Forum Replies Created
-
Posts
-
-
-
-
Hi Tanyawat,
I’m sorry it took a while for me to circle back to answering this question in written form.
1) An answer for this question is highly data-dependent. But the rules of thumb for data training is the more the better. And then you set that aside for training 70-80% of all the data you have.
If there are clear patterns hidden in your data, then not much data is needed for training (as clearly seen in our sample dataset from the lectures). If you find that even the large amount of data could not produce satisfying decision trees, perhaps the right attributes for prediction are not present in your data, or perhaps they were there but there were confounding effects at play, in which case a preprocessing of data might lead to better results.
This idea applies to the appropriate number of attributes. If you have the “right” attributes, then you will not need many of them to predict the data. But, of course, real-life data usually are not perfect. Another thing to consider is the more number of attributes are being fed into the algorithms, the more time it takes and the more complex the resulting decision trees will be.
2) The decision tree algorithm that we use in class for r (rpart) should be able to handle this. For other tools / programming languages / software, you will need to read the documentation of that specific tool/language/software to see if they have this ability embedded.
Regards,
Pimwadee -
Hi Abdillah,
You are right to be concerned about outliers. Outliers are ubiquitous in all kinds of datasets! When we are trying to make sense of something, we probably don’t want to think too much about exceptional cases like outliers. Because outliers can distort our understanding of the nature of the data.
There is an exception if we want to study those extreme cases, then outliers are important. But for clustering, algorithms that are prone to outliers like k-means or hierarchical clustering need to be treated with care. If we are aware that data has outliers, some may choose to play with inclusion / exclusion of those cases to see if clustering results change.
I hope that makes sense. Thanks for a spark of discussion!
Pimwadee
-
-
The dataset I believe has been processed from its original. The data with a bunch of numbers with no B or M diagnosis is in fact described here. http://archive.ics.uci.edu/ml/machine-learning-databases/breast-cancer-wisconsin/breast-cancer-wisconsin.names
An excerpt from the above file is below:
=========================
7. Attribute Information: (class attribute has been moved to last column)# Attribute Domain
— —————————————–
1. Sample code number id number
2. Clump Thickness 1 – 10
3. Uniformity of Cell Size 1 – 10
4. Uniformity of Cell Shape 1 – 10
5. Marginal Adhesion 1 – 10
6. Single Epithelial Cell Size 1 – 10
7. Bare Nuclei 1 – 10
8. Bland Chromatin 1 – 10
9. Normal Nucleoli 1 – 10
10. Mitoses 1 – 10
11. Class: (2 for benign, 4 for malignant) -
Referring to the file “breast-cancer-wisconsin.names” within this data folder (http://archive.ics.uci.edu/ml/machine-learning-databases/breast-cancer-wisconsin/), the class attribute is in the last column.
-
-
Thanks for the question! Yes, the reading assignment 2 of section 1.5 mentions both single linkage and complete linkage, which are two different ways for an agglomerative clustering to merge smaller clusters to bigger clusters. Even though we did not mean to elaborate on their definitions and difference in this course, we are happy to see its discussion on this forum.
To recap, single linkage clustering will merge based on the smallest distance of members within clusters, while complete linkage clustering will do so based on the largest distance of the members within clusters. That yields very different results as you have noticed.
By its definition, the single linkage will give you the merge which spreads locally. It will merge anything (points or clusters) that comes as the next closest objects together. I’m thinking it’s similar to water droplets being merged on a surface. Once smaller droplets form bigger water droplets, they expand to the next closest water droplet to merge. Edge-to-edge distances between clusters are a deciding factor for merging here.
In contrast, the complete linkage looks for a minimum farthest distance between clusters. This makes it not sensitive to outliers, i.e. it will not merge by only looking locally for the closest neighbor, but it will look more globally for who is the most similar neighbor. As a result, the complete linkage tends to give us more compact clusters. That is not to say that the single linkage is without its merit. The single linkage can produce clusters that are more inclusive of special cases.
And you can decide the type of clustering you want for your analysis depending on your goal of analysis as well as the nature of data.
-
-
Hi Navinee,
That is a good question. The short answer is there is no agreeable number in the industry as to which between_SS / total_SS is considered high or low. But let me take you through some examples, which I found on this link. It is helpful to consider between_SS / total_SS altogether with some plots of the data.
Now please consider case 1. This clustering yields between_SS / total_SS to be 47.5%. And the plot is displayed below.
Whereas in case 2, you get between_SS / total_SS to be 71.2 %. Below is the plot of the second case.
In the first case, 47.5% shows a good enough result, the clusters are somewhat separated. But I am not sure if we take the colors away from the plot we can still clearly see the clusters. Whereas in the second case, 71.2% shows more well-separated clusters (there’s more white space between clusters).
So in my opinion, if the between_SS / total_SS is over 70% I usually considered this very good, which means if I were to plot the data I will see the clusters nicely split. If my between_SS / total_SS is between 50-70%, I still consider this quite OK. I might want to verify the cluster qualities with other numbers and maybe with plots of data. But sometimes that’s as much as you can get out of your data. If my between_SS / total_SS is below 30%, then I think the result is not very good.
I hope this helps. If you have more questions, please let us know.
-
Dear Rawinan,
Thanks for your question. Accuracy is the percentage of correct prediction, while error rate is the percentage of the incorrect prediction. Those two numbers are calculated across all classes. Precision tells you how correct the algorithm was in predicting the positive class. Finally, recall tells you how much of the actual positive cases that the algorithm was able to predict as positive. I also did answer in Thai during the session so you may check out the video recording of the class meeting.
Pimwadee
-