کد مقاله کد نشریه سال انتشار مقاله انگلیسی ترجمه فارسی نسخه تمام متن
4944127 1363983 2018 16 صفحه PDF ندارد دانلود رایگان
عنوان انگلیسی مقاله
Robust non-negative sparse graph for semi-supervised multi-label learning with missing labels
کلمات کلیدی
Label recovery; Semi-supervised setting; Missing labels; Semantic structure; Semantic correlation;
موضوعات مرتبط
مهندسی و علوم پایه مهندسی کامپیوتر هوش مصنوعی
پیش نمایش صفحه اول مقاله
Robust non-negative sparse graph for semi-supervised multi-label learning with missing labels
چکیده انگلیسی

•A novel label recovery method based on semi-supervised learning is proposed.•The proposed semi-supervised method learns the label matrix for labeled data imputation and unlabeled data prediction.•We used relational graphs to facilitate the label recovery.•We introduced sparse and nonnegative constraints to enhance multi-label optimization.•The proposed semi-supervised method is thoroughly tested by 5 benchmark multi-label datasets.

In multi-label learning, each instance is assumed to belong to multiple nonexclusive classes among a finite number of candidate categories. Labels are related to certain conceptual space according to their semantic similarities. Most existing approaches that deal with missing labels have the limitations in mining interdependencies among labels in the original incomplete label matrix with missing labels. In addition, semantic gaps are often neglected when features are used to facilitate label recovery. In this paper, we propose a novel label recovery method under a semi-supervised setting. The proposed method can perform label matrix imputation in the labeled space and label matrix prediction in the unlabeled space simultaneously. The semantic structure (label relationships among different instances) and the semantic correlation (label relationships among different labels) are also exploited to increase the robustness to semantic gaps and unreliable label correlations respectively. In formulating the objective function, l1-norm and nonnegative constraints are utilized to capture hidden relational graphs in semantic level and to reveal the annotation structure. An iterative mechanism is introduced to assure all variables are reliable. Intensive simulations were conducted and compared with five widely used multi-label datasets. Obtained results show that the proposed method can achieve highly competitive performance compared to other state-of-art methods.

ناشر
Database: Elsevier - ScienceDirect (ساینس دایرکت)
Journal: Information Sciences - Volume 422, January 2018, Pages 336-351
نویسندگان
, ,