Mobile Gamer Lifetime Value Prediction via Objective Decomposition and Reconstruction

TL;DR

In RTB-driven mobile gaming, user LTV distributions are highly skewed with outliers, hindering accurate prediction. The paper proposes CALTV, a decomposition-and-reconstruction framework that first decomposes $V_{u,i}$ into per-price-category counts via $V_{u,i} = \sum_{m=1}^M C_{m,u,i} V_{m,u,i}$ and uses DR nets to predict $C_{m,u,i}$, followed by reconstructing the final LTV as $\hat{V}_{u,i} = \sum_{m=1}^M V_{m,u,i} \sum_{c=1}^{\hat{C}_m} p_{c,m,u,i} c$. Training relies on truncated counts and cross-entropy losses ${\cal L}_m$, with end-to-end sharing of embeddings; the resulting model, CALTV, shows offline improvements in AULC over baselines and strong online ROI gains via oCPA bidding in TapTap. The approach mitigates outlier impact and improves high-value user discrimination, delivering practical benefits for RTB advertising revenue and advertiser ROI. The decomposition-reconstruction strategy offers a robust pathway for LTV prediction in domains with skewed monetary distributions.

Abstract

For Internet platforms operating real-time bidding (RTB) advertising service, a comprehensive understanding of user lifetime value (LTV) plays a pivotal role in optimizing advertisement allocation efficiency and maximizing the return on investment (ROI) for advertisement sponsors, thereby facilitating growth of commercialization revenue for the platform. However, the inherent complexity of user LTV distributions induces significant challenges in accurate LTV prediction. Existing state-of-the-art works, which primarily focus on directly learning the LTV distributions through well-designed loss functions, achieve limited success due to their vulnerability to outliers. In this paper, we proposed a novel LTV prediction method to address distribution challenges through an objective decomposition and reconstruction framework. Briefly speaking, based on the in-app purchase characteristics of mobile gamers, our model was designed to first predict the number of transactions at specific prices and then calculate the total payment amount from these intermediate predictions. Our proposed model was evaluated through experiments on real-world industrial dataset, and deployed on the TapTap RTB advertising system for online A/B testing along with the state-of-the-art ZILN model.

Figures (6)

• Figure 1: The distribution of the cumulative payment amount of mobile gamers. In lower amount range, a substantial portion of paying users only purchase the initial bundle, as shown by the orange bar. For amount above 10 (depicted by the green bars in the zoomed figure), the distribution roughly follows a log-normal pattern (represented by the red dashed curve) with numerous outliers.
• Figure 2: The schematic representation of the CALTV model. A DNN structure with an embedding layer, several fully connected hidden layers, a layer of decomposition and reconstruction nets to predict the sub-objectives (transaction counts), and an additional reconstruction layer as the output layer to generate the final objective (LTV) by aggregating the intermediate sub-objectives.
• Figure 3: The distribution of model game in-app transaction orders price. Significant portion of transactions occur at specific prices such as CN¥1, CN¥6, CN¥12, CN¥18, CN¥30, CN¥68, CN¥98 represented by the red bars in the top zoomed figure, and CN¥128, CN¥198, CN¥328, CN¥648 in the bottom figure, while transaction orders with other price constitute a negligible portion.
• Figure 4: The distribution of transaction counts of the mobile gamers. Nearly 95% of the mobile gamers are non-spenders, while more than 95% of the paying users make fewer than 5 transactions.
• Figure 5: Comparison of the Lorenz curves between baseline models and the proposed CALTV model. Each curve was plotted by first ranking the evaluation samples in descending order according to the LTV predictions of each model, and then accumulating the payment amount of the samples from the highest ranking one to lowest ranking one. The ground truth curve was derived in the same approach but ranking based on actual payment amount rather than model predictions.