Application of Machine Learning Algorithms in Mobile Game Recharge Monitoring

With the increasing investment in mobile business, the quality of mobile games has become a priority. Traditional recharge monitoring based on year‑over‑year or month‑over‑month fluctuations generates many false alarms because mobile recharge orders are highly volatile.

To address this, the team applied machine learning algorithms to predict future recharge trends, reducing daily alarm counts from hundreds to dozens or even a few, and achieving accurate alerts during game maintenance.

Optimization Goal : Eliminate frequent and severe false alarms while providing concise, clear alerts.

Solution Design :

1. Monitoring Granularity – Monitoring must be precise to the level of individual sub‑packages, as each game has many sub‑packages with distinct recharge patterns.

2. Model Input and Output – Input consists of order counts for specific time intervals (e.g., 2019‑04‑14 12:10). The model outputs the predicted order volume for each 10‑minute window over the preceding hour, enabling timely detection of downward trends.

3. Model Algorithm Selection – The problem is a regression task; after evaluating several regressors, XGBoost was chosen for its superior robustness (lowest MSE) despite a slightly higher execution time compared to a simple decision tree.

4. Model Iterative Updating – Because the model suffers from loss and limited features, it is retrained daily using the most recent two weeks of order data to capture the latest trends.

5. Monitoring Thresholds – Different sub‑packages have varying order volumes, so thresholds are set per order‑volume segment rather than using a single global value.

Model Evaluation – Evaluation metrics include MAE, MSE, and R². MSE is used as the primary metric; a smaller MSE indicates higher precision. The MSE formula is shown in the original Figure 3.

Comparative experiments on a single sub‑package dataset (Figure 4) demonstrate that XGBoost achieves the smallest MSE, confirming its robustness. Although decision trees are the fastest, the monitoring system prioritizes model robustness over speed.

XGBoost Algorithm Overview – XGBoost builds an ensemble of weak learners (typically CART) to form a strong regressor. Each weak learner fits the residuals of the previous learners, gradually reducing loss. The process is illustrated in Figure 5.

Key advantages of XGBoost include regularization to prevent over‑fitting, shrinkage (learning rate) to control each learner’s impact, and support for parallel computation, resulting in fast training.

Future Plans – Current limitations involve sub‑packages with very low order volume and sudden order spikes caused by in‑game events. Planned improvements include extending the monitoring window (e.g., 2‑hour or 1‑day predictions) and adding features such as event flags to better handle spikes.

Overall, the machine‑learning‑driven recharge monitoring system significantly reduces false alarms, provides finer‑grained alerts, and offers a scalable solution for mobile game operations.