Overview

What is cold start?

Most recommendation systems generate candidates by using collaborative filtering, matrix factorization, or deep learning algorithms. These algorithms typically rely on the user-item interaction matrix. In real-world systems, new users and items are continuously added. Because these newcomers lack sufficient historical behavioral data, the system cannot produce accurate recommendations for them or surface them to the right audience. This is known as the cold start problem. Traditional recommendation algorithms—whether used in candidate generation, coarse ranking, or fine ranking—depend heavily on behavioral data that simply does not exist for new users and items. As a result, new items receive few impressions and new user interests cannot be accurately modeled.

For many services, timely exposure of new items is critical to ecosystem health and long-term platform growth. For example, news and information services are inherently time-sensitive—if articles are not surfaced promptly, their value diminishes rapidly. On user-generated content (UGC) platforms, if a creator's new content does not reach enough users quickly, creator motivation declines, which reduces the volume of high-quality content the platform can attract over time. On dating platforms, if new users do not receive adequate attention, the platform struggles to retain newcomers and gradually loses its appeal.

In summary, the cold start problem is a significant challenge in recommendation systems. The following section describes common approaches to address it.

Solutions to the cold start problem

Algorithms and strategies for resolving the cold start problem fall into four categories: generalized, quick, transferable, and few.

Generalized: New items can be generalized by their attributes or categories. For example, a newly published item can be recommended to users who liked items in the same category. A new short video can be recommended to fans of its creator. A new article can be recommended to users interested in the same topic. These are essentially content-based recommendation methods. For better results, systems can also leverage concept-based or topic-based recommendations—matching items not just by category, but also by brand, shop, style, color, and other attributes. Some attributes are explicitly configured by merchants at launch time, while others must be inferred algorithmically, such as article topics that are not explicitly labeled.

Beyond attribute-based or topic-based generalization, another common approach computes embedding vectors for users and items and matches them by vector distance or similarity. Matrix factorization and deep neural network models can generate such embeddings, but training these models still requires behavioral data. Consequently, they cannot produce accurate embeddings for cold start users and items.

For information about the models that can generate embedding vectors for cold start users and items, see DropoutNet: Addressing Cold Start in Recommender Systems.

In essence, the generalization method uses content or attributes of new items to compensate for the lack of historical behavioral data. For example, multimodal information such as images or videos can drive recommendations. A dating platform can score the appearance of a new user and then recommend that user to others who have a preference for such appearance.

Quick: Since cold start items lack historical behavioral data, a natural approach is to rapidly collect interaction data and incorporate it into the recommendation system in near real-time. Traditional models update features on a daily basis, whereas online learning models can update in minutes or even seconds. This approach typically relies on reinforcement learning or contextual bandit algorithms.

For more information, see A Contextual-Bandit Approach to Personalized News Article Recommendation.

Transferable: Transfer learning migrates knowledge from a source domain to a target domain. For example, when a new service has limited samples, data from related services (source domain) can be used to train a model for the new service (target domain). Similarly, a cross-border e-commerce company may have a newly deployed platform in one country with sparse behavioral data. Behavioral data from mature platforms in other countries can train a base model, which is then fine-tuned on the limited local data to achieve good cold start performance. For transfer learning to work effectively, the source and target domains must share sufficient relevance—for instance, platforms in different countries that sell many of the same items.

Few: Few-shot learning trains models with a minimal amount of labeled data. A typical few-shot learning method is meta-learning.

For more information about meta learning, see Cold-start Sequential Recommendation via Meta Learner.