Dual Cold-Start News Recommendation via Neighborhood-Based Transfer Learning

News recommendation has become an important service on mobile devices, and this article focuses on recommending the latest news articles to brand‑new users, defining the Dual Cold‑Start Recommendation (DCSR) problem.

Existing news recommendation methods that rely on users’ historical reading behavior and article content are unsuitable for DCSR because such information is unavailable for new users and new items.

The problem is addressed from a transfer‑learning perspective by leveraging knowledge from the application‑installation domain, assuming that users with similar app‑installation patterns exhibit similar news preferences.

Two domains are defined: the source app domain and the target news domain. In the app domain a triple (u,g,G ug ) indicates that user u has installed a mobile app of type g, forming a user‑type matrix G.

In the news domain a user‑item matrix R records whether a user has read an item, and each item i is associated with a primary category c 1 (i) and a secondary category c 2 (i), yielding a user‑category matrix C after preprocessing.

The goal is to produce a ranked list of new news articles for users who have never read any item, using only category information.

To solve the two challenges of new‑user and new‑item cold‑start, a Neighborhood‑based Transfer Learning (NTL) method is proposed. NTL transfers neighborhood knowledge from the app domain to the news domain, and introduces a category‑level preference to replace the unavailable item‑level preference.

The preference prediction follows the standard neighborhood formula r̂ u,i = (1/|N_u|) Σ_{u'∈N_u} r̂_{u',i} , where N_u is the set of nearest neighbors of user u computed from app‑domain vectors.

For new items, a category‑level preference is estimated as p̂_{c1(i)} = (1/|N_{u',c1(i)}|) Σ_{v∈N_{u',c1(i)}} r̂_{v,i} (and similarly for the secondary category), allowing the final prediction to be expressed as a combination of these terms.

These equations enable the NTL method to address both the new‑user and new‑item cold‑start challenges in news recommendation.