Kyle Polich interviews ML researchers dissecting how recommender systems actually work.
Data Skeptic is a research-minded technology interview show focused here on how recommender systems behave in the real world. Its conversations move from Goodreads ratings and Netflix-style layouts to fairness, eye tracking, human curation, and coordinated attempts to influence recommendation engines.
PRESS PLAY
The show is a one-on-one interview format: host Kyle Polich talks with a single guest per episode about their research or professional work in data science and machine learning. In the sampled run, nearly every episode centers on some facet of recommender systems, suggesting a thematic series rather than one-off topic hopping. Recurring subjects include explainable AI applied to job and product recommendations, disentangled representation learning and whether it actually improves interpretability, the carbon footprint of training graph neural network recommenders, and reproducibility failures in academic ML papers caused by missing hyperparameters. Other episodes look outward from pure ML methods to how recommendation plays out in practice, such as human curators at services like Sling TV shaping candidate pools before algorithms personalize them, or how ratings data from sites like Goodreads reflect reader identity more than book quality. Guests are predominantly early-career researchers from European and North American universities (Copenhagen, Maastricht, Colorado Boulder, Rome, Amsterdam) discussing thesis or dissertation work, plus occasional industry guests with hands-on platform experience. The tone throughout is cautious and methodical, with guests and host frequently qualifying findings ('a cautious yes, with limits') rather than overstating results. The show suits listeners who already have some technical grounding in machine learning and want research-level depth rather than industry news or product announcements. Production is a straightforward two-person conversation with brief sponsor mentions, no narrative framing or sound design beyond the interview itself.
Made for: Data scientists, ML researchers, and graduate students who want research-level discussion of recommender systems, interpretability, and fairness rather than general AI news.
What sets it apart: Favors direct interviews with PhD researchers presenting their own peer-reviewed or thesis work on narrow recommender-systems subfields, rather than broad industry commentary or news roundups.
The Data Skeptic Podcast features interviews and discussion of topics related to data science, statistics, machine learning, artificial intelligence and the like, all from the perspective of applying critical thinking and the scientific method to evaluate the veracity of claims and efficacy of approaches.
Based on 6 episodes we listened to · September 2026
Data Skeptic turns recommender-system research into precise conversations about algorithms, human judgment, environmental cost, and unintended consequences.
Data Skeptic treats recommender systems as systems worth questioning from several directions. The interview-led format usually starts with a research paper or industry problem, then has the host test its assumptions, methods, and practical consequences with the guest.
A recommender stops feeling neutral once you learn how users can organize against it.
Press play if you want
Short reviews from the PlayNext desk, based on the episodes we processed.
A grounded recommender-systems discussion that turns Goodreads ratings into a study of weak signals, reader behavior, and the limits of modeling taste.
Data Skeptic uses Goodreads book ratings as a tidy test case for a messier recommender problem: what can ratings say about books, and what do they mostly say about readers?
A technical but useful look at whether disentangled recommender models are actually easier to interpret.
Data Skeptic turns a slippery recommender systems question into something concrete: when a model learns its own latent features, can those features be separated in a way people can inspect?
A clear look at how coordinated users can push recommender systems toward overlooked people and content.
Data Skeptic treats recommender systems as social machinery, not just engines for personal taste. Kat, a PhD student at MIT EECS studying strategic learning, argues that users can coordinate likes, watches, comments, and similar signals to help underrepresented content reach the…
A measured, research-minded look at modular recommendation systems and the case for more algorithmic choice.
Data Skeptic uses recommender systems to ask a clean product question: what if the algorithm were not owned entirely by the platform? The episode frames today's Netflix-style model as an in-house default, then explores a more modular future where users or communities might…
A grounded look at why explanations in job recommendation systems may not improve human decisions as cleanly as designers expect.
Data Skeptic uses job recommender systems to ask a practical question: what happens when an algorithm explains why it surfaced a role? Roan Schrengerhout's research looks less like a pure matching problem than a study of trust, judgment, and misplaced confidence.
A grounded look at why human curation still matters inside industrial recommender systems.
Data Skeptic uses Corey Zechman's work at Sling TV and his long running music blog to test a stubborn question in recommendation systems: where does human taste still matter? The episode is strongest on "algatorial" work, the mix of algorithmic sorting and editorial judgment.
A practical look at when graph recommender models justify their energy cost, and when smaller choices may be smarter.
Data Skeptic puts recommender systems on an energy budget, using Eco-Aware Graph Neural Networks for Sustainable Recommendations to ask whether larger graph models are worth the compute they consume.
Chicago critics DeRogatis and Kot pair new-album reviews with deep dives into records like Rumours.
Billboard's Keith Caulfield and Katie Atkinson break down the week's Hot 100 and 200 charts.
Ben Brock Johnson and Amory Sivertson trace viral internet stories from Reddit threads to courtrooms.
Chris Black and Jason Stewart riff on fashion, food, and celebrity culture three times weekly.
Are some books better than others?
Disentanglement sounds intuitive.
Recommendations look individual.
Decoupled recommenders are the question.
Job recommendation explanations can add useful friction, but only if people can tell whether the reasons are any good.
This episode explores how Principal Components Analysis (PCA) in recommender systems can lead to unfairness, covering challenges and solutions in building fair systems through an interview with a Cornell researcher.
Human curation still has a job.
This episode explores how eye tracking technology can provide user feedback for recommender systems, addressing challenges like the cold start problem and discussing research on user behavior in recommendation interfaces.
This episode explores the complexities of building recommender systems, contrasting them with standard machine learning classification and discussing hybrid approaches involving collaborative filtering, embeddings, and bandit learning.
An interview about building recommender systems for digital humanities, specifically for the Monasterium.net historical charter collection.
An interview with a researcher about a new Python library for reproducible recommender system analysis.
This episode explores shilling attacks on recommender systems, where malicious users create multiple profiles to manipulate recommendations.
An interview with Rebecca Salganek about fairness in music recommender systems and playlist generation.
An interview with a machine learning researcher about how recommender systems amplify popularity bias and how to serve niche content.
An interview with a doctoral candidate about building sustainable tourism recommender systems to combat over-tourism and exposure bias.
Discussion of a paper on interpretable graph neural network explanations for real estate recommendations.
This episode explores methods to infer the influence of opaque recommender systems on social media, featuring researchers who discuss user modeling without platform exposure data and the societal impacts like polarization.
Training has a carbon cost you can measure.
This episode introduces a new season on recommender systems, featuring a discussion with network scientist Asaf on graph-based approaches.
The host discusses a DIY network analysis project using 10 years of podcast guest data and their scholarly publications.
This episode discusses vulnerable networks and complexity theory in network science, exploring how hubs affect network dynamics.
An interview with Professor Baruch Barzel about the dynamics of networks and universal patterns in network science.
This episode explores how graph data structures and network science can be applied to cybersecurity for detecting fraud and malicious activity by transforming alerts into networks to focus on main problems.
This episode explores how graph neural networks and graph databases can be applied to optimize supply chains, discussing real-world examples like the Suez Canal blockage and the adoption of network approaches in operations research.
This episode discusses using graph neural networks to classify rooms in floor plans and to create a navigation app for robots.
The episode discusses the concept of graph-of-thought and hypergraph-of-thought techniques for improving LLM reasoning.
An exploration of the challenges faced by the graph database industry, focusing on the long-tail distribution of graph use cases and the potential of standardization and LLMs.
This episode introduces a new season on graphs and networks, featuring a network analysis consultant discussing bot farm detection using PageRank.
The hosts reflect on the season about animal intelligence, discussing the process of producing the podcast and the challenges of finding guests.
This episode explores how large language models can be applied to process mining to help businesses analyze event log data and identify operational bottlenecks for optimization.
This episode discusses a new dataset of animal track images and how machine learning, particularly attention maps, can identify species from noisy track images.
This episode discusses the SATBIRD dataset for bird distribution modeling using remote sensing, satellite imagery, and citizen science data, with insights into challenges like imbalanced data.
This episode explores ant intelligence, learning, and colony-level cognition through an interview with an expert.
Interview with Madeline Vilmes about her book 'Computing Skills for Biologists', covering Unix, shell scripting, and transitioning from academia to industry.
This episode discusses how data science and statistical methods are used to model biodiversity, predict species distributions, and identify hotspots.