Assignment 4 Clustering and Classifying Images F23

Guidelines for the Assignment:

In this assignment we will continue to work with materials and algorithms discussed in class. We will use Orange Data Mining to carry out this task and to write a small critical essay.

Part One:

The first exercise is a clustering exercise. You should choose a corpus of images (somewhere between 50-200 would be best). As we discussed in class, it is possible to choose a subset of images from a large dataset found at places like Kaggle, but the topic should be something that you know something about. Some prior knowledge of the subject is important since in the second part you will classify the images into subtypes. Choosing a dataset of images you are not very interested in, or a dataset that you do not have enough knowledge to categorize will make it difficult.

I suggest you use this part of the assignment to do some exploratory data analysis (EDA). Using the pre-set Orange Data Mining workflow I distributed (images2.ows), you can address the following questions:

• How does a built-in algorithm in ODM cluster the data?
• Do other algorithms give you different results?
• What features seem to be most characteristic of the different quadrants of the image plot?
• Using hierarchical clustering, isolate specific clusters to look more closely.

Make screenshots of the main image plot you generate and annotate it with labels (just typed on manually with a program like Preview) which illustrate in your opinion what combination of features are represented in each of the four quadrants of the plot. Caption the image with the relevant information (i.e. Image Plot generated with Orange Data Mining and the XXX algorithm). Also include image plot of specific clusters that can help you to understand how the out-of-the-box algorithms (think of the Monet/Manet video here)

Be sure to comment on how differently you see the images.

Part Two:

In the second exercise you will categorize the images into at least two different groupings. The categorization is carried out by placing them into folders (bearing a descriptive title). This means that for the purposes of your assignment, you should have two or three copies of the same images, but divided across different folders (e.g. size, shape, color OR origin, quantity, size). Your classification system should be sure to have a somewhat equal number of images in every category.

For each of the sets of images, please generate a confusion matrix and assess the ability of the algorithm to predict the categories you have chosen. Remember that you can click on the box where there has been misclassification and visualize the misclassified objects in the viewer.

Guiding questions for this part of the assignment:

• How well did any of the built-in algorithms you chose predict the categories you established?
• Did you go back and adjust any of your categories?
• When you isolate the cases of “false positives”, i.e. mis-predictions, can you understand why the algorithm got them wrong? (Think about the Monet/Manet mis-prediction case we saw in class).
• If you could train your own algorithm what would you aim to teach it?

In your write up, you should include insights and ideas from the new book Distant Viewing (Arnold & Tilton).

Make sure that you include images from your analysis.

1500 words maximum.