Driving style identification is of vital importance for intelligent driving system design and urban traffic management. This study aims to identify and analyze driving styles using large-scale ride-hailing GPS data taking different time periods, traffic, and weather conditions into account. The large-scale GPS data are collected and preprocessed, and then, the k-means clustering is implemented to acquire driving behavior. The modified latent Dirichlet allocation topic approach is applied to extract the driving states as the latent variables behind driving behaviors and finally recognize driving styles. The results show that driving styles are composed of five driving states with different probability combinations. Different driving styles in different situations are further analyzed and compared. When considering the impact of peak periods on the driving style, it indicates that styles tend to be conservative in the morning peak, free and dispersed in the evening peak, and diverse in the off-peak hours. While comparing styles regarding the influence of workdays, drivers act more cautiously and conservatively on weekdays but freer on weekends without the pressure of peak hours. The weather factor is also explored and rainy days are verified to be the resistance of driving so that most drivers become cautious and conservative. Finally, two aberrant driving styles are discovered and countermeasures are suggested to improve traffic safety.