Anomaly Detection and Recommender Systems

Anomaly Detection

1.1 Gaussian (Normal) distribution

• Formulation: $p(x;\mu ,{\sigma }^2)=\frac{1}{\sqrt{2\pi {\sigma }^2}}{e}^{-\frac{(x-\mu {)}^2}{2{\sigma }^2}}$
  where $\mu$ is the mean and ${\sigma }^2$ controls the variance.

• Estimate the mean value for i-th feature: ${\mu }_i=\frac{1}{m}{\sum }_{j=1}^m{x}^{(j)}$

• Estimate the variance for the i-th feature: ${\sigma }_i^2=\frac{1}{m}{\sum }_{j=1}^m(x^{(j)}-{\mu }_j{)}^2$

1.2 Program: Estimate Gaussian

• input: Training dataset matrix X, (mxn matix), with n features
• output: 1. n-dimension vector sigma2 that holds the variances of all the features, 2. n-dimension vector mu that holds the mean of all the features

function [mu sigma2] = estimateGaussian(X)
%ESTIMATEGAUSSIAN This function estimates the parameters of a 
%Gaussian distribution using the data in X
%   [mu sigma2] = estimateGaussian(X), 
%   The input X is the dataset with each n-dimensional data point in one row
%   The output is an n-dimensional vector mu, the mean of the data set
%   and the variances sigma^2, an n x 1 vector
% 
 
% Useful variables
[m, n] = size(X);
 
% You should return these values correctly
mu = zeros(n, 1);
sigma2 = zeros(n, 1);
 
% ====================== YOUR CODE HERE ======================
% Instructions: Compute the mean of the data and the variances
%               In particular, mu(i) should contain the mean of
%               the data for the i-th feature and sigma2(i)
%               should contain variance of the i-th feature.
%


mu = mean(X);
sigma2 = var(X, 1); % 1 -- 1/m rather than 1/(m-1)
mu = mu(:);
sigma2 = sigma2(:);
% =============================================================
% method 2:
% vectorized implementation
mu = 1/m * sum(X);
sigma2 = 1/m * sum((X - repmat(mu, m, 1)).^2);
%repmat(mu, m, 1) means create a mx1 matrix, all the elements is mu 
 
end

1.3 Selecting the threshold, $ϵ$

• use a cross validation set { ( x c v ( 1 ) , y c v ( 1 ) ) , … , ( x c v ( m ) , y c v ( m ) ) } \{(x^{(1)}_{cv},y^{(1)}_{cv}),\ldots, (x^{(m)}_{cv},y^{(m)}_{cv})\} {(xcv(1)​,ycv(1)​),…,(xcv(m)​,ycv(m)​)},
• where the label y=1 corresponds to an anomalous example
• For each cross validation example, we will compute $p(x_{cv}^{(i)})$
• $F_1$ score: tells you how well you’re doing on finding the ground truth anomalies given a certain threshold (how many examples the current threshold classifies correctly and incorrectly)
  • The score is computed using precision (prec) and recall (rec):
  • $F_1=\frac{2(prec)(rec)}{prec+rec}$
  • $prec=\frac{tp}{tp+fp}$
  • $rec=\frac{tp}{tp+fn}$
• true positives: the ground truth label says it’s an anomaly and our algorithm correctly classified it as an anomaly
• false positives: the ground truth label says it’s not an anomaly, but our algorithm incorrectly classified it as an anomaly (右上角)
• false negatives: the ground truth label says it’s an anomaly, but our algorithm incorrectly classified it as not being anomalous.

selectThreshold: select the “best” $ϵ$ based on the F1 score

function [bestEpsilon bestF1] = selectThreshold(yval, pval)
%SELECTTHRESHOLD Find the best threshold (epsilon) to use for selecting
%outliers
%   [bestEpsilon bestF1] = SELECTTHRESHOLD(yval, pval) finds the best
%   threshold to use for selecting outliers based on the results from a
%   validation set (pval) and the ground truth (yval).
%

% given: the cross validation lable yval and possibility of pval

bestEpsilon = 0;
bestF1 = 0;
F1 = 0;

stepsize = (max(pval) - min(pval)) / 1000;
for epsilon = min(pval):stepsize:max(pval)
    
    % ====================== YOUR CODE HERE ======================
    % Instructions: Compute the F1 score of choosing epsilon as the
    %               threshold and place the value in F1. The code at the
    %               end of the loop will compare the F1 score for this
    %               choice of epsilon and set it to be the best epsilon if
    %               it is better than the current choice of epsilon.
    %               
    % Note: You can use predictions = (pval < epsilon) to get a binary vector
    %       of 0's and 1's of the outlier predictions
    
    cvPredictions = (pval < epsilon);
    
    tp = sum ((cvPredictions == 1) & (yval == 1));
    fp = sum ((cvPredictions == 1) & (yval == 0));
    fn = sum ((cvPredictions == 0) & (yval == 1));
    
    prec = tp / (tp + fp);
    rec = tp / (tp + fn);
    F1 = 2 * prec * rec / (prec + rec);

    % =============================================================

    if F1 > bestF1
       bestF1 = F1;
       bestEpsilon = epsilon;
    end
end

end

Recommender System

2.1 Collaborative fitlering objective function and gradient

Data decription:

• The matrix Y (a num_movies x num_users matrix) stores the ratings $y^{(i,j)}$ (from 1 to 5)
• The matrix R is an binary-valued indicator matrix, where R(i,j) = 1 if user j gave a rating to movie i, and R(i,j) = 0 otherwise

The objective of collaborative filtering:

to predict movie ratings for the movies that users have not yet rated, that is, the entries with R(i,j) = 0

Collaborative filtering learning algorithm:

• considers a set of n-dimensional feature vectors (for movie) $x^{(1)},\dots ,x^{(n_m)}$ and ${\theta }^{(1)},\dots ,{\theta }^{(n_u)}$ (for audience)
• prediction for movie i by user j is: $y^{(i,j)}=({\theta }^{(j)}{)}^T{x}^{(i)}$
• the parameters to the function (i.e., the values that you are trying to learn) are X and Theta (two variables to learn)
• cost function: (without regularization)
  $J\left(x^{(i)},\dots ,x^{(n_m)},{\theta }^{(1)},\dots ,{\theta }^{(n_u)}\right)=\frac{1}{2}{\sum }_{(i,j):r(i,j)=1}{\left(({\theta }^{(j)}{)}^T{x}^{(i)}-y^{(i,j)}\right)}^2$

2.2 Collaborative filtering gradient

• $\frac{\partial J}{\partial x_k^{(i)}}={\sum }_{j:r(i,j)=1}\left(({\theta }^{(j)}{)}^T{x}^{(i)}-y^{(i,j)}\right){\theta }_k^{(j)}$
• $\frac{\partial J}{\partial {\theta }_k^{(j)}}={\sum }_{i:r(i,j)=1}\left(({\theta }^{(j)}{)}^T{x}^{(i)}-y^{(i,j)}\right)x_k^{(i)}$

To vectorize the above expression, you can start by indexing into Theta and Y to select only the elements of interest (that is, those with $r(i,j)=1$):
Explain: can set idx = find(R(i,:)==1) to be a list of all the users that have rated movie . This will allow you to create the temporary matrices Theta_temp = Theta(idx,:) and Y_temp = Y(i,idx) that index into Theta and Y to give you only the set of users which have rated the i-th movie
the derivatives for the feature vector fof the i-th movie is:

Regularized cost function

• $J\left(x^{(i)},\dots ,x^{(n_m)},{\theta }^{(1)},\dots ,{\theta }^{(n_u)}\right)=\frac{1}{2}{\sum }_{(i,j):r(i,j)=1}{\left(({\theta }^{(j)}{)}^T{x}^{(i)}-y^{(i,j)}\right)}^2+\left(\frac{\lambda }{2}{\sum }_{j=1}^{n_u}{\sum }_{k=1}^n({\theta }_k^{(j)}{)}^2\right)+\left(\frac{\lambda }{2}{\sum }_{i=1}^{n_m}{\sum }_{k=1}^n(x_k^{(i)}{)}^2\right)$
• $\frac{\partial J}{\partial x_k^{(i)}}={\sum }_{j:r(i,j)=1}\left(({\theta }^{(j)}{)}^T{x}^{(i)}-y^{(i,j)}\right){\theta }_k^{(j)}+\lambda x_k^{(i)}$
• $\frac{\partial J}{\partial {\theta }_k^{(j)}}={\sum }_{i:r(i,j)=1}\left(({\theta }^{(j)}{)}^T{x}^{(i)}-y^{(i,j)}\right)x_k^{(i)}+\lambda {\theta }_k^{(j)}$

完整代码：

function [J, grad] = cofiCostFunc(params, Y, R, num_users, num_movies, ...
                                  num_features, lambda)
%COFICOSTFUNC Collaborative filtering cost function
%   [J, grad] = COFICOSTFUNC(params, Y, R, num_users, num_movies, ...
%   num_features, lambda) returns the cost and gradient for the
%   collaborative filtering problem.
%

% Unfold the U and W matrices from params
X = reshape(params(1:num_movies*num_features), num_movies, num_features);
Theta = reshape(params(num_movies*num_features+1:end), ...
                num_users, num_features);

            
% You need to return the following values correctly
J = 0;
X_grad = zeros(size(X));
Theta_grad = zeros(size(Theta));

% ====================== YOUR CODE HERE ======================
% Instructions: Compute the cost function and gradient for collaborative
%               filtering. Concretely, you should first implement the cost
%               function (without regularization) and make sure it is
%               matches our costs. After that, you should implement the 
%               gradient and use the checkCostFunction routine to check
%               that the gradient is correct. Finally, you should implement
%               regularization.
%
% Notes: X - num_movies  x num_features matrix of movie features
%        Theta - num_users  x num_features matrix of user features
%        Y - num_movies x num_users matrix of user ratings of movies
%        R - num_movies x num_users matrix, where R(i, j) = 1 if the 
%            i-th movie was rated by the j-th user
%
% You should set the following variables correctly:
%
%        X_grad - num_movies x num_features matrix, containing the 
%                 partial derivatives w.r.t. to each element of X
%        Theta_grad - num_users x num_features matrix, containing the 
%                     partial derivatives w.r.t. to each element of Theta
%


% sum(sum(R.*M)) is the sum of all the elements of M for which the corresponding element in R equals 1.

J = 1/2 * sum(sum((R.* ((X*Theta') - Y)).^2));

X_grad = (R .* (X*Theta' - Y)) * Theta;
Theta_grad = (R .* (X*Theta' - Y))' * X;

% With regularization
J = J + lambda/2 * (sum(sum(Theta.^2)) + sum(sum(X.^2)));
X_grad = X_grad + lambda * X;
Theta_grad = Theta_grad + lambda * Theta;
% =============================================================
grad = [X_grad(:); Theta_grad(:)];

end