Learning the Shared Subspace for Multi-task Clustering and Transductive Transfer Classification

Multi-task TSK fuzzy system modeling using inter-task correlation information

“…(18) and (19) can measure the influence of the independence information and correlation information, respectively. Furthermore, we can define an index, i.e., IND/CORR, to reflect the influence of both kinds of information.…”

“…Based on such a principle, multi-task learning methods emerge and are attracting more and more attentions in the fields of machine learning, data mining and pattern recognition. The existing multi-task learning methods can be categorized into three types: (1) Multi-task classification learning [2,[5][6][7][14][15][16]21,22,33,36,43]; (2) Multi-task clustering [3,18,19,23,42,44]; and (3) Multi-task regression learning [8,27,34,37,45]. Although these works have demonstrated the significance of multi-task learning and certain effectiveness in different real-world applications, the current multi-task learning methods still cannot keep up with the real-world requirements, particularly in regression tasks.…”

Effectively Constructing Reliable Data for Cross-Domain Text Classification

“…Pan et al [20] proposed a dimensionality reduction approach to find out this latent feature space, in which supervised learning algorithms can be applied to train classification models. Gu et al [7] learnt the shared subspace among multiple domains for clustering and transductive transfer classification. In their problem formulation, all the domains have the same cluster centroid in the shared subspace.…”

“…There are many papers appear in recent years, and they can be grouped into three types of techniques used for knowledge transfer, namely feature selection based [11,4,29], feature space mapping [20,22,7], weight based [5,6,10].…”

“…These works include feature selection learning [4,29,21,25], transfer learning via dimensionality reduction [20,7], model combination [6,30] and sample selection bias [26,28] et al Unlike previous research, our work is motivated by the observation that although the trained model from source domain may give poor performance over all the target domain data, it might perform well on an elaborately selected subset. Let's look at an intuitive example in Table 1 If we compute the prediction accuracy over all the test documents, we can only obtain a low accuracy 65%.…”