Exploring Interest Similarity Features and Their Combinations for Friendship Recommendation Without Cold Start

While people-to-people recommendation often relies on collaborative filtering, which suffers from cold-start issues, this study intends to propose methods of interest similarity that can provide recommendations with only user-provided data, thus circumventing cold-start issues. Participants in a data collection experiment provided personal data about themselves and information about the type of people they would like to become friends with based on a summary of interests and personal information. This data was used to devise several methods of comparing the users’ interests and information to determine their potential friendship success. The interests were represented as word embeddings and used in both vectorial and image forms. An original score that calculates similarity based on information about desired friends collected from the users’ data is proposed in order to capture interest similarity information based on the users’ perception. These inputs were used both separately and combined to train several neural network models, as well as ensemble models. The proposed methods show an improvement in recommendation quality when compared to the established baseline, particularly when used in ensembles.