Differences in cognitive levels among construction workers create a strong need for targeted personalized safety training. Personalized safety training plays an important role in influencing visual search behavior during hazard recognition. To explore the underlying mechanism, an eye-tracking experiment is conducted with 40 civil engineering students.Mean-shift clustering is used to dynamically define areas of interest (AOIs). MultiMatch is applied to analyze scanpath similarity across five dimensions. The Maximal Information Coefficient (MIC) is used to quantify the association between visual features and recognition performance. A Markov chain model is constructed to calculate gaze transition probabilities and steady-state distributions. In addition, local sequence alignment is employed to identify typical scanpath patterns.The results show that higher recognition accuracy is achieved in the personalized training group. A higher proportion of hazard-oriented gaze and longer average fixation duration are observed. Fixation entropy and average fixation duration exhibit the strongest associations with recognition performance. In the correct group, scanpaths show greater consistency in fixation duration and path length, and gaze transitions are concentrated on key hazard regions, forming a stable structure centered on hazard sources. In contrast, scanpaths in the incorrect group are more dispersed, with a higher steady-state probability assigned to environmental factors. Local sequence alignment reveals that a common scanpath pattern of “worker → hazard source → related elements” is present across different scenarios in the correct group, with a participant coverage rate exceeding 80%.These findings provide a theoretical basis for optimizing construction safety training methods and improving hazard recognition performance.