Last week we learned about using macros in SAS. To be honest, in the beginning, I found it really difficult, but after reviewing it few times even after submitting my assignment, I now feel a bit more comfortable using it. But then again, classes are moving very fast; we’re jumping from one topic to another in a week. Learning a new concept every week is a bit overwhelming, especially when I am unable to keep up with the fast pace of the program. But I also find this exciting, It pushes me to challenge myself. It reminds me why I joined this program. I know it’s hard but I will keep at it.
This week we learnt about two powerful procedures of SAS – PROC UNIVARIATE and PROC FASTCLUS. So here are some things that are basic but I need to make a note of. PROC UNIVARIATE – If the PROC MEANS procedure does not produce the statistic needed for a data analysis then we can use PROC UNIVARIATE as it can do everything that PROC MEANS can do and much more. It provides a wider variety of statistics (than PROC MEANS) like moments, descriptive statistics, quantiles. In addition, it also generates graphs and will help us discover information about the distribution of data as well as identify extreme observations in our data.
PROC FASTCLUS – It performs K-means cluster and other clustering analysis techniques. But we are going to focus on K-means cluster analysis on the basis of distances computed from one or more quantitative variables. I am more excited to learn about PROC Fastclus, as it starts our journey into unsupervised machine learning. Basically it means that we will be able to let the computers identify data on its own and use that identification to cluster similar pieces of data together. It picks up initial random observations called cluster seed, and by default it uses the Euclidean distance (distance between two points). It assumes a center for one of the cluster and finds the shortest distance between each of the observations to the center of the cluster, and finds how well the cluster represents the datasets. PROC Fastclus is designed to find good clusters, however, they may not be the best possible clusters. Moreover, the Fastclus procedure is intended for large data sets, with 100 or more observations.
One of the examples that we used in class for PROC FASTCLUS was to analyze types of beers, where we clustered all the beer data into 3 clusters – high alcohol low bitterness, medium alcohol medium bitterness, and low alcohol high bitterness. We were able to narrow down 2,410 numbers of different beers into 3 categories that sort of captured general variety of beers types in our data set. So, we can now just look into the particular categories that we are interested and find the few selection of beers that we are interested than looking at 2,410 different beers.
Using PROC Fastclus we can build a perfect model. We need to remember that clustering is not an exact science. We can build a perfect cluster by simply increasing the value of K or number of clusters. So, technically, we can have a cluster for each point of observation. But in practical it does not make any sense to do so. There is a trade off how well the data fits the cluster and how well we can define the cluster. Another thing to note is that if we are clustering more than 3 or 4 variables, then we have to really ask does it make sense or can we easily describe the lows and highs of each different things. It is hard to describe when we have 4 to 5 dimensions, it becomes much harder to explain and defend it. We have to be cautioned of clustering more than 10 clusters because we will end up with unintelligible clusters. Therefore, we need to have a fine balance of how many clusters we want so that we can look at the data and have enough clusters to analyze it better.
Another scenario that we looked at is where we separated a data set into 2 sets, one for training set and another set for test. For training set, we would run a model. Once we had a good model, we ran the test set data using that model to predict. I found it very fascinating to learn unsupervised learning where we are not trying to predict a target but we are trying to magically find hidden relationship within the data without trying to force those relationship by having something to try to predict.
Key things for Unsupervised Learning:
1. We don’t have a label or target that we can go learn from.
2. Instead we just look at natural grouping within the data or natural patterns that we find within the data.
Lastly, there is so much to learn. We just touched the basics of machine learning, but it is still very difficult. I often find myself overwhelmed with so much information to process. For this week’s assignment for the last task, i did get stuck and struggled a bit. So, I scheduled a one-on-one video session with my professor to go over some of the conceptual things that i was not clear about. This helped me immensely clear our my confusion and I was able to finish major chunk of the assignment. I still need to clean my codes and write additional comments to finish my assignment for this week, and finishing my assignments, I plan to go back and review the course materials again to firm my understanding of the material better.