Mining social networks for personalized email prioritization

摘要: Email is one of the most prevalent communication tools today, and solving email overload problem pressingly urgent. A good way to alleviate automatically prioritize received messages according priorities each user. However, research on statistical learning methods for fully personalized prioritization (PEP) has been sparse due privacy issues, since people are reluctant share personal importance judgments with community. It therefore important develop evaluate PEP under assumption that only limited training examples can be available, system have data user during testing model This paper presents first study (to best our knowledge) such an assumption. Specifically, we focus analysis social networks capture groups obtain rich features represent roles from viewpoint a particular We also developed novel semi-supervised (transductive) algorithm propagates labels test through message nodes in network. These together enable us enriched vector representation new message, which consists both standard (such as words title or body, sender receiver IDs, etc.) induced receivers message. Using input SVM classifiers predict level obtained significant performance improvement over baseline (without features) experiments multi-user collection. collection: relative error reduction MAE was 31% micro-averaging, 14% macro-averaging.