Experiments
Getting closer to the end….
Really kicked into overdrive this week as we approach the end of SWITCH 2021 Research Experience for Teachers. My main focus was trying out some different experiments with the Wisconsin Breast Cancer data, including reducing the number of features included in my models to only three by looking at those with the highest correlation to the diagnosis. I visualized this using the heatmap function in seaborn and chose to look at radius, perimeter, and concave points as they had the best correlation when looking at the three features which fall under each category: mean, se, and worst. Then I trained seven machine learning models on each set of three features(mean, se, and worst) associated with each category (radius, perimeter, and concave points) and calculated the training and testing accuracy for each. I found that linear SVM had the highest performance for each with only a marginal reduction from utilizing all 31 features included in the total dataset to train the models, where the Random Forest Classifier had the best testing performance. I believe it is possible to get an adequately accurate model for detecting breast cancer even with a significantly reduced number of features in the training and testing dataset given that those specific features are highly correlated with the diagnosis.
I also really got deep into using the seaborn python package to visualize not only data correlation but also to visualize the individual support vectors for my trained SVM models that support drawing the hyperplane to adequately create the maximum margin for separation of features within the dataset. I found using seaborn to have some advantages compared to matplotlib due to the more functional syntax that it implements. I also find the 3D plots produced by seaborn to be more visually appealing. The downside to these visualizations though is when you are using more than two features as getting the 3D model oriented correctly to be able to see adequate separation is somewhat difficult and in some cases not really possible based on the research and work I have done with it so far.
Faculty talks
We also attended a presentation on Designing Effective Posters presented by Dr. Stephen Perry which was very enlightening. He showed many examples and guidelines for creating an engaging, stand-alone poster that tells a story while being informative and attention grabbing. I learned many strategies that not only helped me in designing my poster for the SWITCH presentation but will also be great information I can pass on to my students when they are working on Science Fair or other academic posters. I am confident that Dr. Perry’s presentation is one that I will continue to reference for quite some time.
We also attended a talk from Dr. Dino Di Carlo, faculty at UCLA, on “Lab on a Particle” technology, which involves creating tiny assays for running many tests simultaneously through 3D fabricated particle-sized “flasks”. This technology also has the potential to leverage existing laboratory technology in production and in application. These 3D structures are like a compartment for containing droplets within a larger sample utilizing oil emulsions with the sample to create “dropicles”. This talk was incredibly interesting and also discussed many challenges in the creation and implementation of “Lab on a Particle”. I am excited to see the future applications of Dr. Di Carlo’s work.
