Logistic Regression - THE MATH YOU SHOULD KNOW!

Поділитися
Вставка
  • Опубліковано 7 лис 2024

КОМЕНТАРІ • 127

  • @UnPuntoyComa
    @UnPuntoyComa 3 роки тому +11

    This is the most clearly explained and well developed video about the issue I have seen. Most explanations stop with the maximization of the log likelihood function, and I couldn't find how it is maximized until now. I didn't understand a bit, but it's better to know that something is beyond my comprehension than not knowing what it is. Thank you! Subscribed.

  • @oksaubercool
    @oksaubercool 5 років тому +93

    Very good explanation. Only thing, you're starting really slow, which is perfect, but then when the math gets messy you speed up by 10 times and go by without further explanations. Nonetheless very useful.

    • @calluma8472
      @calluma8472 5 років тому +36

      Yeah anyone who can follow the maths at that speed doesn't need this video I think.

    • @andyd568
      @andyd568 5 років тому +4

      Just pause the video and look up the terms he mentions. I think the problem is if he stops to go into each subtopic the video would become a lecture unto itself.

    • @95Bloulou
      @95Bloulou 4 роки тому +4

      I disagree, I think the speed is really nice during the whole video because it is calculus details that you can study if you want by just pausing the video.

    • @SpecialBlanket
      @SpecialBlanket 4 роки тому +3

      I disagree. Toward the end he's just rearranging the equations.

    • @UnPuntoyComa
      @UnPuntoyComa 3 роки тому +1

      I also considered the speed was adequate. I simply couldn't follow after he mentioned "Taylor series" just because I don't have a clue of what that is, but I get that what he did, if I knew about numerical methods wouldn't be so complicated.

  • @Actanonverba01
    @Actanonverba01 5 років тому +43

    Love the math full proofs! That stuff is rarely shown even in classes. There is just not enough time to... Great stuff!

    • @SpecialBlanket
      @SpecialBlanket 4 роки тому +3

      Yup. I came here after staring at Bishop (the book) for 3 hours failing to get through some of the "trivial" skipped steps.

    • @mktsp2
      @mktsp2 6 місяців тому

      Yeah, most statistics lecturers are loosers

  • @chriskong7418
    @chriskong7418 Рік тому

    Love the maths part. Definitely my hero in ML.

  • @mohammedismail308
    @mohammedismail308 5 років тому +1

    Sometimes the good demonstration is nothing without such one example which is deploying the theory in practice.
    Thanks at all :)

  • @kevinshao9148
    @kevinshao9148 11 місяців тому +3

    THE BEST only 9 min to illustrate Logistic Regression! Really appreciate your brilliant work!

  • @haifasaud1001
    @haifasaud1001 Рік тому

    Yesssssss finally! No one ever gives any significance into the mathematical part

  • @jerrylu532
    @jerrylu532 5 років тому +6

    Oh man, you just saved my course project! Thanks for these great help that really explained how the math works!

  • @qingli1799
    @qingli1799 Рік тому

    Amazing! Parameter estimation in logistic regression has confused me for so long. I know MLE is used to estimate betas in logistic regression. However, the full math proofs really clarify the way! Really appreciate your video!

  • @haraldurkarlsson1147
    @haraldurkarlsson1147 7 місяців тому

    Excellent! Clear and logical explanation of all the steps involved.

  • @hikmatullahmohammadi27
    @hikmatullahmohammadi27 Рік тому

    Thank you for crystal clear explanations.

  • @bambinodeskaralhes
    @bambinodeskaralhes 5 років тому +2

    Thank you very much !!!! The only guy who could make me understand this subject !!!! You are great !!!!!!

  • @christophersolomon633
    @christophersolomon633 4 роки тому

    I find this a really nice video which strikes a good balance between general principles and details (which can be a very tricky thing to do). I had spent some time reading a textbook about the method and had a few uncertainties. This seemed just the ticket to clarify it all.

  • @dm3248
    @dm3248 3 роки тому

    After going through so many videos, finally understood. Thanks!!

  • @professorg000
    @professorg000 2 роки тому

    Excellent job --- congratulations. You sound about 15 years old!!! Even more impressive

  • @stephanschaefer155
    @stephanschaefer155 4 роки тому

    Thank very much. First time I understand how the coefficients sre calculsted. Great!

  • @hnkulkarni
    @hnkulkarni 2 роки тому

    Thank you for this explanation.

  • @Trakushun
    @Trakushun 5 років тому

    Clear explanation and in deep developed. Charming voice and very good structure. Thanks dude!

  • @Saravananmicrosoft
    @Saravananmicrosoft 4 роки тому +1

    Very good explanation, i did that step by step derivative with your material can you do video on maths involved for backward propagation

  • @binyillikcinar
    @binyillikcinar 5 місяців тому

    Decent but NewtonRaphson is not the only numerical method. It could be better to list few alternatives, especially Gradient-Descent to the top of the list. Since it involves single derivative the parameter update rule is much simpler.

  • @mridulavijendran3062
    @mridulavijendran3062 4 роки тому

    why do we maximize the product(in particular) of the probabilities? Is it to exploit the idea that the log of the products are sums and it could also help simplify the calculations of the sigmoid function?
    Edit: How do we know that P(1-P) is a diagonal matrix?

  • @CraftyChaos23
    @CraftyChaos23 3 місяці тому

    07:54 In gradiant of loss function equation there should be X instead of XT

  • @romanwang7562
    @romanwang7562 2 роки тому

    I was able to implement this with a minor difference: I used X.T instead of X for the middle term inside the weight update expression

  • @vaishanavshukla5199
    @vaishanavshukla5199 4 роки тому

    very simplified and good explaination

  • @darasingh8937
    @darasingh8937 3 роки тому

    Great Explanation! Thank you!

  • @vivian_who
    @vivian_who 3 роки тому

    Excellent video!!! I have been looking for something exactly like this... Thanks!

  • @PCD1387
    @PCD1387 3 місяці тому

    can you please tell which book do you follow ?

  • @maryamrastegar6368
    @maryamrastegar6368 5 років тому +1

    thank you. it was very helpful in my exam.

  • @Ltsoftware3139
    @Ltsoftware3139 3 роки тому

    From what I understand, by estimating the beta parameter, we only determine the slope of the sigmoid, but it is still centered on the x-axis. My data is only positive, so in my case, I need another parameter to shift the whole sigmoid to the left or right.

    • @9181shreyasbhatt
      @9181shreyasbhatt 11 місяців тому

      u mean by using e^-(beta0 + beta1 x) instead of e^-(beta1 x) in sigmoid function

  • @haojiang4882
    @haojiang4882 6 років тому +2

    Dude you killing it! Best explanation +1!

    • @haojiang4882
      @haojiang4882 6 років тому

      Subed!

    • @CodeEmporium
      @CodeEmporium  6 років тому

      Thanks a ton! Glad to have you on board! Made a similar video on Kernelisation (and the kernel trick) yesterday. Check it out!

    • @haojiang4882
      @haojiang4882 6 років тому

      @@CodeEmporium Absolutely!

  • @rocavincent2266
    @rocavincent2266 11 місяців тому

    At 6:25, I think there is a sign error for the calculus of beta_{t+1}. You make a subtraction as in gradient descent, whereas we want to maximize the likelihood here. Am I right ?

  • @user-xt9js1jt6m
    @user-xt9js1jt6m 4 роки тому +2

    Nice info.
    How do we obtain standard errors of estimators in logistics regression??
    Kindly guide me

  • @basantmounir
    @basantmounir 5 років тому +4

    Can someone please explain at 5:20 how did we convert the expression into a summation? And how was the log part of the new expression of summation? Also, what do the s and n in the counters represent and is there a relationship between them? Thank you.

    • @dominiccordeiro9257
      @dominiccordeiro9257 5 років тому +4

      you take the log of a product. Then convert it to sums of logs. for example: log (a*b) = log(a) + log(b)

  • @Actanonverba01
    @Actanonverba01 5 років тому +1

    Hey, do you have a book reference for the math you show here? Awesome work! :)

  • @name6297
    @name6297 4 роки тому

    The explanation was really good. Can you suggest a couple of math courses that help better visualize what I've seen here?.Thanks :-)

  • @sijiahuang6936
    @sijiahuang6936 Рік тому

    Hi, I just want to ask, in the last formula, should it be "X^T" instead of "X"? I mean the middle "X" in (X^T * W*X)^(-1) X (Y-Y').

  • @kkkk150984
    @kkkk150984 4 роки тому +4

    How the powers come yi and 1-yi at 5:19 in video please clarify..

    • @giacomopauletti5099
      @giacomopauletti5099 4 роки тому

      I am watching the video rn ... did you find the answer to your question? If you did, pls tell me the answer because I have been struggling from such a long period

    • @Ltsoftware3139
      @Ltsoftware3139 3 роки тому +2

      For the part p(xi)^yi -> if yi = 1 we will remain with p(xi), but if yi=0 that element will not have any impact.
      For the second part (1-xi)^(1-y1) it's exactly the opposite. If yi = 1 then we will have (1-xi)^0, so the term will not have any impact. If yi=0 then we will have (1-xi)^1=1-xi.
      Basically, raising to the power of yi filters(get's rid of) all the elements where yi=0, and raising to the power of (1-yi) filters all the elements where yi=1.

    • @dm3248
      @dm3248 3 роки тому

      @@Ltsoftware3139 thanks!!

  • @shivampradhan6101
    @shivampradhan6101 4 роки тому

    I watched the whole playlist but didn't understand much of the maths.what should I do

  • @henrychen1544
    @henrychen1544 4 роки тому +3

    Hi, I was wondering where the term yi came from and what do you mean by s in yi
    thank you

  • @Tyokok
    @Tyokok 5 років тому +2

    Great explain in such short 9 min. Subscribed! One question: in your video, you finally got formula beta(t+1) = beta(t) + ......., so how you set up beta(t=0) the initial value of beta to start your iteration? Thank you very much in advance!

    • @CodeEmporium
      @CodeEmporium  5 років тому

      Thanks for hopping board! You can randomly initialize your parameters ( the beta vector ).

    • @Tyokok
      @Tyokok 5 років тому

      @@CodeEmporium I see. Many thanks for the reply!

    • @Tyokok
      @Tyokok 11 місяців тому

      @@deepanshudashora5887 hope it's not too late if you ask me or the poster. linear regression for linear model predict a value, logistic regression for classification problem.

  • @JDMathematicsAndDataScience
    @JDMathematicsAndDataScience 9 місяців тому

    Great. I’ve never heard of this pronunciation of matrix.

  • @bevansmith3210
    @bevansmith3210 5 років тому +1

    one word: thankyou!

  • @areejnasser6664
    @areejnasser6664 6 років тому +3

    Great explanation

  • @rithealeang6217
    @rithealeang6217 3 роки тому

    Don’t quite understand when you said remove y as it is independent to beta and no gradient term with p(x)x. Any explanation thank

  • @juntong8488
    @juntong8488 5 років тому +1

    Thanks, very clear.

  • @1UniverseGames
    @1UniverseGames 3 роки тому

    What will happen if we use (-1,1) instead of (0,1) in logistic function, what kind of equations it will give? Any video or source to study this?

    • @विशालकुमार-छ7त
      @विशालकुमार-छ7त 3 роки тому

      1/(1+e^(-bx)) always lie between 0 and 1. There is no choice to use anything, the function is chosen in such a way that it always lie between 0 and 1.

  • @sahilchaturvedi593
    @sahilchaturvedi593 6 років тому +1

    Best explanation on youtube. Thanks :)

  • @yulinliu850
    @yulinliu850 6 років тому +1

    Excellent! Thanks a lot!

  • @mahammadodj
    @mahammadodj Рік тому

    Could anyone explain why P(1-P) is written as W at 8:12 ?

  • @vaishanavshukla5199
    @vaishanavshukla5199 4 роки тому

    great explaination !!!!!

  • @malepatirahul7339
    @malepatirahul7339 3 роки тому

    in loglikelihood function how was the seventh step calculated

  • @marx427
    @marx427 4 роки тому

    Omg i was looking for this ! ❤️

  • @shardx191
    @shardx191 4 роки тому +1

    i dont understand at 4:40 ,
    what does s in yi=1 means ? how does it relate to the P notation

    • @mahammadodj
      @mahammadodj Рік тому

      it means the independent variable is 1 in dataset.

  • @moisessoto5061
    @moisessoto5061 5 років тому

    Can we do a linear regresion of the logit to the explanatory variables and get the probabilities from the fitted logit?

    • @CodeEmporium
      @CodeEmporium  5 років тому

      Linear regression expects the outcome y to be continuous - not categorical

  • @mkaberli
    @mkaberli 5 років тому +2

    You could drop the music.

  • @akshatjain1699
    @akshatjain1699 4 роки тому +1

    hey, i am having difficulty in implementing the formula in python. the matrix inside the inverse bracket is singular matrix. how do I solve this

  • @hareedyhareedy2863
    @hareedyhareedy2863 5 років тому +1

    Can you please create another with examples

  • @cuysaurus
    @cuysaurus 5 років тому +1

    at 8:21 is it X^T (Y-Yhat^(t)) instead of X(Y-Yhat^(t))? in the very last line.

  • @Tyokok
    @Tyokok 6 місяців тому

    Dear Sir, if I may have 2 questions here: 1) 7:25, how did you remove y_i as it's independent? yi can be opposite signs, how can it be removed like 1? 2) at 7:58 in matrix representation why you convert p(x_i) in different way? or it really doesn't matter, cuz you will substitute beta_i in sigmoid function at each iteration?
    Many Thanks!

  • @utkarshsingh2675
    @utkarshsingh2675 Рік тому

    perfecto!

  • @aayusmaanjain9854
    @aayusmaanjain9854 Рік тому

    Can someone please explain why the yi and (1-yi) term goes to the exponent in the second equation when we combine the product at 5:20

    • @CodeEmporium
      @CodeEmporium  Рік тому

      We took the logarithm of the equation. A property of logarithms is the exponent term can be written as a product. And we took the logarithm in the first place since we want to just find the betas that maximize the value of L. The values of betas remain the same if we maximize the log of the equation too (a property called “monotonically increasing functions”)

  • @animeshsharma7332
    @animeshsharma7332 4 роки тому +3

    7:38 from where that goddamn transpose arrived

    • @CheetahDFurious20
      @CheetahDFurious20 4 роки тому +2

      Hi.. Actually In terms of matrix we cannot multily any matix by itself as such i.e. if you consider X is a matrix and if you want to calculate X * X then we cannot do it as such because order of matrix i.e. m x n wont allow us untill and unless it is sqaure matix else we have to transpose it and then multiply it.. X (m x n) * X ^T (n x m) = XX^T (m x m). Hope this final representation would help you... Thanks Happy learning...

    • @sorvex9
      @sorvex9 3 роки тому

      @@CheetahDFurious20 Thanks bro

  • @azaira010
    @azaira010 4 роки тому

    20 times speedy lecture for me..... i am a noob in deep learning

  • @musarratazim7940
    @musarratazim7940 Рік тому

    Why you don't estimate alpha? ?you only consider Beta in logistic regression model

    • @CodeEmporium
      @CodeEmporium  Рік тому

      Sorry. What is alpha in this case?

    • @musarratazim7940
      @musarratazim7940 Рік тому

      In logistic regression parameter alpha is also also present (book gujrati econometrica and walepole introduction to statistics) but in this case why you not take it.

  • @hayatt143
    @hayatt143 6 років тому

    Awesome Explanation. I was looking for an answer to this question. Please help.
    In Logistic Model ,a coefficient has value 1.6. This means that each unit change in the corresponding predictor variable multiples the odds of the outcome by how much?

    • @sominya
      @sominya 6 років тому

      e^1.6

    • @hayatt143
      @hayatt143 6 років тому

      can you give the formula to calculate this..? coz options are
      a) 2.75
      b)3.95
      c)4.75
      d)4.95

  • @shahnawazfingertips5367
    @shahnawazfingertips5367 6 років тому +3

    dude we can use gradient descent instead of newton raphson

  • @shubhijain2706
    @shubhijain2706 4 роки тому

    Please someone help me with this, I am lil confused whether Y hat at 7:54 and P at 7:58 are same?

  • @AlexSmith-tr9hc
    @AlexSmith-tr9hc 5 років тому

    "Numerically encoding classes looses meaning" - looses? Did you mean "loses" at 0:31?

  • @anverHisham
    @anverHisham 3 роки тому

    Thanks a lot :-)

  • @user-xt9js1jt6m
    @user-xt9js1jt6m 4 роки тому

    Wt if there are two parameter?
    Alpha and beta??

    • @ColinXYZ
      @ColinXYZ 3 роки тому

      You’d just do the same steps, but derive for your other variable instead .

  • @himeshph
    @himeshph 5 років тому

    Hidden gem

  • @oluwole635
    @oluwole635 4 роки тому

    Please I have a presentation on logistic regression and the part of the Hessian Matrix where we applyed the gradient, can someone please explain to me. I got all other thing including the matrices but only that. Please help ASAP.

  • @bin5156
    @bin5156 5 років тому +1

    It was sooo helpful!

  • @ltbd78
    @ltbd78 6 років тому

    Thanks

  • @rishhabhnaik2298
    @rishhabhnaik2298 5 років тому

    How did we remove yi at 7:27 ?

    • @ssshukla26
      @ssshukla26 4 роки тому

      yi is independent of beta.

  • @tasnimyusof7079
    @tasnimyusof7079 6 років тому +1

    Hi, could you share also if there are few variables.. How to get every coefficient for the variable.. Let say it have 5 variables. Thankssss 😁

  • @louerleseigneur4532
    @louerleseigneur4532 4 роки тому

    merci

  • @nebiyathawi7457
    @nebiyathawi7457 4 роки тому

    hello,its is nice man

  • @faisalsal1
    @faisalsal1 10 місяців тому +2

    The background music is distracting.

  • @kumaravelk1091
    @kumaravelk1091 6 років тому +1

    Content is good... but going too fast..

    • @CodeEmporium
      @CodeEmporium  6 років тому

      Kumaravel K thanks for the feedback. I'll pace myself better in future videos

  • @Felicidade101
    @Felicidade101 6 років тому

    if you are here I recommend you check out this video too, ua-cam.com/video/yIYKR4sgzI8/v-deo.html its from StatsQuest. Super good channel.

  • @Dave-nz5jf
    @Dave-nz5jf 5 років тому

    lol it's loses not looses.

  • @farooq8fox
    @farooq8fox 5 років тому

    I lost it at 5:50, Ill comeback when im smarter

    • @ninjawarrior_1602
      @ninjawarrior_1602 5 років тому +2

      Bro not a worry see until that moment he has just simplified the equation and now he just want to maximise the function
      Additionally
      Then he is using Taylor's series expansion and truncating that to two terms

  • @rohanreddymelachervu3498
    @rohanreddymelachervu3498 2 роки тому

    Why video in chinese??

  • @Areeva2407
    @Areeva2407 4 роки тому

    You are a Good Tutor but content is very Basic ..
    No Solved Examples ,,, Purpose not solved.
    Please also add Learning Outcomes at the beginning so that we can save our time.