1) Pre-training models can help solve the cold start problem in recommendation systems by providing useful features from side information like text, images, or knowledge graphs to make recommendations for new users or items. 2) There are two types of pre-trained models - feature-based models that collect features for users and items, and fine-tuning models that are initially pre-trained on user-item data and then fine-tuned for specific recommendation tasks. 3) Experiments on movie recommendation datasets showed that pre-trained models like BERT4Rec and Caser improved recommendations over non-pre-trained models, especially when external knowledge was incorporated, demonstrating that pre-training can help address cold start issues.