Personalized Recommendation of Game Cosmetic Items: From Popularity to Latent Factor Models

In online games, visual items like character costumes, weapon skins, and mounts are purely aesthetic and do not affect gameplay stats, making their recommendation a challenge that requires converting subjective beauty into objective features.

The first practical approach for cold‑start scenarios relies on item popularity combined with recent in‑game behavior, assuming that a player’s recent hero usage correlates with their desire to purchase that hero’s costumes.

While popularity‑based recommendations are simple, they suffer from bias due to varying release dates and item counts, and they cannot balance personal interest against herd behavior, limiting exposure of long‑tail items.

The second approach draws parallels with fashion e‑commerce, using user demographics, browsing history, and purchase records to infer style preferences, but game contexts lack many of these signals.

The third approach creates explicit features by tagging items: professional artists can provide design‑oriented descriptors, and crowdsourced player tagging can capture perceived aesthetics.

Beyond manual tags, a latent factor model (LFM) can learn hidden features from the player‑item interaction matrix, enabling the calculation of preference scores via matrix factorization.

Training data are constructed from purchase (positive) and equally sized non‑purchase (negative) samples, and gradient descent optimizes the latent matrices; the resulting scores rank items for each player.

Although LFM lacks the interpretability of manual tags, it is effective when explicit features are scarce, especially for offline recommendation of low‑frequency cosmetic purchases.

In conclusion, successful game‑item recommendation must blend quantitative methods with an understanding of player aesthetics, adapting algorithms to each game’s economy and player base.