In-Game Content Recommendation
“Which in-game items should we show this player?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
Which in-game items should we show this player?
The average game store shows 200+ items but only 3% get meaningful engagement. A game earning $40M in IAP where 80% of revenue comes from 5% of the catalog has massive untapped potential. Generic featured-item rotations ignore that a player who just unlocked a new character class wants complementary gear, not random skins. Personalized stores lift conversion 30-50% but require understanding the intersection of player progression, inventory gaps, and social trends.
Quick answer
The best in-game recommendation systems connect player inventory, purchase history, progression state, and social data in a graph model. The strongest signals come from inventory gaps (what the player is missing in a set), guild purchase trends (what is trending in their social circle), and complementary item ownership. Graph ML lifts store conversion by 30-50% over static featured rotations.
Approaches compared
4 ways to solve this problem
1. Static featured rotations
Manually curate a rotating set of featured items in the store, occasionally themed around events or seasons.
Best for
Simple to manage and ensures every player sees your highest-margin items. Good for small catalogs under 50 items.
Watch out for
Ignores player context entirely. A warrior does not want mage gear, regardless of how well it sells globally. Conversion rates plateau fast.
2. Collaborative filtering
Recommend items based on what similar players purchased, using player-item interaction matrices (matrix factorization, ALS).
Best for
Works well when you have rich purchase history and want to surface items that similar spenders enjoyed.
Watch out for
Cold-start problem for new items and new players. Cannot incorporate inventory state, progression context, or social trends. Recommends popular items, not contextually right items.
3. Content-based filtering with item metadata
Match item attributes (class, rarity, category) to player profile attributes and recommend items that fit their character build.
Best for
Ensures recommendations are class-appropriate and rarity-appropriate. Easy to reason about and debug.
Watch out for
Too narrow. Recommends items the player already has equivalents for. Misses cross-category discovery like a warrior player who would buy a cosmetic skin.
4. KumoRFM (relational graph ML)
Connect players, inventory, store catalog, and purchase history into a graph. The GNN learns inventory gaps, guild trends, and item affinity patterns across the entire player-item ecosystem.
Best for
Highest conversion lift. Captures inventory set completion, social purchasing trends, and complementary item ownership in one model.
Watch out for
Requires a structured item catalog with clear relationships (sets, classes, categories). Unstructured item data limits the graph signal.
Key metric: SAP SALT benchmark: relational graph ML achieves 91% accuracy vs 75% for XGBoost on flat tables in item recommendation and propensity tasks.
Why relational data changes the answer
Item recommendation in games is a multi-table problem. The right recommendation depends on what the player owns (inventory table), what they are progressing toward (achievement and level data), what their guild is buying (social purchase data), and what complements their existing loadout (catalog relationship data). Collaborative filtering flattens this into a player-item matrix and loses all context: it does not know the player is missing one piece to complete an epic armor set.
Relational models connect inventory to catalog to social graph and learn that players who own a specific epic sword and are missing the matching helm, whose guild members recently bought that helm, purchase it at 5x the base rate. This is not a feature you can hand-engineer into a flat table because the signal depends on the specific combination of what you own, what exists, and what your friends just bought. On games with 200+ item catalogs, graph ML lifts store conversion by 30-50% over collaborative filtering, turning the store from a revenue leak into a personalized shopping experience.
Static store rotations are like a bookstore that puts the same bestsellers on every shelf regardless of who walks in. A great bookstore clerk knows you just finished Book 2 of a trilogy (inventory gap), that your book club is reading a specific author (social signal), and that you prefer hardcovers (preference data). Graph ML is that clerk, working across your entire player base simultaneously.
How KumoRFM solves this
Graph-learned player intelligence across your entire game ecosystem
Kumo connects players, inventories, the item catalog, and purchase histories into a graph where item affinity propagates through ownership patterns and social influence. It learns that players who own specific item combinations and just reached a new progression tier purchase complementary items at 5x the base rate. The model captures trending items within guilds, seasonal preference shifts, and inventory-gap signals that collaborative filtering alone cannot detect.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
PLAYERS
| player_id | level | class | total_spend |
|---|---|---|---|
| PLR401 | 45 | Warrior | $34.99 |
| PLR402 | 22 | Mage | $0.00 |
| PLR403 | 60 | Archer | $129.50 |
INVENTORY
| inv_id | player_id | item_id | acquired_date | source |
|---|---|---|---|---|
| INV01 | PLR401 | SWORD_EPIC_3 | 2025-02-20 | Purchase |
| INV02 | PLR401 | SHIELD_RARE_7 | 2025-02-25 | Drop |
| INV03 | PLR402 | STAFF_COMMON_1 | 2025-02-18 | Starter |
STORE_CATALOG
| item_id | name | category | price_usd | rarity |
|---|---|---|---|---|
| HELM_EPIC_2 | Dragonbone Helm | Armor | $4.99 | Epic |
| STAFF_EPIC_1 | Arcane Focus | Weapon | $7.99 | Epic |
| SKIN_LEG_5 | Phoenix Wings | Cosmetic | $14.99 | Legendary |
PURCHASE_HISTORY
| purchase_id | player_id | item_id | timestamp |
|---|---|---|---|
| PH01 | PLR401 | SWORD_EPIC_3 | 2025-02-20 |
| PH02 | PLR403 | HELM_EPIC_2 | 2025-02-15 |
| PH03 | PLR403 | SKIN_LEG_5 | 2025-02-22 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
RANK TOP 5 STORE_CATALOG.ITEM_ID FOR EACH PLAYERS.PLAYER_ID PREDICT BOOL(PURCHASE_HISTORY.*, 0, 7, days)
Prediction output
Every entity gets a score, updated continuously
| PLAYER_ID | RANK | ITEM_ID | ITEM_NAME | PURCHASE_PROB |
|---|---|---|---|---|
| PLR401 | 1 | HELM_EPIC_2 | Dragonbone Helm | 0.72 |
| PLR401 | 2 | SKIN_LEG_5 | Phoenix Wings | 0.41 |
| PLR402 | 1 | STAFF_EPIC_1 | Arcane Focus | 0.38 |
Understand why
Every prediction includes feature attributions — no black boxes
Player PLR401 -- Warrior, Level 45, $34.99 spent
Predicted: 72% purchase probability for Dragonbone Helm
Top contributing features
Inventory gap (armor set completion)
Missing helm slot
33% attribution
Guild trending purchase
4 guild members bought
22% attribution
Complementary item ownership
Has matching sword
19% attribution
Price vs avg purchase
Within range ($4.99)
14% attribution
Time since last purchase
10 days
12% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about in-game content recommendation
How do you personalize an in-game store?
Connect player inventory, purchase history, progression data, and social connections into a relational model. Rank items by predicted purchase probability for each player. The strongest personalization signals are inventory gaps (completing a set), guild trends (items trending in the player's social circle), and complementary ownership (items that pair well with what they already have).
What is the conversion lift from personalized game stores?
Graph ML-powered personalization lifts store conversion by 30-50% over static featured rotations and 15-25% over collaborative filtering. For a game earning $40M in IAP, a 30% lift adds $12M in annual revenue.
How do you solve the cold-start problem for new items?
Relational models connect new items to the catalog graph through attributes (class, rarity, set membership) and can predict affinity before any purchase data exists. A new epic warrior helm is immediately connected to players who own other pieces of that set. Collaborative filtering has no history to work with and must wait for organic discovery.
Should you use collaborative filtering for game recommendations?
Collaborative filtering is a reasonable baseline for players with purchase history but misses contextual signals. It does not know a player is one piece from completing a set, or that their guild is on a cosmetic-buying spree. For games with rich progression systems and social features, graph ML significantly outperforms collaborative filtering.
What data do you need for in-game recommendation ML?
Four tables: player profiles, inventory (what each player owns), store catalog (item attributes, set membership, pricing), and purchase history. Adding social connections and guild data provides the trend signals that lift conversion further. The catalog structure (sets, classes, complementary items) is especially important for graph ML.
Bottom line: A game earning $40M in IAP that personalizes its store to each player lifts conversion by 30%, adding $12M in annual revenue. Kumo learns inventory-gap signals, guild purchase trends, and item affinity patterns that static featured rotations and basic collaborative filtering miss.
Related use cases
Explore more gaming use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
