The intellectual property (IP) mapping problem is a non-deterministic polynomial-time (NP)-hard problem in network-on-chip (NoC) designs. To explore the vast solution space intelligently, this paper investigates leveraging knowledge stored in the probability model to guide the search process. This paper first introduces a new probability model based on the ordering of IPs, namely the graph and attention-based neural network (GANN). By considering the ordering of IPs rather than the ordering of NoC topology nodes, the GANN facilitates placing the IPs with high communication demand at neighboring NoC topology nodes. Simulation results demonstrate that the GANN outperforms the message passing-attention network (MAN). Furthermore, a discrete particle swarm optimization-based intelligent mapping algorithm (DIMA) is proposed. The DIMA enhances the performance of the discrete particle swarm optimization algorithm in two ways. Firstly, during the initialization phase, the initial positions of the particles are generated by the proposed GANN and the existing MAN to improve the global search capability. Secondly, during the search phase, a new position updating formula for particles is designed using the continuously updated GANN, thus enhancing the local search capability. Simulation results show that the DIMA achieves an average of 5.81% reduction in the communication cost compared to the ATSRP.