name: layout-general layout: true class: left, top --- count: false background-image: url('img/stat.jpg') background-size: cover class: top, center, title-slide <img src="img/X.png" align="middle" height="110" width="100"> .white[........] <img src="img/logo_lmd.png" align="right" height="90"> <img src="img/logo_ens-remove.png" align="left" height="90"> <hbr> <hr class="L1"> <hbr> ## Frequentist vs Bayesian Statistics <hr class="L2"> #### Opening Talk of MAC Seminar of Statistics <img src="img/mac.png" align="middle" height="130"> <hbr> ### Sothea .textsc[Has], PhD --- ## What's statistics? - Statistics is a branch of mathematics that deals with .stress[collecting], .stress[organizing], .stress[analyzing], .stress[interpreting], .stress[presenting] data, .stress[designing experiments/surveys], and .stress[drawing valid conclusions from data]. .center[ <img src="img/def_stat.jpg" align="middle" height="320"> ] --- count: false ## What's statistics? - Statistics is a branch of mathematics that deals with .stress[collecting], .stress[organizing], .stress[analyzing], .stress[interpreting], .stress[presenting] data, .stress[designing experiments/surveys], and .stress[drawing valid conclusions from data]. <h0br> ## Why do we need it? <hbr> -- .pull-left[<h0br> We need it in everything: - Decision making - Controlling risks - Maximizing profits - Comprehension... ] .pull-right[<h0br> <img src="https://d1jnx9ba8s6j9r.cloudfront.net/blog/wp-content/uploads/2019/02/Statistics-Applications-Math-And-Statistics-For-Data-Science-Edureka-768x484.png" align="middle" height="230"> ] --- template: inter-slide class: left, middle count: false ##
.bold-blue[Outline] <br> .hhead[I. Frequentist statistics] <br> .hhead[II. Bayesian statistics] <br> .hhead[III. Summary and do & don't] --- template: inter-slide class: left, middle count: false ##
.bold-blue[Outline] <br> .section[I. Frequentist statistics] <br> .hhead[II. Bayesian statistics] <br> .hhead[III. Summary and do & don't] --- ## Consider an example <hbr> πͺ Coin toss: `\(n=100\)` and observed `\(55\)` Heads & `\(45\)` Tails. -- - What's your estimate of `\(p=\mathbb{P}(\text{Head})\)`? <hbr> -- > π There exists a .stress[fixed] parameter `\(p\)` generating the observations. -- - Statistical model: triple `\((S, {\cal A}, {\cal P})\)` - `\(S=\{0,1\}\)`: sample space - `\({\cal A}=\{S,\{1\},\{0\},\emptyset\}\)`: a `\(\sigma\)`-algebra or tribu on `\(S\)` - `\({\cal P}=\{\mathbb{P}_{p}:p\in [0,1]\}\)`: a family of probability distributions ([Bernoulli](https://en.wikipedia.org/wiki/Bernoulli_distribution)): <hbr> `$$\forall p \in [0,1]: \mathbb{P}_p(X=1)=p=1-\mathbb{P}_p(X=0).$$` > β οΈ This model is .stress[identifiable] i.e., `\(p\to\mathbb{P}_p\)` is one-to-one. -- <hbr> - Observations: `\(X_1,...,X_n\overset{\text{iid}}{\sim} \mathbb{P}_p\)` for some .stress[fixed] `\(p\in [0,1]\)`. -- .center[ π€ How to find such a `\(p\)`? ] --- ## Consider an example <hbr> .center[ <hbr> π€ How to find such a `\(p\)`? ] <hbr> -- ### Likelihood method <hbr> - Likelihood: probab of observing `\(H=55\)` and `\(T=45\)`. -- .left[ <hbr> `$$\begin{align}L(p)&=\mathbb{P}_p(X_1=x_1,...,X_{100}=x_{100})\\ &\overset{\text{iid}}{=}\prod_{i=1}^n\mathbb{P}_p(X_i=x_i)\\ &=\left(\prod_{i:x_i=1}\mathbb{P}_p(X_i=x_i)\right)\left(\prod_{i:x_i=0}\mathbb{P}_p(X_i=x_i)\right)\\ &=p^{55}(1-p)^{45}\end{align}$$` ] --- count: false ## Consider an example <hbr> .center[ <hbr> π€ How to find such a `\(p\)`? ] <hbr> ### Likelihood method <hbr> - Likelihood: probab of observing `\(H=55\)` and `\(T=45\)`. .left[ <hbr> `$$\begin{align}L(p)&=p^{55}(1-p)^{45}\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\ \ \ { }\end{align}$$` ] <iframe src='https://www.desmos.com/calculator/znfrid5qjg' style="width:100%; height:230px;" border: 0px none; title="Likelihood"> </iframe> --- name: CI count: false ## Consider an example <hbr> .center[ <hbr> π€ How to find such a `\(p\)`? ] <hbr> ### Likelihood method <hbr> - Likelihood: probab of observing `\(H=55\)` and `\(T=45\)`. .left[ <hbr> `$$\begin{align}L(p)&=p^{55}(1-p)^{45}\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\ \ \ { }\end{align}$$` ] -- - `\(L(p)\)` maximized at `\(\hat{p}=0.55\)` (.stress[Maximum Likelihood Estimator]). <hbr> -- <hbr> ### How confident are we? <hbr> -- - Confidence interval at `\(\beta=0.95\)` i.e., `\(\mathbb{P}[p\in\text{CI}(\beta)]=\beta\)`. Estimated by: <hbr> -- `$$\text{CI}(0.95)=\left[\hat{p}-\sqrt{\hat{p}(1-\hat{p})}\frac{1.96}{\sqrt{n}},\hat{p}+\sqrt{\hat{p}(1-\hat{p})}\frac{1.96}{\sqrt{n}}\right]=[0.452, 0.648]$$` <hbr> ---- <h1br> .small[π [See Appendix 1](#appendix1)] --- ## Consider an example <hbr> .pull-left-60[ <hbr> - Likelihood estimator: `\(\hat{p}=0.55\)`. - Estimated `\(\text{CI}(0.95)=[0.452,0.648]\)`. ] -- .pull-left[ <br-2> ### What does it mean? <hbr> ] .center[<img src="img/CI.gif" align="middle" width="60%">] --- ## Consider an example <hbr> .pull-left-60[ <hbr> - Likelihood estimator: `\(\hat{p}=0.55\)`. - Estimated `\(\text{CI}(0.95)=[0.452,0.648]\)`. ] .pull-right-40[ <h1br> π‘ .stress[If we repeat this ] `\(100\)` .stress[times, we expect that] `\(p\in\text{CI}(0.95)\)` .stress[around 95 times]. ] .pull-left[ <br-2> ### What does it mean? <hbr> ] .center[<img src="img/CI.gif" align="middle" width="60%">] --- count: false ## Consider an example <hbr> .pull-left-60[ <hbr> - Likelihood estimator: `\(\hat{p}=0.55\)`. - Estimated `\(\text{CI}(0.95)=[0.452,0.648]\)`. ] .pull-right-40[ <h1br> π‘ .stress[If we repeat this ] `\(100\)` .stress[times, we expect that] `\(p\in\text{CI}(0.95)\)` .stress[around 95 times]. β οΈ The only source of uncertainty is the data `\(X_j\)`s. ] .pull-left[ <br-2> ### What does it mean? <hbr> ] .center[<img src="img/CI.gif" align="middle" width="60%">] --- # Frequentist approach <hbr> - Key points: > - There's a fixed parameter producing the observations. - Suitable for long run with large observations. - The uncertainty is due to sampling error only. -- - Statistical model: `\((S, {\cal A}, {\cal P})\)` - Estimation: maximum likelihood, moment method,... - Uncertainty analysis: test or confidence interval,... - Interpretation, inference, conclusion. <hbr> -- ### Challenge <hbr> > Propose a statistical model for **hatched eggs** π£, **birth weights** πΆ and **dice game** π². ??? - `\(S=(0,\infty)\)` - `\({\cal A}={\cal B}(S)\)` - `\(\mathbb{P}_{\theta}={\cal N}\)` or `\(\chi^2\)`. ### π€ What a statistician might do - Probablity: `\(X_1,...,X_N\sim^{iid}{\cal B}(p), i.e., X_i=\begin{cases}1&\text{Head, prob }p\\ 0&\text{Tail, prob }1-p\end{cases}\)` - Likelihood: `\(L(p|x_1,...,x_N):=\mathbb{P}_p(\text{Observing such observations})\)` - Find `\(\hat{p}:=\arg\max_{p}L(p|X_1,...,X_N)\)` - Confidence interval `\(\text{CI}(\alpha)\)`: `\(\mathbb{P}(p\in \text{CI}(\alpha))\geq \alpha\)`. .left[ <img src="img/job1.png" align="middle" height="150"> <img src="img/job3.png" align="middle" height="150"> ] <br> .left[ <img src="img/job2.png" align="middle" height="235"> <img src="img/job4.png" align="middle" height="235"> ] --- template: inter-slide class: left, middle count: false ##
.bold-blue[Outline] <br> .hhead[I. Frequentist statistics] <br> .section[II. Bayesian statistics] <br> .hhead[III. Summary and do & don't] --- ## Recall Bayes's formula <hbr> - If `\(O,H\)` are two events with `\(\mathbb{P}(O)\times\mathbb{P}(H)>0\)`, then .pull-left[ <hbr> `$$\underbrace{\mathbb{P}(H|O)}_{\text{Posterior}}=\frac{\overbrace{\mathbb{P}(O|H)}^{\text{Likelihood}}\times\overbrace{\mathbb{P}(H)}^{\text{Prior}}}{\underbrace{\mathbb{P}(O)}_{\text{Marginal}}}$$` .right[[<img src="https://upload.wikimedia.org/wikipedia/commons/e/e1/ThomasBayes.png" align="middle" height="250">](https://upload.wikimedia.org/wikipedia/commons/e/e1/ThomasBayes.png)] ] .pull-right[ .center[<img src="img/Bayes.jpg" align="middle" height="150">] <hbr> .left[[<img src="img/Bayes_info.png" align="middle" height="250">](https://en.wikipedia.org/wiki/Thomas_Bayes)] ] --- count: false ## Recall Bayes's formula <hbr> - If `\(O,H\)` are two events with `\(\mathbb{P}(O)\times\mathbb{P}(H)>0\)`, then .pull-left[ <hbr> `$$\underbrace{\mathbb{P}(H|O)}_{\text{Posterior}}=\frac{\overbrace{\mathbb{P}(O|H)}^{\text{Likelihood}}\times\overbrace{\mathbb{P}(H)}^{\text{Prior}}}{\underbrace{\mathbb{P}(O)}_{\text{Marginal}}}$$` ] .pull-right[ .center[<img src="img/Bayes.jpg" align="middle" height="150"> ] ] <h1br> ### Explanation: <hbr> > - Prior `\(\mathbb{P}(H)\)`: our initial belief in the .stress[Hypothesis] `\(H\)`. <vbr> - Likelihood `\(\mathbb{P}(O|H)\)`: given `\(H\)`, how likely for `\(O\)` to occur. <vbr> - Marginal `\(\mathbb{P}(O)\)`: chance that `\(O\)` would occur in general. <vbr> - Posterior `\(\mathbb{P}(H|O)\)`: `\(O\)` is observed, how likely for `\(H\)` to occur? --- ## Covid test accuracy during the peak <hbr> .pull-left[ - During the .stress[peak]: Cov+ `\(=2\%\)`, how one is likely to be infected if he/she was tested positive?] .pull-right[ .center[<img src="img/covid1.jpg" align="middle" height="140" width="300">] ] -- <br> `$$\begin{align}\mathbb{P}(\text{Cov}+|T+)&=\frac{\mathbb{P}(T+|\text{Cov}+)\mathbb{P}(\text{Cov}+)}{\mathbb{P}(T+)}\\ &=\frac{\mathbb{P}(T+|\text{Cov}+)\mathbb{P}(\text{Cov}+)}{\mathbb{P}(T+|\text{Cov}+)\mathbb{P}(\text{Cov}+)+\mathbb{P}(T+|\text{Cov}-)\mathbb{P}(\text{Cov}-)}\\ &=\frac{1\times 0.02}{1\times 0.02+0.001\times 0.98}\approx 0.95\end{align}$$` -- - I think we can trust the test! <img src="https://t4.ftcdn.net/jpg/03/23/71/91/240_F_323719173_LcDYRnQuNiaBrRmRzsY4u6JWjj4IjvRv.jpg" align="middle" width="40"> --- ## Covid test during quarantine <hbr> .pull-left[ - During .stress[quarantine]: Cov+ `\(=0.2\%\)`, how one is likely to be infected if he/she was tested positive?] .pull-right[ .center[<img src="img/covid2.jpg" align="middle" height="140" width="300">] ] <br> `$$\begin{align}\mathbb{P}(\text{Cov}+|T+)&=\frac{1\times 0.002}{1\times 0.002+0.001\times 0.998}\approx 0.667\end{align}$$` -- - .stress[False positive] `\(\mathbb{P}(\text{Cov}-|T+)\approx 1/3\)`. -- - Wait! What! π€ I'm having another test! -- > π‘ Prior information can alter the posterior! --- ## Bayesian statistics <hbr> > π Parameter is a .stress[random variable] conditioned on observations. <hbr> -- ### Back to coin toss example <h0br> .pull-left[ - Same model except for `\({\cal P}\)`: - Prior belief: `\(p\sim \mathbb{P}_{\text{pr}}\)` on `\([0,1]\)` - `\(O=\{\)`H `\(=55\)` & T `\(=45\}\)`, ] --- count: false ## Bayesian statistics <hbr> > π Parameter is a .stress[random variable] conditioned on observations. <hbr> ### Back to coin toss example <h0br> .pull-left[ - Same model except for `\({\cal P}\)`: - Prior belief: `\(p\sim \mathbb{P}_{\text{pr}}\)` on `\([0,1]\)` - `\(O=\{\)`H `\(=55\)` & T `\(=45\}\)`, `$$\begin{align}\overbrace{\mathbb{P}(p|O)}^{\text{Posterior}}&=\frac{\overbrace{\mathbb{P}(O|p)}^{\text{Likelihood}}\times\overbrace{\mathbb{P}_{\text{pr}}(p)}^{\text{Prior}}}{\mathbb{P}(O)}\\ &\propto \mathbb{P}(O|p)\times \mathbb{P}_{\text{pr}}(p)\\ &\propto p^{55}(1-p)^{45}\times\mathbb{P}_{\text{pr}}(p)\end{align}$$` ] --- count: false ## Bayesian statistics <hbr> > π Parameter is a .stress[random variable] conditioned on observations. <hbr> ### Back to coin toss example <h0br> .pull-left[ - Same model except for `\({\cal P}\)`: - Prior belief: `\(p\sim \mathbb{P}_{\text{pr}}\)` on `\([0,1]\)` - `\(O=\{\)`H `\(=55\)` & T `\(=45\}\)`, `$$\begin{align}\overbrace{\mathbb{P}(p|O)}^{\text{Posterior}}&=\frac{\overbrace{\mathbb{P}(O|p)}^{\text{Likelihood}}\times\overbrace{\mathbb{P}_{\text{pr}}(p)}^{\text{Prior}}}{\mathbb{P}(O)}\\ &\propto \mathbb{P}(O|p)\times \mathbb{P}_{\text{pr}}(p)\\ &\propto p^{55}(1-p)^{45}\times\mathbb{P}_{\text{pr}}(p)\end{align}$$` ] .pull-right[ <br-4>
] ??? https://mspeekenbrink.github.io/sdam-book/ch-Bayes-factors.html --- ## Bayesian statistics <hbr> > π Parameter is a .stress[random variable] conditioned on observations. <hbr> ### Credible interval <h0br> .pull-left[ - For `\(\beta\in[0,1]\)`, `\(\text{CrI}(\beta)\)` satisfies: `$$\underbrace{\mathbb{P}(p\in\text{CrI}(\beta)|O)}_{\text{Posterior}}=\beta.$$` ] .pull-right[ <br-4>
] --- count: false ## Bayesian statistics <hbr> > π Parameter is a .stress[random variable] conditioned on observations. <hbr> ### Credible interval <h0br> .pull-left[ - For `\(\beta\in[0,1]\)`, `\(\text{CrI}(\beta)\)` satisfies: `$$\underbrace{\mathbb{P}(p\in\text{CrI}(\beta)|O)}_{\text{Posterior}}=\beta.$$` - Many ways to define it: - Below is as likely as above. - Minimal size. - Centered at the mean. ] .pull-right[ <br-4>
] --- count: false ## Bayesian statistics <hbr> > π Parameter is a .stress[random variable] conditioned on observations. <hbr> ### Credible interval <h0br> .pull-left[ - For `\(\beta\in[0,1]\)`, `\(\text{CrI}(\beta)\)` satisfies: `$$\underbrace{\mathbb{P}(p\in\text{CrI}(\beta)|O)}_{\text{Posterior}}=\beta.$$` - Many ways to define it: - Below is as likely as above. - Minimal size. - Centered at the mean. > .purple[Uniform]: [0.453,0.641]. ] .pull-right[ <br-4>
] --- count: false ## Bayesian statistics <hbr> > π Parameter is a .stress[random variable] conditioned on observations. <hbr> ### Credible interval <h0br> .pull-left[ - For `\(\beta\in[0,1]\)`, `\(\text{CrI}(\beta)\)` satisfies: `$$\underbrace{\mathbb{P}(p\in\text{CrI}(\beta)|O)}_{\text{Posterior}}=\beta.$$` - Many ways to define it: - Below is as likely as above. - Minimal size. - Centered at the mean. > .purple[Uniform]: [0.453,0.641]. ] .pull-right[ <br-4>
<h0br> π‘ For the chosen prior and the observed observations, there is a `\(\beta %\)` chance that `\(p\in\text{CrI}(\beta)\)`. ] --- ## Bayesian statistics <hbr> > π Parameter is a .stress[random variable] conditioned on observations. <hbr> ### Credible interval <h0br> .pull-left[ - For `\(\beta\in[0,1]\)`, `\(\text{CrI}(\beta)\)` satisfies: `$$\underbrace{\mathbb{P}(p\in\text{CrI}(\beta)|O)}_{\text{Posterior}}=\beta.$$` - Many ways to define it: - Below is as likely as above. - Minimal size. - Centered at the mean. > .purple[Uniform]: [0.453,0.641]. ] .pull-right[ <br-4>
<h0br> π‘ For the chosen prior and the observed observations, there is a `\(\beta %\)` chance that `\(p\in\text{CrI}(\beta)\)`. β οΈ This uncertainty is due to the parameter itself. ] --- # Bayesian statistics <hbr> - Key points: > - The parameter is modeled by a random variable. - It is updated through iteration of data collection process. - The uncertainty is due to the parameter itself. -- - Statistical model: `\((S, {\cal A}, {\cal P})\)` with prior `\(\theta\sim\mathbb{P}_{\text{pr}}.\)` - Computation of .stress[posterior] through Bayes's rule. - Uncertainty analysis: tests or credible intervals,... - Interpretation, inference, conclusion (according to prior/observations). -- ### Important remark > β οΈ Influence of the prior must be taken into account! ??? - `\(S=(0,\infty)\)` - `\({\cal A}={\cal B}(S)\)` - `\(\mathbb{P}_{\theta}={\cal N}\)` or `\(\chi^2\)`. --- template: inter-slide class: left, middle count: false ##
.bold-blue[Outline] <br> .hhead[I. Frequentist statistics] <br> .hhead[II. Bayesian statistics] <br> .section[III. Summary and do & don't] --- .pull-left[ <hbr> ## Frequentist <hbr> - Parameter is fixed, data is random: i.e., `\(\mathbb{P}(x_1,..,x_n|p)\)`. - Uncertainty analysis (**test/CI**) is interpreted w.r.t data. - For long run and large observations. ] .pull-right[ ## Bayesian <hbr> - Data is fixed, parameter is random: i.e., `\(\mathbb{P}(p|x_1,..,x_n)\)`. - Uncertainty analysis (**test/CrI**) is interpreted w.r.t parameter conditioned on data & prior. - Works well for small observations and can be iteratively updated. ] -- .pull-left-60[<br-4> ## Do and don't <hbr> - .stress[Do] define data and questions then choose the appropriate statistical analysis. - .stress[Do] investigate the influence of the prior when performing Bayesian analysis. - .stress[Do] ensure that your interpretation is correct with `\(p\)`-value, **CI** or **CrI**. ] .pull-right-40[ <br> - .stress[Don't] assume that Bayesian approach will solve insufficient or poor-quality data! Data quality is the most important thing in every analysis. ] --- ## π Further reading .center[ [<img src="img/ref1.png" align="middle" height="270">](https://www.redjournal.org/action/showPdf?pii=S0360-3016%2821%2903256-9) [<img src="img/ref3.jpg" align="middle" height="270">](http://ndl.ethernet.edu.et/bitstream/123456789/39363/1/Leonhard%20Held.pdf) ] ## π₯οΈ Slides - Link: [https://hassothea.github.io/files/teaching/MAC/Freq_Bayes_Stat.html](https://hassothea.github.io/files/teaching/MAC/Freq_Bayes_Stat.html) --- background-image: url('img/MAC_no_bg.jpg') background-size: cover <h1br> ### π£ Coming up next... <h0br> .pull-left[ <h0br> .center[<img src="img/Samnang.jpg" align="middle" height="100">] **Samnang**: .stress[Multidimensional Gaussian Distribution] (**June, 2024**) .center[<img src="img/samai.jpg" align="middle" height="100">] **Samai**: .stress[Linear & Quadratic Discriminant Analysis] (**July, 2024**) ] .pull-right[ <h0br> .center[<img src="img/sreyneath.jpg" align="middle" height="100">] **Sreyneath**: .stress[Logistic Regression] (**August, 2024**) .center[<img src="img/kimsia.jpg" align="middle" height="100">] **Kimsia**: .stress[Gradient Descent Algorithm] (**September, 2024**) ] <h0br> .center[<img src="img/Samreth.jpg" align="middle" height="100"> **Samreth**: .stress[Principal Component Analysis] (**October, 2024**) ] --- count: false name: appendix1 ### Appendix 1 : Confidence Interval Approximation <hbr> - .stress[Objective]: if `\(X_1,...,X_n\overset{\text{iid}}{\sim} \mathcal{B}(p)\)` and for `\(n\)` large enough, one has `$$\mathbb{P}\left(p\in\text{CI}\right)\underset{n\to\infty}{\to}1-\alpha$$` with `$$\text{CI}=\left[\sin^2\left(\arcsin(\sqrt{\hat{p}})-\frac{q_{1-\alpha/2}}{2\sqrt{n}}\right),\sin^2\left(\arcsin(\sqrt{\hat{p}})+\frac{q_{1-\alpha/2}}{2\sqrt{n}}\right)\right]$$` -- - .stress[In practice], we can approximate using Student's distribution: `$$\frac{\sqrt{n}(\hat{p}-p)}{\hat{\sigma}}\underset{n\to\infty}{\leadsto} T(n-1)\underset{n\text{ large}}{\approx} {\cal N}(0,1)$$` where `\(\hat{\sigma}^2=\frac{1}{n-1}\sum_{i=1}(X_i-\overline{X}_n)^2=\hat{p}(1-\hat{p})\)`. This leads to our result! --- count: false ### Appendix 1 : Confidence Interval Approximation <hbr> - .stress[Sketch of proof of objective]: - **Central Limit Theorem**: `\(X_1,...X_n\)` are iid r.v with `\(\mathbb{E}(X_1)=\mu, \mathbb{V}(X_1)=\sigma^2<\infty\)` then `$$\frac{\sqrt{n}(\overline{X}_n-\mu)}{\sigma}\overset{\cal L}{\to}{\cal N}(0,1).$$` - `\(\Delta\)`**-method**: if `\(R_n\overset{\cal L}{\to} \theta_0\)` with rate `\(1/r_n\)` such that `$$r_n(R_n-\theta_0)\overset{\cal L}{\to}{\cal N}(0,g(\theta_0))$$` and if `\(f:\Theta\to \mathbb{R}\)` is differentiable at `\(\theta_0\in\Theta\)` then `$$r_n\left(f(R_n)-f(\theta_0)\right)\overset{\cal L}{\to}{\cal N}(0,g(\theta_0).[f'(\theta_0)]^2).$$` - The proof is completed using `\(f(x)=2\arcsin(\sqrt{x}).\)` .small[π [Back to main slide](#CI)] --- exclude:true count: false template: inter-slide class: left, middle count: false .center[# References]<hbr> 📚 [Linder, T. (2002). Learning-Theoretic Methods in Vector Quantization. In: GyΓΆrfi, L. (eds) Principles of Nonparametric Learning. International Centre for Mechanical Sciences, vol 434. Springer, Vienna.](https://link.springer.com/chapter/10.1007/978-3-7091-2568-7_4) 📚 [Banerjee, S. Merugu, I.S. Dhillon, and J. Ghosh. Clustering with Bregman divergences. Journal of Machine Learning Research, 6:1705β1749, 2005](chrome-extension://oemmndcbldboiebfnladdacbdfmadadm/https://jmlr.org/papers/volume6/banerjee05b/banerjee05b.pdf) 📚 [A. Fischer. Quantization and clustering with Bregman divergences. Journal of Multivariate Analysis, 101(10):2207β2221, 2010.](chrome-extension://oemmndcbldboiebfnladdacbdfmadadm/https://www.lpsm.paris/_media/users/fischer/quantization_and_clustering_with_bregman_divergences.pdf) 📚 [S. Has, A. Fischer, and M. Mougeot. Kfc: A clusterwise supervised learning procedure based on the aggregation of distances. Journal of Statistical Computation and Simulation, 0(0):1β21, 2021. doi: 10.1080/00949655.2021. 1891539.](https://www.tandfonline.com/doi/abs/10.1080/00949655.2021.1891539)
[https://github.com/hassothea/KFC-Procedure](https://github.com/hassothea/KFC-Procedure) <h0br> .pull-right[ # Thank you π€ ]