*My takeaways:* 1. DIstance 9:30 2. k-means algorithm 17:03 - How to choose k 23:57 - Unlucky initial centroids 25:56 - An example 28:58 - Scale data into the same range 37:11
The explanation level of this video is by far the best I have ever watched. Prof. Guttag does a very good job in explaining every concept more clearly.
it just struck me, after all these lecture videos, that professor Guttag is actually using a classic positive reinforcement technique to make the students more attentive and responsive in class by giving out candies for correct answer. lol! and i am not sure if its the result of this or something else but the students seem wayyy too eager to answer questions in this paticular lecture video!
Great lecture ! I attended the Clustering lecture by prof Ayan Seal today (even though I dont have the course : Introduction to Data Science) , he didn't focus a lot on code, but had similar things to share about clustering !
When we are clustering the airports, the professor only stopped to think about linkage when he arrived at Denver. Shouldn't we have thought about it since the beginning of the clustering? If so, we could have gotten (BOS, SF) instead of (BOS, NY) for the first iteration using complete linkage.
Since in the first iteration there are a number of clusters equal to the number of cities, wouldnt complete linkage be the same as single linkage, given there is only one point of measurement for each cluster? I didnt go back to check, but perhaps after the second iteration there wouldve been some different answers?
The guys who down voted this video must had nothing better to do. The lecture was nicely paced and I think he already made the problem as clear as it can get. Anyway, that was a great lecture. A big thank you to Professor Guttag and the MIT OpenCourseWare team.
Main issues of K-Means : choosing the number of clusters (k) and data scaling: But what if one wants to apply weights to the features (parameters)? Should you just multiply the features with the desired coefficients?
29:06 Isn’t the heart attack example a case of supervised learning, since we have the labels? 1:59 At the start of the lecture, the professor mentioned clustering as an example of unsupervised learning
What are some some methods to evaluate the quality of the clusters, if we do not have an outcome variable? In the example they were evaluated based in part based on whether the subjects in the cluster died at a higher rate. What do I do if I don't have an outcome to look at, only characteristics? For context, I'm creating cognitive style groups based on user data for an insurance company, and these styles will be later used for morphing, churn etc. but do not have an outcome variable per se.
Ok, by minute 7 my mind is wondering if there's going to be a bonus assignment to find the probability that Professor Guttag will correctly throw you the piece of candy on the first try. The odds of you catching it greatly increase the closer your sit to the front center of the room.
Thanks for an amazing lecture! @29:35 it tries to cluster data into two groups and see if it correctly differentiated people who dies of heart attack and those that didn't. To me this is using clustering for classification task, if yes, when would someone use clustering rather than classification?
To show new mean is zero: mean = (sum of vals)/n, vals is then transformed (by code) such that centered_normalized_vals = new_vals = (vals - mean)/sd, so new_mean = (sum of (new_vals - mean))/n = (sum of vals - n*mean) / (sd*n) = (n*mean - n*mean) / (sd*n) = 0. To show sd becomes 1: originally sd is calculated as root ((1/n)*sum over every val of ((val - mean)^2)), vals is transformed (by code) such that centered_normalized_vals = new_vals = (vals - mean)/sd, so new_sd = root ((1/n)*sum over every new val of ((new_val - new_mean)^2)), new_mean was previously shown to be zero, so this means new_sd = root ((1/n)*sum over every new val of (((val - mean)/sd)^2)) = root ((1/n)*(1/sd^2)*sum over every val of ((val - mean)^2)), since (by our initial definition) sum over every val of ((val - mean)^2) = n*sd^2, this tells us that new_sd = root ((1/n)*(1/sd^2)*n*(sd^2)) = new_sd = 1.
Pretty sure he is talking about the number of comparisons which need to occur to create the group, n-squared meaning the number of comparisons is on the order of the square of the number of objects to compare, and n-cubed on the order of the cube of the number of objects to compare. Sort of like big-O notation.
This is not always the case (depends on the code), but it might help to think of n-squared as 2 nested loops and n-cubed 3 nested loops. For instance, in a n-squared algorithm you have n items where, for each item, you make n comparisons. Imagine a really big n.
"Clustering" is usually taught to "signal" "alumni" that anyone "in their *network*" can't learn and be good at some skills because some Terrorists in their "*network*" may be affected andor effected.
As the video description states, the course name is "Introduction to Computational Thinking and Data Science" as it was taught in the Fall of 2016 by the Massachusetts Institute of Technology. For more information, see the course on MIT OpenCourseWare at: ocw.mit.edu/6-0002F16.
at 46:50 the professor mentions “has pretty good specificity, or positive predictive value, but its sensitivity is lousy.” can someone explain how specificity = ppv? im assuming: ppv = tp/(tp+fp) specificity = tn/(tn+fp) doesnt ppv = precision?
Hello, I come from wet lab and I am not familiar with machine learning. But I am really interested in this topic since I want to apply machine learning to my research in plant genetics. I have watched this video several times but still I have not gotten all the things the professor mentioned. I wonder if the author or anyone can share the lecuter or books in this topic. It will mean alot to me. Thank you in advance.
A late response but here we go. I would suggest you read the 100 pages machine learning book, it doesn't really really assume any background but of course having it makes things easier.
The textbook is: Guttag, John. Introduction to Computation and Programming Using Python: With Application to Understanding Data. 2nd ed. MIT Press, 2016. ISBN: 9780262529624. See the Readings section for more details: ocw.mit.edu/6-0002F16. Best wishes on your studies!
14:20 the distance from Denver to Seattle is 1307 and the distance from Denver to Boston is 1949, so why he clustered Denver to Seattle instead of Boston when using Complete linkage? should it not be clustered to the greatest distance?
No, for complete linkage you look up, as you said the greatest distance of each cluster to the datapoint and then cluster it with the smallest out of these distances.
Take care, students, with democrat teachers in computer science classes. They don't care to play with you and call you a dumb if you are a republican and later ask you to choose who is the dumb and who is the smart. I hope you grades doesn't be influencied by you political bias.
