Source-linked AI summary
Book Readership During Movie Releases: An Exploratory Analysis
Sushobhan Parajuli, Vittoria Vineis, Samira Vaez Barenji, Michael D. Ekstrand
TL;DR
The paper examines whether movie adaptations create exogenous, time-bounded relevance shifts that recommender systems may miss before interaction data reflects them. It links Goodreads readership with movie releases and evaluates event-centered readership patterns and recommendation scores. Adapted books show readership spikes around release, while existing recommenders provide no statistical evidence of systematically different scores for recently adapted books.
Problem
Exogenous item-specific events may temporarily change recommendation relevance before sufficient interaction data accumulates, but whether existing recommenders capture movie-release effects is unknown.
Method
The study links Goodreads interactions to Wikidata and TMDB movie-book pairs, analyzes readership with event studies and DiD, and estimates covariate-adjusted associations in model-specific recommendation scores.
Results
Movie-adapted books show pronounced readership increases around release across popularity deciles, while personalized models show no evidence of systematically higher or lower scores for recently adapted books.
Takeaways & Limitations
Movie releases are associated with measurable readership spikes, but existing collaborative filtering models do not show statistical evidence of accounting for their impact on book consumption.
Takeaways & Limitations
Inference is limited by residual confounding, approximate SUTVA, heterogeneous adaptations, possible unreflected upcoming events, and few treated books or matched groups.
Abstract
from arXiv · showhide
Exogenous events can temporarily change the relevance of items in recommender systems, but these shifts are often not visible in historical interaction data until after users have already responded. In book recommendation, movie adaptations provide a clear example of such events: the release of a movie based on a book can temporarily increase attention to the source text and change its relevance for some readers. We examine this phenomenon using a large Goodreads dataset matched to movie release dates. We find a clear spike in readership around the release month, and then we evaluate existing recommendation models to understand how they rank movie-adapted books around the movie release date.
1 Introduction
Recommender systems rarely incorporate exogenous item-specific events, although movie releases can create a time-bounded increase in attention to adapted books before standard models can respond. The paper asks whether readership changes around adaptations and whether existing recommenders surface or score these books differently.
- Movie adaptations can renew attention to source books from new readers, returning fans, and audiences comparing adaptations with their sources.
- Prior research finds book interest increases around movie releases, with effects strongest immediately before and after release.
- Movie releases provide a meaningful, time-bounded demand signal before sufficient new interaction data accumulates for standard recommenders to respond.
- The study asks whether adaptations change readership across popularity levels and whether recommenders assign different scores to recently or imminently adapted books.
2 Data
The study combines a large Goodreads interaction dataset with linked movie-book pairs and defines a release-centered window for identifying active adaptations.
- The UCSD Book Graph contains over 228 million user-book interactions across 876,145 users and 1.52 million books from January 2007 to November 2017.
- Linking Wikidata, TMDB, and Goodreads data yields 287 movie-book pairs for US movies released between 2007 and 2017 with at least 1,000 cumulative TMDB votes.
- A book is active when an interaction or recommendation falls from one month before through five months after the movie release.
3 Movie Release and Book Readership
Event-study and Difference-in-Differences analyses compare adapted books with readership-matched controls around release dates. Adapted books show release-window readership spikes across popularity levels, with heterogeneous relative gains and limited evidence of pre-trend differences.
- 3 Movie Release and Book Readership: The analysis uses event-study visualizations and a Difference-in-Differences framework to characterize readership timing and heterogeneity without claiming causal identification.
- Event-Study Design: 261 adapted books are compared with up to five nearest-neighbor controls across pre-release readership deciles and three consecutive six-month windows.
- Event-Study Design: Matching reduces pre-release readership SMD from 2.63 to 0.22, improving balance between treated and control books.
- Event-Study Analysis: Across all ten popularity deciles, adapted books exhibit pronounced readership increases around release that are absent among matched controls.
- Movie-Release Effect: The placebo estimate is DiD = −48,617 with p = 0.14, providing limited evidence of differential pre-trends before treatment.
- Heterogeneous Effects: Relative readership increases for low-popularity books are approximately five to six times larger than for the highest-popularity decile.
- The readership pattern suggests movie-adaptation status may be informative for recommendation, motivating analysis of recommender scores.
4 Movie Release and Recommendation Scores
The study tests whether standard recommenders surface movie-adapted books during release windows and whether adaptation status changes their scores relative to matched books. Retrieval analysis shows adapted books are often missed or ranked low, while ATT estimates provide no systematic score difference across models.
- Recommendation evaluation: The evaluation analyzes rankings and recommendation scores for active movie-adapted books relative to comparable non-adapted books.Four recommenders are trained before January 1, 2017 and evaluated during the following two months.
- Retrieval and ranking: 33,631 interactions involving 24 active movie-adapted books show that adapted books are frequently missed or buried in candidate lists.Popularity retrieves the highest interaction share and concentrates retrieved interactions near the top, while personalized models achieve higher top-100 rates.
- Score-effect estimation: The score analysis estimates book-equal ATT using matched groups, comparing each treated book with matched controls for users receiving scores for both.Matching uses coarsened exact stratification followed by within-stratum nearest-neighbor refinement; uncertainty is estimated with a matched-group block bootstrap.
- Score-effect estimation: For the three personalized models, ATT estimates are small and their 95% confidence intervals include zero, providing no evidence of systematic score differences.The popularity baseline has the largest standardized value, but its interval also includes zero and is based on only G=15 matched groups.
- Assumptions and limitations: The ATT estimates have causal interpretation only under conditional ignorability, overlap, and SUTVA.Residual confounding, approximate SUTVA, heterogeneous adaptations, unreflected upcoming adaptations, and limited treated-book counts constrain inference.
5 Conclusion
Movie releases are associated with measurable readership spikes in adapted books, but the study finds no statistical evidence that collaborative filtering models systematically account for movie-release effects. The authors therefore motivate event-aware recommendation strategies and evaluation across pooled release windows.
- Movie releases are associated with measurable readership spikes in adapted books.
- The study finds no statistical evidence that collaborative filtering models account for or fail to account for movie-release impact on book consumption.A modest effect cannot be ruled out.
- The authors motivate adding adaptation status and time-to-release as temporal item features or applying post-hoc re-ranking during release windows.Because individual events are rare at inference time, evaluation likely requires pooling many historical release windows.