show that a powerful Transformer learning model can provide high accuracy for UVM page prefetching. We then perform analysis to interpret this Transformer model and derive several insights that allow us to design a simpler model to match the unconstrained model's accuracy with orders of magnitude lower cost. We evaluate this simplified model on a set of 11 memory-intensive benchmarks from popular benchmark suites. Our solution outperforms the state-of-the-art UVM framework, improving the performance by 10.89%, improving the device memory page hit rate by 16.98% (89.02% vs. 76.10% for prior art), and reducing the CPU-GPU interconnect traffic by 11.05%. According to our proposed unified metric, which combines the accuracy, coverage, and page hit rate, our solution is approaching the ideal prefetching scheme more than the state-of-the-art design (0.90 vs. 0.85, with the perfect prefetcher of 1.0).