Coursera.MachineLearning.Week9

Machine Learning Week9 : Anomaly Detection & Recommender Systems
GMM - 混合高斯模型算法

Anomaly Detection

1. density estimation(密度估计)

1.1 概率模型

密度估计
Anomaly detection example : Fraud detection & Manitoring

1.2 Gaussian Distribution【Normal distribution】

Gaussian distribution
Gaussian distribution examples
Parameter estimation : 对样本数据进行处理获得参数估计

1.3 Algorithm

对不同的特征,独立分布


Density estimation
Anomaly detection algorithm
Anomaly detection example
2. Building an anomaly detection system

2.1 Developing and evaluating an anomaly detection system

real-number evaluation
Training CV and Test sets
Question
Algorithm evaluation

2.2 Anomaly Detection vs Supervised Laerning

Anomaly Detection vs Supervised Laerning
examples

2.3 Choosing what features to use

某一特征的直方图(hist)
看起来像高斯分布,可以直接加入作为输入特征;有偏锋的可以尝试使用log(x+c)、x.^c等。

Non-gaussian festures
example

从判断错误的样本中观察是否可加入新的特征


Get a new feature
3. Multivariate Gaussian Distribution(多元高斯分布)

3.1 Multivariate Gaussian Distribution

特征不满足独立分布时,使用多元高斯分布
Multivariate Gaussian Distribution
Multivariate Gaussian examples.1
Multivariate Gaussian examples.2
Multivariate Gaussian examples.3
Multivariate Gaussian examples.4
Multivariate Gaussian examples.5
Multivariate Gaussian examples.6

3.2 Anomaly Detection using the Multivariate Gaussian Distribution

Get parameters μ and Sigma, Sigma likes in PCA
Steps for MGD
Relationship between MGD and original model

Original model 计算量较小,在m较少时也适用;MGD能自动捕捉特征间的相关性,n越大计算量越大,一般在m远大于n时可以考虑。
(很少出现此情况)当MGD的Sigma是奇异矩阵时(不存在逆),可能是 m>n 或者 在特征中存在重复或冗余的特征。


When to use MGD or original model
Question review
Question review


Recommender Systems

1. Predicting Movie Ratings

1.1 Problem Formulation

Example

1.2 Content Based Recommendations

n=2表示电影有两个特征,预测用户对电影的评分。


Content Based Recommender systems
Problem formulation
Optimization objective
Optimization Algorithm & Gradient descent update
2. Collaborative Filtering(协同过滤)

2.1 Collaborative Filtering
协同过滤自行学习需要使用的特征

Problem motivation
Optimization algorithm
基本的协同过滤算法

2.2 Collaborative Filtering Algorithm
合并上述两个J(),此时不需要x0=1这个固定的特征值(同时不需要θ0),因为如果系统需要一个永远为1的特征值,会在算法运行中自动调整得出。


协同过滤算法优化目标
Collaborative filtering algorithm
Question
3. Low Rank Matrix Factorization

3.1 Vectorization : Low Rank Matrix Factorization

Collaborative filtering
公式
Finding related movies

3.2 Implementation Detail : Mean Normalization

Users who have not rated any movies
Mean Normalization
Question.1
Question.2
Question.3
Question.4

你可能感兴趣的:(Coursera.MachineLearning.Week9)