Two prediction tasks
The project considers two different questions: whether a user will read a book and what rating they might give it. Both require extracting signals from observed interactions without assuming every user or title has rich history.
How the models work
Started with popularity-based heuristics and examined user–book signals, including Jaccard similarity between audiences of different books.
Modeled the global mean together with regularized user and item bias terms. The accompanying project writeup also explores a latent-factor matrix factorization approach.
Used user averages, item averages, or the global average to produce fallback rating predictions when information is missing.
What I learned
Recommendation systems must balance personalization against sparse interaction histories. The project compares simpler, more interpretable baselines with richer collaborative signals; the materials provided do not establish a reliable headline evaluation score, so no accuracy claim is shown here.
Explore the work
Read the original project materials behind this case study.
Gambling & Loss Aversion