Tag‑Based Precise Advertising Recommendation Algorithm for Low‑Volume Channels

The article describes the background of game advertising recommendation: traditional search cannot satisfy users' game discovery needs, and users often cannot express their interests with keywords. Personalized recommendation is needed to help users find relevant game ads among massive ad materials.

In the advertising market, the buy‑side faces two main challenges: (1) extremely small traffic for many channels (often fewer than 100 clicks per day), leading to highly sparse data and low‑dimensional effective features; (2) limited user attribute information, especially for new users, causing cold‑start problems. These motivate the development of a tag‑based precise delivery algorithm called "Xianzhi".

Algorithm Principle

The system uses three types of tags: material tags (created by artists), user tags (derived from request time and cookies), and tag click‑through rate (CTR) matrices. Material tags are one‑hot encoded into a matrix T , while user attributes are processed into a matrix A . Historical logs are used to compute a tag‑CTR matrix C that reflects the probability of a user clicking a material under each attribute.

When a user requests an ad, the engine retrieves the relevant material set M = [C1, C2, …, Cm] and the user attribute vector A . The recommendation score for each material is calculated as R_m = Σ (weight_i × feature_i) , and the material with the highest R_m is served.

Algorithm Optimization

1. Click‑Data Confidence : Small‑sample channels produce unreliable CTRs. The Wilson interval is applied to obtain a confidence‑adjusted CTR, improving estimation for low‑traffic materials.

2. Time‑Window Weighting : User interests evolve over time. Clicks from recent time windows are given higher decay weights, and the weighted CTR is computed using the Wilson‑adjusted values.

3. Dimension Expansion : To alleviate sparsity, time slots are merged into four periods and city dimensions are aggregated to provinces. This reduces zero entries in the CTR matrix.

4. Weight Determination : Objective weighting based on coefficient of variation (standard deviation divided by mean) is used to assign relative importance to each attribute (e.g., UID, browser). The resulting weights improve model CTR by 9.7% in A/B tests.

5. Tag Accuracy : Manual tags suffer from inconsistency, missing tags, and duplication, leading to identical recommendation scores for different materials. The article proposes a deep‑learning‑based automatic tagging solution (described in a subsequent article).

Algorithm Effect

Offline calculations generate tag‑CTR matrices hourly; online serving retrieves the appropriate matrix for the current request and ranks materials. A/B testing shows that the optimized algorithm increases click‑through rate by roughly 10% compared with manual placement.

Current Issues and Outlook

The system still faces data sparsity (e.g., 16.7% of user‑tag request counts are zero), extremely low‑traffic channels, and limited user attribute collection due to cookie deletion. Future directions include expanding user attributes via platform‑wide data integration, aggregating channel data based on cosine similarity of user attributes, and employing deep learning for automatic material tagging.