- Belief Thresholds
This paper explores the social dynamics leading to rapid belief change by using an agent-based model written in Netlogo. Underlying concepts implemented include preference falsification, diverse individual thresholds for change, social clustering, and dissent punishment. The ODD protocol is used to describe the model design.
- Genre Classification for Movies
The paper implemented and analysed the use of decision tree and logistic regression models on a dataset of textual ‘tags’ and audio-visual features from each movie’s trailer.
Decision tree models performed poorly even after restricting the depth of the tree, and it is put forth that this is primarily due to the large number of features, limiting the effectiveness of the model’s greedy algorithm. Feature reduction techniques may prove useful for future work.
The logistic regression model outperformed the baseline, though the unbalanced nature of the dataset caused significant over-prediction in high-frequency classes. Weighting the classes during training did not improve overall accuracy. More instances of the low-frequency classes may have made a greater difference. It was also noted that a significant portion of movies share two genres, and that some genres may be a subset of others and therefore a multi-label output may allow the classifier to match reality more closely.