Thisis an open book/Internet homework, but you should not interact with your classmatesThis is a bonus problem set with a single harddeadlineYouwill receive a non-transferrable 2% bonus for using LATEXThere are 2 problems over 3Theproblems are assigned the same weight for the overallDue Monday April 29, 2019 11:59pm EST (Monday May 6, 2019 for DistanceLearning Students)Pleaseturn in your solution using the LATEX template as the first page代写数学考试K-means clustering with the Euclidean distance inherently assumes that each pair of clusters is linearly separable, which may not be the case in practice. In this problem you will derive a strategy for dealing with this limitation that we did not discuss in class. Specifically, you will show that like so many other algorithms we have discussed in class, K-means can be “kernelized.” In the following,we consider a dataset {xi}N .Let zij≜ 1 xi j , where j denotes the jth  Show that the rule for updating thejth cluster center mj given this cluster assignment can be expressed asNmj = αijxi (1)i=1by expressing αij as a function of the zij’s.Giventwo points x1 and x2, show that ∥x1 − x2∥2 can be computed using only linear com-binations of inner products.Given the results of the previous parts, show how to compute ∥xi−mj∥2 using only(linearcombinations of) inner products between the data points {xi}N .Describehow to use the results from the previous parts to “kernelize” the K-means clustering algorithm described inIn this problem we consider the scenario seen in class, where x is drawn uniformly on [−1, 1] and y = sin(πx), for which we are given N = 2 training samples. Here, we will consider an alternative approach to fitting a line to the data based on Tikhonov regularization. Specifically, we lety = y1 y2A = 1 x11 x2] θ = [ b] (2)We will then consider Tikhonov regularized least squares estimators of the formθˆ ≜ (A⊺A + Γ⊺Γ)−1A⊺y. (3)Howshould we set Γ to reduce this estimator to fitting a constant function (i.e., finding an h(x) of the form h(x) = b)? (Hint: For the purposes of this problem, it is sufficient to set Γ in a way that just makes a 0. To make a = 0 exactly requires setting Γ in a way that makes the matrix A⊺A + Γ⊺Γ singular, but note that this does not mean that the regularized least-squares optimization problem cannot be solved; you must just use a different formula than the one in (3).Howshould we set Γ to reduce this estimator to fitting a line of the form h(x) = ax + b that passes through the observed data points (x1, y1) and (x2, y2)?(Optional)Play around and see if you can find a (diagonal) matrix Γ that results in a smaller risk than either of the two approaches we discussed in  You will need to do this numeri- cally using Python or MATLAB. Report the Γ that gives you the best results. (You can restrict your search to diagonal Γ to simplify this.)最先出自315代写 cs代写 作业代写 数学代写合作：315代写

所有的数学代写范围：essayghost为美国、加拿大、英国、澳洲的留学生提供数学代写、math代写、数学作业代写、math作业代写、数学代考、math代考等留学生数学作业代写、exam代考服务。