Frequent and high utility itemset mining is an important task in data mining, and the mined itemsets are measured by two metrics, support and utility. Among a series of methods used to solve such problems, evolutionary multi-objective methods provide a set of high-quality solutions to meet the needs of different users, as well as avoiding the problem of difficulty in determining the thresholds of support and utility in traditional algorithms. The existing multi-objective algorithms are encoded with 0-1 and the dimensionality of the decision space is proportional to items in the dataset. As a result, the curse of the dimensionality problem can occur in high-dimensional datasets. Therefore, this paper designs an itemset reduction strategy to reduce the search space by reducing the unimportant items. According to this strategy, the paper proposes a high-dimensional frequent and high utility multi-objective evolutionary algorithm for itemset mining based on itemset reduction (IR-MOEA), where a learning-based population restoration strategy is proposed to adjust the evolutionary direction for over-reduced or under-reduced individuals. In addition, an initialization strategy is proposed to generate sparse solutions that facilitate evolution. Finally, experimental results on datasets show that this algorithm outperforms the existing state-of-the-art multi-objective optimization algorithms for mining frequent and high utility itemsets, especially on high-dimensional datasets.