Leveraging Peripheral Behavioral Signals to Predict Sparse Monetization Decisions: The Parallel Sequence Network

Data scarcity in labeled outcomes poses a fundamental challenge in predicting infrequent yet economically consequential human decisions, such as purchasing, subscribing, or gifting. Because these monetization outcomes are rare and costly to observe at the individual level, digital platforms often face severe shortages of reliable supervision signals. At the same time, platforms routinely collect abundant peripheral behavioral data that are unlabeled and only weakly correlated with monetization outcomes. Existing supervised prediction models that depend heavily on monetization labels tend to overfit to a small group of known spenders, struggle to generalize to users who have never monetized before, and suffer from delayed feedback and extremely imbalanced distributions. These limitations highlight a growing need for IS research on how to extract monetization-relevant signals from large-scale behavioral traces. To address this challenge, we propose a Parallel Sequence Network (PSN) that leverages abundant non-monetary engagement data to complement sparse monetization signals. Within the live-streaming context, PSN separately models short-term viewing behaviors to capture emerging interests and long-term gifting behaviors to preserve stable preference trajectories. A unified latent representation balances exploitation (repeated gifting to familiar channels) with exploration (interest in new channels) while mitigating the outsized influence of dominant high spenders. Furthermore, a two-phase transfer learning structure enables knowledge from viewing sequences to enhance monetization prediction for users with little or no monetization history. We evaluate the PSN framework using large-scale field data from a leading live-streaming platform in China. Results show that PSN significantly outperforms state-of-the-art baselines, improving key metrics (e.g., Hit Rate) by 3.6% to 191.7%. PSN demonstrates strong performance in early-stage spender identification and remains robust under severe class imbalance and temporal misalignment. We additionally demonstrate practical improvements in downstream tasks, including channel discovery and cold-start recommendations. This study contributes to IS research by establishing a scalable and generalizable approach for inferring value-creating intentions from weakly correlated peripheral behaviors, moving beyond outcome-only learning. Our findings also inform platform monetization strategies, offering data-driven guidance for resource allocation, traffic distribution, and sustainable growth in digital markets.