This paper proposes a meta-structure-based heterogeneous personalized space-embedded recommendation system. The system algorithm uses a meta-structure-based random wandering strategy for heterogeneous personalized space for node sequence generation, which selects the next node type and chooses the next node through personalized probability. Then, node embedding learning is carried out through the heterogeneous Skip-Gram algorithm after obtaining node sequences. A nonlinear fusion function transforms the learned embedding vectors with different meta-structures and then integrates them into a matrix decomposition model for rating prediction. Pre-processing the user check-in datasets LastFM and Urban for a platform, the number of check-ins varies significantly, with some records exceeding 1,000 while others are only in the single digits. A comparison of recommender system performance shows that MPHSRec outperforms the comparison method in all metrics, with a recall of 0.2415 on the Top 20. This model analyzes the impact on cold-start users with a number of data and item interactions of less than 10 and verifies the validity as well as the feasibility of the methodology proposed in this paper.