To reveal the dominant microscopic factors controlling the mechanical properties of compacted loess, a series of direct shear, consolidation, and scanning electron microscopy (SEM) tests were conducted under varying conditions of dry density and moisture content, and multivariate statistical analysis methods were introduced to systematically investigate the macroscopic mechanical behavior and microstructural evolution characteristics of the compacted loess. The results indicate that as dry density increases, macropores and mesopores are collapsed and transformed into small pores and micropores, causing the structure to be converted from a loose skeleton type to a dense flocculated type. Conversely, as moisture content increases, the soil skeleton is softened and the channels of macropores and mesopores are interconnected, resulting in a decrease in shear strength and an increase in compressibility. Through principal component dimensionality reduction, the area ratio of small and micropores along with the directional probability entropy, which dominate the shear strength, were quantitatively extracted to constitute the densification factor. Meanwhile, the area of meso and macropores and the average pore diameter, which control the compressibility, were extracted to form the pore scale factor. The principal component regression model established upon this basis can effectively eliminate the information redundancy of microstructural variables, and is verified to possess good generalized predictive capabilities (R2 > 0.86). Simultaneously, the quantitative hierarchical identification of three typical microstructures of compacted loess—loose, transitional, and dense types—was realized via cluster analysis. The evolution of compacted loess is revealed by this study to be fundamentally a continuous process from pore refinement to structural densification, and ultimately to strength enhancement. A novel theoretical pathway is thereby provided for the microscopic quantitative evaluation of compaction quality and complex multidimensional data analysis in loess high-fill engineering.