候選標記信息感知的偏標記學習算法
doi: 10.11999/JEIT181059 cstr: 32379.14.JEIT181059
-
國家數(shù)字交換系統(tǒng)工程技術研究中心 ??鄭州 ??450002
基金項目: 國家自然科學基金(61601513)
Candidate Label-Aware Partial Label Learning Algorithm
-
National Digital Switching System Engineering & Technological R&D Center, Zhengzhou 450002, China
Funds: The National Natural Science Foundation of China (61601513)
-
摘要: 在偏標記學習中，示例的真實標記隱藏在由一組候選標記組成的標記集中?，F(xiàn)有的偏標記學習算法在衡量示例之間的相似度時，只基于示例的特征進行計算，缺乏對候選標記集信息的利用。該文提出一種候選標記感知的偏標記學習算法(CLAPLL)，在構建圖的階段有效地結合候選標記集信息來衡量示例之間的相似度。首先，基于杰卡德距離和線性重構，計算出各個示例的標記集之間的相似度，然后結合示例相似度和標記集的相似度構建相似度圖，并通過現(xiàn)有的基于圖的偏標記學習算法進行學習和預測。3個合成數(shù)據(jù)集和6個真實數(shù)據(jù)集上實驗結果表明，該文方法相比于基線算法消歧準確率提升了0.3%～16.5%，分類準確率提升了0.2%～2.8%。
-
關鍵詞:
- 偏標記學習 /
- 弱監(jiān)督學習 /
- 消歧 /
- 杰卡德距離 /
- 線性重構
Abstract: In partial label learning, the true label of an instance is hidden in a label-set consisting of a group of candidate labels. The existing partial label learning algorithm only measures the similarity between instances based on feature vectors and lacks the utilization of the candidate labelset information. In this paper, a Candidate Label-Aware Partial Label Learning (CLAPLL) method is proposed, which combines effectively candidate label information to measure the similarity between instances during the graph construction phase. First, based on the jaccard distance and linear reconstruction, the similarity between the candidate labelsets of instances is calculated. Then, the similarity graph is constructed by combining the similarity of the instances and the label-sets, and then the existing graph-based partial label learning algorithm is presented for learning and prediction. The experimental results on 3 synthetic datasets and 6 real datasets show that disambiguation accuracy of the proposed method is 0.3%～16.5% higher than baseline algorithm, and the classification accuracy is increased by 0.2%～2.8%. -
表 1 候選標記信息感知的偏標記學習算法偽代碼
輸入：偏標記數(shù)據(jù)集$D = \left\{ {({X_i},{S_i})|1 \le i \le m} \right\}$，最近鄰樣本數(shù) $k$，標記相似度權重$\alpha $ 訓練階段： 1 對特征矩陣${\text{X}} \in {{\text{R}}^{m \times d}}$進行Z-score歸一化； 2 根據(jù)式(1)求${{\text{w}}_j}$； 3 根據(jù)${{\text{w}} _j}$構建相似度圖${G_i}(V,E)$； 4 switch v； case Jaccard：根據(jù)式(3)計算${{\text{u}}_j}$，并構建候選標記集相似度 圖${G_{\rm{c}}}(i,j)$, (CAP-J算法)； case linear：根據(jù)式(4)計算${{\text{u}}_j}$，并構建候選標記集相似度 圖${G_{\rm{c}}}(i,j)$, (CAP-L算法)； end switch 5 根據(jù)式(7)計算最終相似度圖$G(i,j)$； 6 結合現(xiàn)有圖模型偏標記學習算法進行消歧，得到消歧結果 $\mathop D\limits^ \wedge = \left\{ {({X_i},{{\widehat y}_i})|1 \le i \le m} \right\}$； 測試階段： 7 對于未見示例${x^*}$，根據(jù)式(8)計算得分類結果； 輸出：消歧結果$\mathop D\limits^ \wedge = \left\{ {({X_i},{{\widehat y}_i})|1 \le i \le m} \right\}$和分類結果${y^*}$。 
下載: 導出CSV
表 2 基線算法和本文算法復雜度比較
算法復雜度 實際復雜度 基線算法 $O({d^{\,\; 2} }{n^3}\lg (n))$ $O({d^{\,\; 2} }{n^3}\lg (n))$ 本文算法(CAP-J) $O({d^{\,\; 2} }{n^3}\lg (n) + (s + 1){k^2})$ $O({d^{\,\; 2} }{n^3}\lg (n))$ 本文算法(CAP-L) $O({d^{\,\; 2} }{n^3}\lg (n) + (sk + 1){k^2})$ $O({d^{\,\; 2} }{n^3}\lg (n))$
下載: 導出CSV
表 3 真實偏標記數(shù)據(jù)集的特征
數(shù)據(jù)集 樣本數(shù) 特征數(shù) 類別標記數(shù) 候選標記數(shù) 平均 最小 最大 Lost 1122 108 16 2.23 1 3 Birdsong 4998 38 13 2.18 1 4 MSRSCv2 1758 48 23 3.16 1 7 FG-NET 1002 262 78 7.48 2 11 Yahoo! News 22991 163 219 1.91 1 5 Soccer Player 17472 279 171 2.09 1 11
下載: 導出CSV
表 4 合成偏標記數(shù)據(jù)集的特征
數(shù)據(jù)集 樣本數(shù) 特征數(shù) 類別標記數(shù) 參數(shù)設置 Ecoli 336 7 8 p={0.1, 0.2, 0.3, 0.4,0.5, 0.6, 0.7, 0.8} r={1, 2, 3, 4, 5} Movement 360 90 15 CTG 2126 21 10
下載: 導出CSV
表 5 不同算法在真實偏標記數(shù)據(jù)集上的消歧準確率(%)
數(shù)據(jù)集 消歧準確率(mean±std.) Lost MSRCv2 BirdSong FG-NET Soccer Player Yahoo! News PLKNN 67.54±0.09 51.00±0.09 68.69±0.04 11.06±0.13 52.60±0.02 66.06±0.02 CAP-JKNN 73.60±0.10 62.19±0.08 77.14±0.04 14.71±0.15 69.55±0.01 80.00±0.02 CAP-LKNN 73.38±0.13 61.88±0.09 76.67±0.04 14.81±0.17 69.22±0.02 79.78±0.05 PLKNN(監(jiān)督) 84.93±0.04 73.07±0.02 84.29±0.14 14.94±0.05 90.65±0.03 91.21±0.03 IPAL 84.01±0.15 70.58±0.15 83.61±0.04 15.28±0.19 67.65±0.03 84.99±0.05 CAP-JIPAL 85.58±0.17 71.25±0.20 84.22±0.04 15.40±0.19 67.94±0.02 85.33±0.04 CAP-LIPAL 85.39±0.24 70.92±0.12 84.40±0.05 14.86±0.17 67.89±0.07 85.21±0.03 IPAL(監(jiān)督) 85.43±0.32 76.43±0.22 85.92±0.10 15.53±0.18 71.43±0.05 86.43±0.06 LALO 75.05±1.24 59.42±0.89 78.14±0.75 15.92±0.69 – – CAP-JLALO 76.80±1.11 59.48±1.09 78.02±0.81 15.69±0.75 – – CAP-LLALO 80.22±1.08 59.72±0.82 78.24±0.64 15.76±0.94 – – LALO(監(jiān)督) 84.53±1.53 60.04±1.14 79.25±0.88 16.13±0.62 – –
下載: 導出CSV
表 6 不同算法在真實偏標記數(shù)據(jù)集上的分類準確率(%)
數(shù)據(jù)集 消歧準確率(mean±std.) Lost MSRCv2 BirdSong FG-NET Soccer Player Yahoo! News PLKNN 61.48±0.78 44.12±0.36 64.66±0.23 5.58±0.42 49.55±0.04 58.30±0.06 CAP-JKNN 64.01±0.65 46.35±0.38 66.01±0.26 6.24±0.38 50.77±0.09 61.18±0.05 CAP-LKNN 63.58±0.72 46.14±0.48 65.88±0.21 5.74±0.56 50.43±0.09 60.50±0.12 PLKNN(監(jiān)督) 69.26±0.48 51.33±0.30 68.49±0.13 6.98±0.21 54.26±0.05 61.53±0.08 IPAL 73.18±0.79 53.08±0.33 71.09±0.33 5.28±0.55 54.84±0.10 65.88±0.14 CAP-JIPAL 73.95±0.68 53.35±0.50 71.34±0.30 5.45±0.60 55.00±0.10 66.02±0.16 CAP-LIPAL 73.44±0.68 52.61±0.71 71.60±0.26 5.89±0.57 54.46±0.18 66.02±0.18 IPAL (監(jiān)督) 75.04±0.82 55.71±0.46 72.05±0.27 5.95±0.62 55.38±0.13 66.83±0.15 LALO 72.15±3.04 50.13±2.03 72.99±1.54 6.11±1.61 – – CAP-JLALO 73.02±2.88 49.23±2.10 73.00±1.62 5.96±1.19 – – CAP-LLALO 74.84±2.20 50.27±3.19 73.37±1.50 6.76±1.64 – – LALO(監(jiān)督) 76.68±2.19 52.31±2.49 74.87±1.26 7.03±1.29 – –
下載: 導出CSV
-
HüLLERMEIER E and BERINGER J. Learning from ambiguously labeled examples[J]. Intelligent Data Analysis, 2006, 10(5): 419–439. doi: 10.3233/IDA-2006-10503 SONG Jingqi, LIU Hui, GENG Fenghuan, et al. Weakly-supervised classification of pulmonary nodules based on shape characters[C]. The 14th International Conference on Dependable, Autonomic and Secure Computing, The 14th International Conference on Pervasive Intelligence and Computing, The 2nd International Conference on Big Data Intelligence and Computing and Cyber Science and Technology Congress, Auckland, New Zealand, 2016: 228–232. TANG Caizhi and ZHANG Minling. Confidence-rated discriminative partial label learning[C]. The 31st AAAI Conference on Artificial Intelligence, San Francisco, USA, 2017: 2611–2617. TODA T, INOUE S, and UEDA N. Mobile activity recognition through training labels with inaccurate activity segments[C]. The 13th International Conference on Mobile and Ubiquitous Systems: Computing, Networking and Services, Hiroshima, Japan, 2016: 57–64. YU Fei and ZHANG Minling. Maximum margin partial label learning[J]. Machine Learning, 2017, 106(4): 573–593. doi: 10.1007/s10994-016-5606-4 LUO Jie and ORABONA F. Learning from candidate labeling sets[C]. The 23rd International Conference on Neural Information Processing Systems, Vancouver, Canada, 2010: 1504–1512. ZHANG Minling and YU Fei. Solving the partial label learning problem: An instance-based approach[C]. The 24th International Conference on Artificial Intelligence, Buenos Aires, Argentina, 2015: 4048–4054. FENG Lei and AN Bo. Leveraging latent label distributions for partial label learning[C]. The Twenty-Seventh International Joint Conference on Artificial Intelligence, Stockholm, Sweden, 2018: 2107–2113. COUR T, SAPP B, and TASKAR B. Learning from partial labels[J]. Journal of Machine Learning Research, 2011, 12: 1501–1536. ZHOU Zhihua. A brief introduction to weakly supervised learning[J]. National Science Review, 2018, 5(1): 44–53. doi: 10.1093/nsr/nwx106 TOLDO R and FUSIELLO A. Robust multiple structures estimation with J-linkage[C]. The 10th European Conference on Computer Vision, Marseille, France, 2008: 537–547. DUA D and TANISKIDOU E K. UCI machine learning repository[EB/OL]. http://archive.ics.uci.edu/ml, 2017. ZENG Zinan, XIAO Shijie, JIA Kui, et al. Learning by associating ambiguously labeled images[C]. 2013 IEEE Conference on Computer Vision and Pattern Recognition, Portland, USA, 2013: 708–715. GUILLAUMIN M, VERBEEK J, and SCHMID C. Multiple instance metric learning from automatically labeled bags of faces[C]. The 11th European Conference on Computer Vision, Heraklion, Greece, 2010: 634–647. ZHANG Minling, ZHOU Binbin, and LIU Xuying. Partial label learning via feature-aware disambiguation[C]. The 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, USA, 2016: 1335–1344. BRIGGS F, FERN X Z, and RAICH R. Rank-loss support instance machines for MIML instance annotation[C]. The 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Beijing, China, 2012: 534–542. LIU Liping and DIETTERICH T G. A conditional multinomial mixture model for superset label learning[C]. The 25th International Conference on Neural Information Processing Systems, Lake Tahoe, USA, 2012: 548–556. ZHANG Minling, YU Fei, and TANG Caizhi. Disambiguation-free partial label learning[J]. IEEE Transactions on Knowledge and Data Engineering, 2017, 29(10): 2155–2167. doi: 10.1109/TKDE.2017.2721942 ZHANG Minling and YU Fei. Solving the partial label learning problem: An instance-based approach[C]. The 24th International Conference on Artificial Intelligence, Buenos Aires, Argentina, 2015: 4048–4054. -
圖(9) / 表(6)
計量
- 文章訪問數(shù): 2834
- HTML全文瀏覽量: 1509
- PDF下載量: 72
- 被引次數(shù): 0
下載:
