• topic
• news
• recommendation
• graph
• content

Content based recommendation algorithms for news articles fail to capture human knowledge into account. This is both due to two facts: topic annotators are still a somewhat recent research field, and more research effort went into development of methodologies for user-based approaches. The work of this thesis is focused on topic based recommendation on news, where topics are real world things or concepts, as extracted from a public knowledge base (Wikipedia). The proposed methodology applies salient topic detection, knowledge graphs, graph embedding, and ranking methodologies to recommend items according to the human knowledge that is carried by the input news. We evaluate this approach on real world data coming from European Broadcaster, and on a well-known dataset for this task, and acknowledge that it improves results over state-of-the-art solutions by up to 10% for F1 metrics.