A cold-start situation exists when a recommender system doesn’t have sufficient historical information about a user or an item.
User cold-start problem
When a user has provided one or just a few more feedback, possible solutions are based on:
- cross-domain information, specially for domains strongly related like books and movies. Obviously, it requires information at least about the target users’ interests in the source domain, namely the domain used for generating recommendation in the target domain. See Cross-Domain Recommender Systems
- personality information, based on the assumption that users with similar personality traits have similar interests. Such information could be explicitly required to the users or inferred from their behaviour. See Alleviating the new user problem in collaborative filtering by exploiting personality information
- hybrid recommender systems, combining collaborative and content information, and if possible also external content information, such as metadata extracted from DBpedia. This is possible by linking the items to external resources. See Exploiting Linked Open Data in Cold-start Recommendations with Positive-only Feedback
- preference elicitation by means of active learning, namely requiring directly cold-start users to provide some information. See Active Learning in Recommender Systems and Alleviating the new user problem in collaborative filtering by exploiting personality information
- social network data, assuming that people have similar tastes to their social network friends. See the answer of Suvash Sedhain in Is there a recommender system using social media resources (e.g. Facebook likes, keywords on twitter, etc.)?
In the most extreme situation, where no feedback and additional information are available or can be elicited, you can only use popularity-based or random recommendations.
However, when you have enough feedback for a user, additional information may become not so useful or even misleading.
A rigorous cross-validation strategy for evaluation in cold-start scenario is proposed in Evaluating recommender behavior for new users.
Item cold-start problem
An item is in the cold start situation when no user has yet interacted with it, or just a few ones. Therefore, this problem particularly affects collaborative filtering systems, but not content based ones, since they do not need historical information about items. In this case I see only two solutions:
- hybrid recommender systems, since, as said before, content information may be useful to find connections among “old” and new items. See Methods and metrics for cold-start recommendations and Cold-Start Item and User Recommendation with Decoupled Completion and Transduction
- asking users for help, useful when there is no content information. it requires intelligent methods to choose the right users. See ExcUseMe: Asking Users to Help in Item Cold-Start Recommendations