Recommender systems typically use
- item and user embeddings (content-based),
- user-item interactions (collaborative filtering) or
- a combination of both (hybrid)
to recommend relevant items to users. Absence of any of the above data leads to a cold start problem.
We will list down the different types of cold start problems arising from this and how it is typically addressed.
- Item/Product cold start- usually encountered in collaborative filtering method since a new item with no interaction history will fail to be recognized by the recommendation system thus making them cold start items. One intuitive and common way to solve this by using item's features & meta-data (embeddings) to recommend items - which is what content-based filtering essentially is.
- User/Visitor Cold Start- creating personalized recommendations for new visitors (or inactive users) without the knowledge of their behavior or preferences leads to user cold start problem. Few ways to deal with them -Survey method - a technique called preference elicitation can be used to generate initial embeddings through onboarding surveys. In order to keep these addition steps in an onboarding journey short and avoid potential dropouts, active learning is used whose main goal is to guide the user in the preference elicitation process in order to ask them to rate only the items that for the recommender point of view will be the most informative ones.Demographic method - we can create user embeddings by categorizing users into stereotypical descriptions such as age, gender, location etc.Popular method- system serves the most popular items to the user that have been well-received or frequently chosen by other users.
- Community / Marketplace Cold Start - This refers to the scenario where there is no information on user behavior and item features due to the startup of the system (imagine a completely new marketplace that didn't exist before). In such cases, methods like early adopter programs are applied where a specific group of users (internal or external) are exposed to the product to gain initial knowledge on user behavior.
One thing all the above three cold start problems have in common is lack of user-item interaction data and hence their mitigation strategies also have some commonalities. Different common ways to deal with them are-
- Hybrid Recommender System - using collaborative filtering for warm items and users (warm = enough interaction data) and content based filtering for cold items and users. Although you may run in to problems in the case when it is difficult to collect item features for embedding.
- Knowledge-based Systems - Some businesses may have sufficient prior knowledge about their user behavior to create rule-based system, which is typically referred to as knowledge-based system.
- Active learning - It is a technique that involves selecting a set of representative items which are most informative for users and asking them to provide feedback which is in turn used to improve the accuracy of the recommendations.
- Context-aware Systems - These systems typically consider context factors such as time, location etc. to make recommendations. You can think of Google maps using contextual information to recommend the best possible route.
- Feature Hashing Trick - hashing trick is applied to map the high-dimensional feature vectors to a lower-dimensional space with a fixed number of dimensions. It handles the cold start problem by assigning multiple users (or categories of sparse variables) to the same latent representation simultaneously helping in memory conservation.
You can read more about these here -
