User attribute discovery with missing labels. (January 2018)

• Record Type: Journal Article
• Authors: Cong, Yang; Sun, Gan; Liu, Ji; Yu, Haibin; Luo, Jiebo
• Is Part Of: Pattern recognition. Volume 73(2018:Jan.)
• Page Start: 33
• Page End: 46
• Publication Date: 2018-01
• Subjects: User attribute -- Smart sensor -- Multi-task learning -- Semi-supervised learning -- Missing labels -- Low rank; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.07.012
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a new problem, i.e., user attribute discovery via smart sensor data. We design a new a semi-supervised multi-task learning model (S2MTL) for user attribute discovery with missing label. To reduce the model complexity of high-dimensional data, we learn the mapping feature dictionary and attribute space information simultaneously. We also build a new smart building dataset. Abstract: In this paper, we focus on user attribute analysis by recasting such a problem as a multi-task learning issue, where each attribute is considered as an independent task. In comparison with traditional data analysis, the missing labels problem broadly presents for smart sensor data due to some objective / subjective factors, where the label incompleteness increases the difficulty significantly. Therefore, we design a semi-supervised multi-task learning model (S2MTL) to handle the missing labels issue. For modeling, we integrate the matrix factorization to learn the mapping feature dictionary and attribute space information simultaneously, and adopt the pairwise affinity similarity to incorporate the unlabeled data information, where the low rank property and model efficiency can be well controlled. For model optimization, we convert our model as two individual convex subproblems with one non-smooth, and implement an alternating direction method to generate an efficient optimal solution. State-of-the-art models have validated the effectiveness and efficiency of our proposed model via extensive experiments and comparisons, on two public datasets and our new smart building dataset.