Book Ratings and Recommendations

March 27, 2026·39 min·2 clips

Hannes, who uses LLMs freely for coding, refuses to let them write a single sentence of his fiction — and explains exactly why.

Goodreads supplies the data. Users mark books as read, reading, or want to read, rate them from one to five, and connect through a friend network, so the paper has more to work with than title metadata. The host starts with a simple hunch. Maybe people rate books highly because they like certain genres, follow familiar authors, or react to visible features. Hannes narrows that down. He is looking for hidden variables, including page count, but he treats them as weak signals, not the secret to aesthetic appreciation. Page count becomes the useful test case. Hannes says individual readers can show statistically meaningful links between length and rating, while making clear those links are not close to overwhelming. There is no silver bullet. If page count alone produced a 0.7 or 0.8 correlation, he says, that would describe a strangely mechanical reader who loves all short books and hates all long ones. That example carries the conversation. It keeps predictive convenience separate from any serious claim about whether a book is good. The host's skepticism is methodological. He asks whether missing variables could explain the weak relationship, then gives Hannes room to explain why many plausible predictors still look too small to build around. The average reader is the trap. Writing toward statistical sweet spots means chasing a vague aggregate person instead of the actual target reader group or genre group an author cares about. The exchange stays grounded in the data. Recommender data is useful, but it does not cleanly answer whether a book is better in a broad aesthetic sense. Near the end, Hannes turns briefly to his own fiction, shaped by psychology and academic life, and points listeners toward academic fiction, magical realism, campus stories, his website, and University of Amsterdam contact routes.

As heard by us

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?

Read the full review in PlayNext →

Why you'd press play

You want Goodreads data to help explain why some books land and others do not.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Data Skeptic