簡單的名詞解釋上的區別是
discriminative model 估計的是條件概率分布(conditional distribution, posterior class probabilities)
p(class|observed x)
generative model 估計的是聯合機率分布（joint probability distribution）
p(observed x, class)，之所以稱為Generative model，就是因為根據Joint probabilities，利用各種sampling方法，就可以產生人工資料(Synthetic data)。
Christoper M. Bishop, Pattern Recognition and Machine Learning Section 1.5.4, p.43有提到這兩者的區別。在這裡作者Bishop也點出了第三種學習方法，discriminative function.
Sampling方法，可在此書的第11章找到各式方法。
在目前的研究上，
有學者研究綜合兩者的優點的Hybrid方法，
- Hybrid Generative-Discriminative Models for Speech and Speaker Recognition (2002)
- Multi-Conditional Learning: Generative/Discriminative Training for Clustering and Classfication(2006)
- Bias-Variance tradeoff in Hybrid Generative-Discriminative models(2007)
- A hybrid generative/discriminative approach to text classification with additional information(2007)
- A mixed generative-discriminative framework for pedestrian classification(2008)
- Combining Evidence from a Generative and a Discriminative Model in Phoneme Recognition(2008)
也有學者提出透過Discriminative model去學習generative model.
- Learning Generative Models via Discriminative Approaches(2007)
兩者的比較
- Generative versus discriminative methods for object recognition(2005)
