Fall Semester 2018
It’s great if you come up with your own ideas! However, here are some suggestions, in case you are looking for inspiration or would like to calibrate the scope of your project. While most of these are free-form projects, you can also do paper re-implementations or re-evaluations.
The main goal of these is to reimplement the ideas of the paper. For other current Vis publications, see the IEEE Conf. on Visualization. Consider to use the SFU Library Bookmarklet if you want to access non-free publications from home.
One of the difficulties with machine learning is to really understand how an algorithm/algorithm family works. The goal here would be to pick a particular algorithm/algorithm family and help the user to better understand it by visualizing their behaviour. One way (but a promising way) is to expose their parameters and create lots of different results of the algorithms by varying these parameters. The summary of the results gives an overview of what this “black box” is capable of. Pick your favourite algorithm/algorithm family (SVM, clustering, Deep Learning, Neural networks, etc.) and develop such a tool.
Interesting aspects:
In the last years Deep Networks, a special kind of artificial neural network with many layers, have revolutionised many fields such as Natural Language Processing or Computer Vision.
For image classification the Deep Networks are able to distinguish 1000s of different classes, unfortunately it’s not always clear for which type of class (e.g. dogs) the network works better and for which it doesn’t. In classic machine learning there’s the concept of confusion matrices which are a way to organise classification and mis-classification results in a simple matrix. While standard visualizations of these matrices are still usable up to about 12 classes, they unfortunately won’t scale up to matrices of size 1000x1000 as encountered in modern Computer Vision datasets.
Your job is to create new visualisations that scale to very large confusion matrices and enable an computer vision expert to understand the classification accuracy of his current algorithm, i.e, a convolutional neural network.
There has been a deluge of open data by various government and governmental organization over the last few years. While this is admirable, what good is all this data doing if the common citizen is not being able to understand, explore, nor learn from this data. Hence, the goal is to develop a tool (ideally) web based that helps people to explore such data. One of the challenges will be to gear this tool toward a broad set of people, hence you cannot assume a great visual literacy (a problem the New York times has been struggling with and perhaps is providing some ideas for). Further, it is unrealistic to provide a universal tool where all types of data can be explored with and all questions can be answered with. Hence, it’ll be important to narrow your focus on specific aspect of civic life. There are quite a number of open data sources that you can choose from:
A nationally representative sample of eighth-graders were first surveyed in the spring of 1988. A sample of these respondents were then resurveyed through four follow-ups in 1990, 1992, 1994, and 2000. On the questionnaire, students reported on a range of topics including: school, work, and home experiences; educational resources and support; the role in education of their parents and peers; neighborhood characteristics; educational and occupational aspirations; and other student perceptions. The .xls file contains 2000 records of students’ responses to a variety of questions and at different points in time. The codebook explains the question and answer codes.
This data set is Florida election data from the CMU Statistical Data Repository (Note: when downloading these files, be sure to use the correct “save-file” operation for your browser. IE tends to add extra characters that confuse the programs.)
This contains roll call data from the 108th House of Representatives: data about 1218 bills introduced in the House and how each of its 439 members voted on it. The data covers the years 2003 and 2004. The individual columns are a mix of information about the bills and about the legislators, so there’s quite a bit of redundancy in the file for the sake of easier processing in Tableau.
Have you ever wanted to find more information on government spending? Have you ever wondered where federal contracting dollars and grant awards go? Or perhaps you would just like to know, as a North-american citizen, what our neighbour’s government is really doing with their money.
For a number of years, the Vis, InfoVis, and VAST conferences have created a visualization contest. For each contest a problem scenario together with the relevant data sets have been provided to the research community and a price has been awarded to the best visualization. Some of the problems have been quite challenging. However, for the most part, these are great problems to work on.
While we are not expecting you to create winning entries to these visualization challenges, these are often well thought out problems that are fun and solvable. See whether any are of interest to you.
Two contributed datasets are part of two separate competitions:
To use this as a project idea, you do not need to submit your solution to the competition nor do you have to follow the given analysis objectives. However, feel free toregisterand submit by Nov 30th.