A Dynamic Ensemble Framework for Mining Textual Streams with Class Imbalance

Algorithm 1

Clustering forest.

Output: Class label of the testing instance 𝑥

Input:  𝐷_𝑡: The data chunk at each time stamp

   𝑀: The number of CTs in CF𝑓^𝐸

   𝑀_max: The maximum number of CTs selected in the clustering forest 𝑓^𝐸_max

   𝑓^𝐸: The current ensemble model

   𝐸_𝑡: Misclassified subset at the 𝑡-th time stamp

   𝑅_𝑡: Rare-class subset at the 𝑡-th time stamp.

   𝑃^𝑅_𝑡: The Rare-class subset as a proportion of the training chunk; 𝑝^𝑅_𝜃: The threshold of 𝑝^𝑅

 (1) at the 𝑡-th time stamp

 (2) built the training set by Algorithm 2

 (2) create a new CT 𝑓_𝑡 using 𝐷^train_𝑡

 (3) if 𝑡<𝑀_max

 (4) 𝑓^𝐸←𝑓^𝐸∪𝑓_𝑡

 (5) Endif

 (6) if 𝑡>𝑀_max

 (7) 𝑓^𝐸←𝑓_{𝑡−𝑀max+1}∪⋯∪𝑓_𝑡

 (8) Endif

 (9) compute the accuracy weight of each clustering tree

 (10) UPDATE(𝑓^𝐸) based on the adaptive selection method

 (11) obtain the misclassified subset 𝐸_𝑡 and the rare-class subset 𝑅_𝑡

 (12) for each testing instance 𝑥 do:

 (13) PREDICT(𝑥) by the voting method.

 (14) Endfor.