Educating people about vaccination tends to target vaccine acceptance and reduction of hesitancy. Social media provides a promising platform for studying public perception regarding vaccination. In this study, we harvested tweets over a year related to vaccines from February 2018 to January 2019. We presented a framework comprising of two-stage classifier to: (1) classify the tweets as relevant or non-relevant and (2) categorize them in terms of pro-vaccination, anti-vaccination, or neutral outlook. We found that the classifier was able to distinguish clearly between anti-vaccination and pro-vaccination tweets, but also misclassified many of these as neutral. Using Latent Dirichlet Allocation, we found that two topics were sufficient to describe the corpus of tweets. These dealt with: (1) consequences of vaccination/non-vaccination, and (2) promotion of vaccination/non-vaccination. Finally, using the NRC emotion lexicon, we found practically significant differences in emotions expressed about vaccination between vaccine outlooks, but no practically significant temporal differences across a year. By analyzing negative emotions in people opposing vaccination, healthcare administrates can employ target intervention programs addressing these tweets.

Related Publications and Presentations:

• Mahajan, R., Romine., W., Miller., M., & Banerjee., T. (2019). Analyzing Public View towards Vaccination using Twitter. Proceeding from the IEEE International Conference on Big Data (IEEE Big Data 2019). Los Angeles, CA.
• Presentation at The College of Science and Mathematics Festival of Research – 2019

Choosing to not teach evolution can have an outsized impact in that all affected students will likely have a poor understanding of evolution. Emotional factors play a role in that decision. Thus, we develop and utilize a novel assessment which measures emotional experience in light of real-world evolution education scenarios. We presented a diverse sample of 296 science teachers 4 pro-evolution and 8 anti-evolution scenarios and asked them to rate their levels of joy, anger, sadness, fear, disgust, shame, and guilt elicited by that scenario. We extracted the most important factors in the responses using factor analysis, explored the validity of the subscales using Rasch analysis and found the most important factors driving emotional dispositions using stepwise regression. We extracted 3 factors that fit well with the Rasch model: (1) pro- evolution orientation (positive emotions on pro-evolution and negative emotions on anti-evolution scenarios), (2) anti-evolution orientation (negative emotions on pro-evolution and positive emotions on anti-evolution scenarios), and (3) feelings of regret over anti-evolution events (shame and guilt on anti- evolution scenarios). Acceptance of evolution facts and a non-theistic religious orientation were positively related to pro-evolution experience. Anti-evolution experience was predicted by lack of microevolution acceptance and lack of teacher preparation. Feelings of regret around anti-evolution statements were driven by acceptance of evolution facts and lower levels of teacher preparation. This work advances our knowledge of how teachers relate affectively to the theory of evolution and offers a framework for scientific evaluation of teachers’ emotions as they pertain to professional development or practice.

Related Publications and Presentations:

• Romine, W., Mahajan, R., & Todd, A. (September 2018). Measuring science teachers’ emotions around evolution using real-world scenarios. Presentation at the Three Rivers Evolution Event (TREE), University of Pittsburgh.
• Romine, W., Mahajan, R., & Todd, A. (March 2019). Measuring science teachers’ emotions around evolution using real-world scenarios. Presentation at the NARST 2019 Conference.

Wearable sensor devices are growing in popularity and are a potential source of data which could be used predict physical impairments as a result of cognitive disorders such as dementia. Such devices allow for the non-intrusive collection of heart rate, breathing rate, acceleration, and cadence, along with other metrics. In this study, we apply machine learning techniques to data collected from a Hexoskin vest with sensors measuring acceleration, cadence abdominal and thoracic respiration, minute ventilation, breathing rate, electrocardiography (ECG), and heart rate. The data was collected from elderly participants in the Fels Longitudinal Study wearing the vest while performing a short physical performance battery (SPPB) walk. Machine learning techniques such as support vector regression and convolutional neural networks were applied to predict the Montreal Cognitive Assessment (MoCA) scores of the subjects (regression) and determine if they may have some form of cognitive impairment reflected in their physical performance (classification) based on the estimated MoCA score. The Support Vector Regression model overestimated the true MoCA scores, so strategies such as introducing truncating or rounding off MoCA scores and removing outliers were implemented for mitigating the tendency of marginal misclassification. The performance of Convolution Neural Network and Support Vector Regression was similar for classification. The Convolution Neural Network had a large number of spuriously low predictions as the dataset was small. Thus, the network trained was always underfitting and unable to generalize well to new data.

One can perceive substantial changes in people’s needs and fashion in today’s trendy world. The companies suffer a lot to withstand the whimsical demands of customers thus sometimes succumbing to loss instead of profits. In such a case, predicting business sales by accumulating historical data trend marks as a logical step. A sales forecast is an estimate of the number of goods and services one can realistically sell over a period, the cost of goods and services and the estimated profit. Providing an estimation allows business owners to prepare the inventory in advance. The dataset was retrieved from Kaggle which provided seventeen sales estimation features from January 2013 to October 2015 consisted of one million entries from five stores. Data were pre-processed and exploratory analysis such as examining the correlation among features was performed to identify essential features. It was seen that promotion offers highly benefitted the sales for stores, especially during holidays and weekends. Feature selection techniques such as entropy gain, examining effects of features on each other, forward and backward selection method suggested that promotion and competition features are best suitable for estimating the sales. We applied certain simple to complex machine learning algorithms such as linear regression, random forest and gradient boosting to predict the outcome of sales for different stores. The outcome was evaluated using mean squared error, root mean squared error and root mean square percentage error. The best performing model was concluded to be random forest using a train-validation-test split of 60:20:20 for both with interaction features and without interaction features.