Realtime Email Delivery Failure Prediction Using the One-vs-All Classifier
G. Jasmine Beulah, Abhijeet Singhai, Rashmi Ranjan Parida · 2018
Due to assorted reasons, when a sender sends an email to a receiver, the email does not get delivered. The non-delivery of an email is said to be a bounce and a bounce memo typically known as a failed Delivery Status Notification (DSN) is directed to the sender's server. Bounces can ensue because of temporary failures or permanent issues. Permanent bounces should be evaded at any cost and temporary/transient bounces should be abated. When bounces are numerous, the existing Email Service Providers (ESPs) may denounce the sender as a likely spammer. The failure messages sent by the different ESPs are unlike both in the format and in the message content, in-addition to sending ambiguous bounce messages for the same scenario. Also, when quite a few ambiguities are involved, the relay servers obscure the identification of the original reason of bounce. Prediction of bounce reason is therefore vital to proactively avoid being listed as a spammer in any public Remote Black Holes(blacklists). This paper proposes to use the One-vs-All Multiclass classifier with the best fit, the Two-Class Locally Deep Support Vector Machine to predict the bounce category by comparing the performance of the various binary classifiers in the One-vs-Multiclass classifier setup.