In recent years, diversity in recommender systems have become increasingly an essential dimension for evaluating the effectiveness of recommendations. However, many existing recommendation techniques are challenged by information overload with the widespread use of recommender systems in many real-world applications. In this paper, we propose a new diversified recommendation approach, namely DRN2V, based on rich constructed graphs and Network Embedding technology. Specifically, we construct a knowledge graph of two sub-graphs, the User-Item subgraph that represents the interactions between users and items and the Item-Category subgraph which uses the item categorization to enrich the network structure. Afterwards, we use Node2vec algorithm to capture the complex latent relationships between users and items from the constructed knowledge graph. Moreover, to propose personalized and relevant predictions for each user, a new formula was proposed based on category coverage and users' preferences for categories. The experimental results demonstrate the significant outperforms of our approach over several embedding-based methods and recommendation algorithms including both traditional and diversity-oriented algorithms) regarding accuracy and diversity.
展开▼
