Using Posters to Recommend Anime and Mangas in a Cold-Start Scenario

Jill-Jênn Vie, Florian Yger, Ryan Lahfa, Basile Clement, Kévin Cocchi, Thomas Chalumeau, Hisashi Kashima

Introduction

Recommender systems are the backbone of modern content discovery, yet they consistently struggle with the "cold-start" problem: predicting user interest for items with zero or very few ratings. In the anime and manga industries, where thousands of niche titles are released annually, traditional Collaborative Filtering (CF) often fails to surface new works. While metadata like genre can help, it is frequently sparse. This paper introduces BALSE (Blended Alternate Least Squares with Explanation), a hybrid model that leverages the visual information in promotional posters to bridge the gap between new releases and user preferences.

The BALSE Architecture

The BALSE framework integrates deep learning-based visual analysis with traditional matrix factorization. The system is designed to transition smoothly from content-based recommendations to collaborative filtering as an item gains popularity.

The model consists of four primary components:

  1. Illustration2Vec: A VGG-16 based CNN that extracts semantic tags from posters.
  2. ALS (Alternate Least Squares): A standard collaborative filtering block for "warm" items with existing ratings.
  3. LASSO (Least Absolute Shrinkage and Selection Operator): A regularized regression model that maps user preferences to specific visual tags.
  4. The Steins Gate: A sigmoid-based blending function that weights the predictions of ALS and LASSO based on the number of ratings an item has received.

Visual Semantic Extraction

A core insight of this research is that anime posters are highly stylized and rich in "tropes." Unlike live-action movie posters, which may focus on actor names, anime posters are designed to communicate the "vibe" and specific character archetypes (e.g., "magical girl," "mecha," "school life") immediately.

By using Illustration2Vec, the system converts raw pixels into a vector of weighted tags. The LASSO component then identifies which of these tags a specific user tends to favor. This approach provides a significant advantage in explainability; the system can justify a recommendation by highlighting the specific visual elements (like "red hair" or "swords") that align with a user's history.

Experimental Results

The authors evaluated BALSE using the Mangaki dataset, which includes over 330,000 ratings from 2,000+ users across nearly 10,000 items. The performance was measured using Root Mean Squared Error (RMSE).

In the most challenging "cold-start" scenario (items with zero ratings), BALSE achieved an RMSE of 1.345, a marked improvement over the baseline ALS score of 1.500. Even as items moved from "cold" to "warm," BALSE maintained a slight edge or parity with traditional methods, proving that visual data provides a valuable prior without degrading overall system accuracy.

The "Steins Gate" blending mechanism proved effective, with a learned threshold ($\gamma$) indicating that even a single community rating allows the system to begin incorporating collaborative data.

Conclusion and Future Work

BALSE demonstrates that promotional posters are a viable and rich data source for overcoming the item cold-start problem in stylized media. By combining the "black box" of deep learning with the sparsity-promoting LASSO, the authors created a system that is both accurate and interpretable. Future research may explore more modern CNN architectures, deeper analysis of artistic styles (beyond simple tags), and the application of similar logic to the "user cold-start" problem.

Summary updated

BibTeX
@inproceedings{vie2017using,
  author = {Jill-Jênn Vie and Florian Yger and Ryan Lahfa and Basile Clement and Kévin Cocchi and Thomas Chalumeau and Hisashi Kashima},
  title = {{Using Posters to Recommend Anime and Mangas in a Cold-Start Scenario}},
  booktitle = {MANPU@ICDAR},
  year = {2017},
  url = {https://openreview.net/forum?id=Tlt_0pYq4Mk},
}
Figure