Introduction

Label Distribution Learning is a novel machine learning paradigm. A label distribution covers a certain number of labels, representing the degree to which each label describes the instance. LDL is a general learning framework which includes both single-label and multi-label learning as its special cases.

Further details about LDL can be found in the following paper.

X. Geng. Label Distribution Learning. IEEE Transactions on Knowledge and Data Engineering (IEEE TKDE), 2016, 28(7): 1734-1748.

Our alogrithms can be used freely for academic, non-profit purposes. If you intend to use it for commercial development, please contact us.

In academic papers using our codes and data, the following references will be appreciated:

[1] X. Geng. Label Distribution Learning. IEEE Transactions on Knowledge and Data Engineering (IEEE TKDE), 2016, 28(7): 1734-1748.

[2] X. Geng, C. Yin, and Z.-H. Zhou. Facial Age Estimation by Learning from Label Distributions. IEEE Transactions on Pattern Analysis and Machine Intelligence (IEEE TPAMI), 2013, 35(10): 2401-2412.

Applications of LDL

Facial Age Estimation

  1. X. Geng, Q. Wang, and Y. Xia. Facial Age Estimation by Adaptive Label Distribution Learning. In: Proceedings of the 22nd International Conference on Pattern Recognition (ICPR’14), Stockholm, Sweden, 2014, pp. 4465 - 4470.
  2. X. Geng, C. Yin, and Z.-H. Zhou. Facial Age Estimation by Learning from Label Distributions. IEEE Transactions on Pattern Analysis and Machine Intelligence (IEEE TPAMI), 2013, 35(10): 2401-2412.
  3. X. Geng, K. Smith-Miles, Z.-H. Zhou. Facial Age Estimation by Learning from Label Distributions. In: Proceedings of the 24th AAAI Conference on Artificial Intelligence (AAAI’10), Atlanta, GA, 2010, pp. 451-456.

Head Pose Estimation

  1. X. Geng and Y. Xia. Head Pose Estimation Based on Multivariate Label Distribution. In: Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition (CVPR’14),Columbus, OH, 2014, pp. 1837-1842.

Pre-release Prediction of Movies

  1. X.Geng and P.Hou. Pre-release Prediction of Crowd Opinion on Movies by Label Distribution Learning. In: Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI'15), Buenos Aires, Argentina, 2015, 3511-3517.

Multi-label Ranking

  1. X. Geng and L.-L Luo. Multilabel Ranking with Inconsistent Rankers. In: Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition (CVPR’14), Columbus, OH, 2014, pp. 3742-3747.

Multi-label Learning

  1. P. Hou, X. Geng and M.-L. Zhang. Multi-Label Manifold Learning. In: Proceedings of the 30th AAAI Conference on Artificial Intelligence (AAAI'16), Phoenix, AZ, 2016, in press.
  2. Y.-K. Li, M.-L. Zhang and X. Geng. Leveraging implicit relative labeling-importance information for effective multi-label learning. In: Proceedings of the 15th IEEE International Conference on Data Mining (ICDM'15), Atlantic City, NJ, 2015, 251-260.

Crowd Counting

  1. Z. Zhang, M. Wang, X. Geng. Crowd Counting in Public Video Surveillance by Label Distribution Learning. Neurocomputing, 2015, vol. 166: 151-163.

Matlab Code

We have implemented four LDL algorithms, namely IIS-LLD, BFGS-LLD, CPNN and LDSVR. To help you start working with LDL, we provide three demos (See iisllddemo.m, bfgsllddemo.m, cpnndemo.m, ldsverdemo.m) in this package.

Once downloaded, unzip the compressed folder to see the structure of LDL package. You'll see:

                
bfgsllddemo       -   The example of BFGSLLD algorithm.
bfgslldtrain      -   The training part of BFGSLLD algorithm.
bfgsprocess       -   Provide the target function and the gradient.
cpnn              -   The implementation of CPNN structure.
cpnndemo          -   The example of CPNN algorithm.
cpnnpredict       -   To predict the distribution by trained CPNN structure.
cpnntrain         -   The training part of CPNN algorithm.
drawdistribution  -   Draw the label distributions for comparision.
euclideandist     -   Calculate the average Euclidean distance between the predicted label
                      distribution and the real label distribution.
fidelity          -   Calculate the average fidelity between the predicted label
                      distribution and the real label distribution.
fminlbfgs         -   Finds a local minimum of a function of several variables.
iisllddemo        -   The example of IISLLD algorithm.
iislldtrain       -   The training part of IISLLD algorithm.
intersection      -   Calculate the average intersection between the predicted label
                      distribution and the real label distribution.
kernelmatrix      -   calculate the kernel matrix of vectors (samples) between two data matrices.
kldist            -   Calculate the average Kullback-Leibler divergence between the predicted label
                      distribution and the real label distribution.
ldsvrdemo         -   The example of LDSVR algorithm.
ldsvrpredict      -   To predict the distribution by trained LDSVR model.
ldsvrtrain        -   The training part of LDSVR algorithm.
ldsvrmsvr         -   The LDSVR's Multioutput SVR.
lldpredict        -   Calculate the predicted distribution of the instance X by weights.
sorensendist      -   Calculate the average Sorensen's distance between the predicted label
                      distribution and the real label distribution.
squaredxdist      -   Calculate the average squared x*x distance between the predicted label
                      distribution and the real label distribution.
Download

Data Sets

We have collected 14 real-world LDL data sets. Here are some statistics about them:

No. Dataset #Examples(n) #Features(q) #Labels(c)
1 Yeast-alpha 2,465 24 18
2 Yeast-cdc 2,465 24 15
3 Yeast-elu 2,465 24 14
4 Yeast-diau 2,465 24 7
5 Yeast-heat 2,465 24 6
6 Yeast-spo 2,465 24 6
7 Yeast-cold 2,465 24 4
8 Yeast-dtt 2,465 24 4
9 Yeast-spo5 2,465 24 3
10 Yeast-spoem 2,465 24 2
11 Human Gene 30,542 36 68
12 Natural Scene 2,000 294 9
13 SBU_3DFE 2,500 243 6
14 Movie 7,755 1,869 5
Download

There are some other LDL datasets provided by other researchers, such as

No. Dataset
15 fbp5500
16 RAF_ML
17 Twitter_ldl and Flickr_ldl