Blog post

by Sohail Khan. Netflix knows what you like, but does it need to know your politics too? We often face a frustrating choice: share our data and be tracked, or protect our privacy and lose personalization. But what if there was a third option? This article begins by introducing the concept of the privacy-utility trade-off, then explores the methods behind strategic data distortion, a technique that lets you subtly tweak your data to block sensitive inferences (like political views) while still maintaining useful recommendations. Finally, it looks ahead and advocates for a future where users, not platforms, shape the rules, reclaiming control of their own privacy.

by Jonathan Pérez. Reflecting on the complexities of the human experience is paramount to conducting research. Jonathan Pérez, through his exploration of a conspiracy subreddit, reflects on his experience trying to find the human behind the datum. Jonathan critiques the harmful effects of dehumanizing rhetoric and the researcher’s responsibility to navigate ethical implications. In doing so, he establishes three guiding rules to support researchers seeking to humanize their analysis: 1) a researcher must always find the story behind the data; 2) a researcher must protect themselves; 3) a researcher must still humanize participants (even those who perpetuate harmful narratives).

by Sohail Khan. Language models are becoming integral to daily life as trusted sources of advice. While their utility has expanded from simple tasks like text summarization to more complex interactions, the empathetic quality of their responses is crucial. This article explores methods to assess the emotional appropriateness of these models, using metrics such as BLEU, ROUGE, and Sentence Transformers. By analyzing models like LLaMA in mental health dialogues, we learn that while they suffer through traditional word-based metrics, LLaMA's performance in capturing empathy through semantic similarity is promising. In addition, we must advocate for continuous monitoring to ensure these models support their users' mental well-being effectively.

by Jaewon Saw. Imagine perfecting a favorite recipe, then sharing it with others, only to find their results differ because of small changes in tools or ingredients. How do you ensure the dish still reflects your original vision? This challenge captures the principle of stability in data science: achieving acceptable consistency in outcomes relative to reasonable perturbations of conditions and methods. In this blog post, I reflect on my research journey and share why grounding data work in stability is essential for reproducibility, adaptability, and trust in the final results.