Unsupervised Learning

2018 · Machine Learning

Github Repo

Principal Components Analysis (PCA) of colored faces

• The mean value of all 415 faces

• Use numpy.linalg.svd and np.linalg.eig to implement PCA
• The first four eigenfaces of the first four largest eigenvalues

Visualization of Chinese word embedding

• Use gensim to implement word embedding
• Use TSNE to implement dimension reduction
• Plot the result 2D figure

Image clustering

• Determine if two images come from the same dataset
• The result of different feature extraction methods

	Public set score	Private set score
TSNE to all vectors to 2 dimension	0.02954	0.02910
Autoencoder+DNN 128 64 epochs = 10, k-means	0.52631	0.52531
Autoencoder+DNN 128 64 epochs = 200, k-means	0.96402	0.96237

\[F1-\text{score}=2\cdot \frac{p\cdot r}{p+r}\]

where $p=\frac{tp}{tp+fp}$, $r=\frac{tp}{tp+fn}$

	prediction positive	prediction negative
ground true positive	true positive (tp)	false negative (fn)
ground true negative	false positive (fp)	true negative (tn)

• Visualize all features with their labels

Result

• Achieved 225/333 (Top $68\%$) rank in the Kaggle competition