/Project_2-Distance-Metrics

The second project

Primary LanguagePython

Project_2-Distance-Metrics

  • 处理数据集
  • KNN 分类：Deep Learning Feature
  • 尝试不同的距离的度量
    • 切比雪夫距离
    • 欧氏距离
    • 曼哈顿距离
    • 余弦距离
  • 至少用一种metric learning的方法，来提高KNN的结果
  • 写报告
    • 实验越多越充分
    • 注意参考文献和引用

参数说明：

  • 需安装包metric_learn，API地址：http://metric-learn.github.io/metric-learn/metric_learn.rca.html
  • mltype：选择的metric learn
    • None: 不使用任何metric learn
    • MMC: Mahalanobis Metric for Clustering
    • LMNN: Large Margin Nearest Neighbor
    • NCA: Neighborhood Components Analysis
    • LFDA: Local Fisher Discriminant Analysis
    • MLKR: Metric Learning for Kernel Regression
    • ITML: Information Theoretic Metric Learning
    • LSML: Least Squares Metric Learning
    • SDML: Sparse Determinant Metric Learning
    • RCA: Relative Components Analysis
  • distance：KNN使用的距离的类型
    • euclidean
    • manhattan
    • chebyshev
    • cosine
  • neighbors：KNN中的K

任务分配

姓名 任务
李东岳 None, MMC, LMNN
李杰宇 NCA, LFDA, MLKR
王皓轩 ITML, report
陈鸿滨 LSML, SDML, RCA
  • 代码可能会有问题，不过应该都只是API上调用的问题，大家看文档改一下就可以了，有的可能要用带_Supervised()的那个函数
  • 有的算法会有爆栈的可能
  • 对每个人负责的算法，那几个参数可以用gridsearch来调，结果尽量高过None
  • 尽量在这周之内弄完

实验结果

euclidean manhattan chebyshev cosine
None 88.94 88.08 78.50 89.84
MMC - - - -
LMNN 90.48 89.49 76.22 90.59
NCA - - - -
LFDA 84.60 64.49 89.99 89.74
MLKR - - - -
ITML - - - -
LSML 79.87 9.02 19.98 85.87
SDML - - - -
RCA
  • None（distance=euclidean，neighbors=50）结果为86.45，neighbors可以考虑先用gridsearch，针对每种距离选取一个最好的neighbors的取值，然后所有的度量学习方法在过KNN的时候都使用这个值，如果有时间的话再做它对acc的影响的实验

Results: Li Dongyue

KNN

Neighbors_Num euclidean manhattan chebyshev cosine
1 87.59 87.44 75.97 88.31
3 88.37 87.87 77.02 89.17
5 88.94 88.08 78.50 89.84
8 88.84 87.85 78.56 90.19
10 88.85 87.63 78.45 90.24
15 88.53 86.97 78.36 90.07
20 87.96 86.42 77.97 89.78
30 87.36 85.76 77.01 89.54
40 86.90 84.91 76.10 89.16
50 86.45 84.30 75.08 89.00
100 84.47 81.84 72.29 87.62

PCA+KNN

Neighbors_Num euclidean manhattan chebyshev cosine
1 87.58 78.99 84.53 87.84
3 88.36 77.77 85.92 88.94
5 88.93 76.40 86.99 89.30
8 88.84 74.60 87.40 89.33
10 88.85 73.44 87.34 89.41
15 88.52 70.96 87.19 89.62
20 87.95 68.73 87.05 89.17
30 87.36 64.72 86.60 89.01
40 86.90 61.53 86.04 88.64
50 86.44 58.84 85.62 88.19
100 84.46 48.99 83.91 86.77

LMNN+KNN

Neighbors_Num euclidean manhattan chebyshev cosine
1 88.76 87.78 74.19 88.61
3 89.52 88.74 74.81 89.95
5 90.48 89.49 76.22 90.59
8 90.60 89.36 76.83 90.91
10 90.52 89.60 76.79 91.00
15 90.66 89.38 76.61 91.06
20 90.37 88.98 76.10 90.96
30 89.91 88.54 75.42 90.71
40 89.57 87.93 74.55 90.61
50 89.20 87.47 73.95 90.27
100 87.71 85.87 71.40 89.61

PCA+LMNN+KNN

Neighbors_Num euclidean manhattan chebyshev cosine
1 88.78 86.52 85.55 88.43
3 89.53 86.80 86.99 89.41
5 90.48 87.41 88.11 89.81
8 90.61 87.24 88.59 90.13
10 90.53 87.14 88.75 90.17
15 90.66 86.66 88.86 90.31
20 90.37 86.29 88.86 90.09
30 89.90 85.46 88.59 89.89
40 89.57 84.43 88.21 89.54
50 89.20 84.54 88.06 89.18
100 87.71 80.43 86.49 88.26

LDFA+KNN

Neighbors_Num euclidean manhattan chebyshev cosine
1 84.95 67.64 87.74 88.92
3 84.94 65.61 89.22 89.33
5 85.16 63.87 89.87 89.71
8 85.18 62.24 90.22 89.20
10 84.99 61.40 90.15 89.08
15 84.52 59.98 90.29 88.86
20 84.34 58.59 90.24 88.64
30 83.88 56.32 90.12 88.25
40 83.52 54.97 90.07 87.72
50 83.29 53.78 89.83 87.32
100 81.70 50.03 89.58 85.88

LSML+KNN

Neighbors euclidean manhattan chebyshev cosine
1 15.61 16.71 19.82 82.58
3 9.96 10.86 18.60 84.25
5 7.99 9.00 20.00 85.86
8 6.90 7.88 20.74 86.35
10 6.19 7.03 20.62 86.67
15 5.28 5.83 20.28 86.42
20 5.18 5.49 20.30 86.23
30 4.86 5.07 19.63 84.97
40 4.68 4.98 19.01 83.54
50 4.50 4.88 18.40 82.00
100 3.82 4.42 16.58 73.78