To hopefully make it clearer to everyone, the iris labels were hidden from the clustering algorithms and then only used after the fact to see how well the clusters recovered the true labels. So the task was still unsupervised because the supervision was hidden. This is a standard technique for evaluating unsupervised machine learning algorithms.
If you use the PAM clustering algorithm with an outlier, as in your example, is it possible that PAM would assign the outlier to its own group? In other words, is it possible that the lowest error would be achieved by assigning the outlier to one group and everything else to another group?
Title of the video should be : clustering (with Kmeans and pam). I thought we would have some insights on graph theory (PMFG), on hierarchical (with complete and ward linkage), on fuzzy etc... Great video on Kmeans anyway !
Only at 11 min in, but erm confused. You're clustering, but you don't have any idea of what you're clustering on? You aren't clustering on a dimension? I.e., music genre What gives you a red/blue division in the first place - "it just looks like here's a cluster... And that, there, is a cluster. Now, let's make 'em fit..."? I'm not understanding.
Clustering is just defining groups based on similarity. That similarity could be based on one or on any number of attributes. It is unsupervised which means there are no labels in the data. Now, the algorithms label each point with the same label as the closest cluster center, thereby labeling each cluster. Then the cluster centers are adjusted to better reflect the cluster members, and then the labels are updated based on the new cluster centers (which can assign points to different clusters compared to the previous round). Repeat until convergence, which is when no points are assigned to a different cluster and when the cluster centers stay the same. This is when the "best fit" is achieved.
Check out the full Data Analysis Learning Playlist: th-cam.com/play/PLzH6n4zXuckpfMu_4Ff8E7Z1behQks5ba.html
13:58 "It's worked pretty well; it's not perfect."
I feel like that should be the slogan for this course and for data science in general.
To hopefully make it clearer to everyone, the iris labels were hidden from the clustering algorithms and then only used after the fact to see how well the clusters recovered the true labels. So the task was still unsupervised because the supervision was hidden. This is a standard technique for evaluating unsupervised machine learning algorithms.
Thanks - was wondering
I love videos like these. I believe that one day people won't say "I learned that on college", but "I went on TH-cam and it was all there" instead.
I love this guy! He is amazing at these things.... Thanks friend this is a lifetime charity. Everyone is going to learn from this at free of cost.
I always click on a new Mike Pound Computerphile video at first sight.
How meta. This WAS the video that was recommended for me to watch.
You're gonna be so happy after google recommends you the Saw movie! A whole movie dedicated to a SAW!
best series ever! :) thanks so much for this
Hi, thanks for these great videos.
At 13:35, I'm missing maximise_diag function, in which package is it?
I am having the same problem, did you get an answer?
I can't find any function for this too. ill come back if i find it.
Hello. I'm having same issue, did you get an answer?
Same question...No answers yet
Wood turning videos are fascinating. Can’t wait to see what the TH-cam algorithm recommends next for me.
Just submitted my last paper for my masters degree that HEAVILY relies on clustering algorithms. I really wish this video was released 2 years ago.
Also please consider making a video on big o complexity. I think it would go really good with this topic.
If you use the PAM clustering algorithm with an outlier, as in your example, is it possible that PAM would assign the outlier to its own group? In other words, is it possible that the lowest error would be achieved by assigning the outlier to one group and everything else to another group?
Title of the video should be : clustering (with Kmeans and pam). I thought we would have some insights on graph theory (PMFG), on hierarchical (with complete and ward linkage), on fuzzy etc... Great video on Kmeans anyway !
I watch
Computerphile videos the same way Mike watches woodturning videos.
Thanks so much for this, a great look at clustering
Is there a degree to which the dimensions are weighted when you cluster? Or would you apply the weighting to your data before clustering them?
I assume you would do this with the dimensions that come out of PCA, which implies that they have already been weighted.
If you ever wondered about the expression "high brow", here's your example.
Depending on which language(s) you're using, DBSCAN libraries can be worth a look
Very clear explanation
I purchased a wood turning lathe
Could I use kmeans for hyperspectral images ?
Brilliant, cheers!
This video was recommended to me after watching an MIT lecture on clustering lol
We all end up watching the wood turning videos, lol
If TH-cam can't subtitle properly a British accent, I have no hope in Artificial Intelligence.
The joy of wood? Dr Mike could become the Nick Offerman of the UK.
PSA: TH-cam uses a mix between association analysis and clustering analysis
:) thanks
Only at 11 min in, but erm confused. You're clustering, but you don't have any idea of what you're clustering on? You aren't clustering on a dimension? I.e., music genre
What gives you a red/blue division in the first place - "it just looks like here's a cluster... And that, there, is a cluster. Now, let's make 'em fit..."? I'm not understanding.
Clustering is just defining groups based on similarity. That similarity could be based on one or on any number of attributes. It is unsupervised which means there are no labels in the data. Now, the algorithms label each point with the same label as the closest cluster center, thereby labeling each cluster. Then the cluster centers are adjusted to better reflect the cluster members, and then the labels are updated based on the new cluster centers (which can assign points to different clusters compared to the previous round). Repeat until convergence, which is when no points are assigned to a different cluster and when the cluster centers stay the same. This is when the "best fit" is achieved.
K nearest neighbors...
PCA got almost equally wrong result 😂
I remember firefly.
Jared aka Donald Dunn spotted