User Embedding Model for Mobile Game LTV Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for predicting the lifetime value (LTV) of users playing mobile F2P games lack a deep and broad understanding of various user types and their actions, and are unable to effectively utilize multiple data sources, leading to inaccurate LTV predictions.
Innovation Solution
A computing system employing an artificial neural network-based user embedding model that integrates subscription, status, and action information to predict LTV, utilizing a complex architecture including deep neural networks, autoencoders, convolution layers, Time2Vec layers, and multi-layer transformers to process time-series data and concatenate features for accurate predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing LTV prediction methods (RFM, BYTD, simple machine learning) are used, then the prediction process is simple and fast, but the understanding of user types and actions is shallow and inaccurate
Solution Approach 1:
The patent segments user data into multiple distinct types (subscription information, status information, action information) and processes each type through dedicated neural network branches. This segmentation allows the system to capture nuanced user characteristics while maintaining manageable model complexity through modular architecture.
Solution Approach 2:
The patent transforms traditional tabular user data into multi-dimensional embeddings using deep neural networks, autoencoders, and transformer layers. This dimensional transformation enables the model to capture complex user behavior patterns and relationships that shallow methods cannot detect, significantly improving LTV prediction accuracy.
2Loss of information
If multiple data sources are integrated, then comprehensive user understanding is achieved, but the curse of dimensionality occurs making prediction difficult
Solution Approach 1:
The patent extracts essential user characteristics from multiple data sources by processing each data type through specialized neural network components. The embedding layers and autoencoders extract latent features that capture the most important user behavior patterns while filtering out redundant information, effectively managing dimensionality.
Solution Approach 2:
The patent introduces embedding layers as intermediary representations between raw multi-source data and the final LTV prediction. These embeddings serve as a bridge that transforms high-dimensional, heterogeneous data into a unified, lower-dimensional feature space that preserves essential information while facilitating efficient processing.
3Measurement precision
If traditional feature processing methods are used, then the model is easy to implement, but it cannot capture complex user action patterns and time-series behavior
Solution Approach 1:
The patent replaces traditional mechanical feature engineering methods with deep learning-based automatic feature extraction. Neural networks, particularly the transformer and autoencoder components, automatically learn complex user action patterns from raw data without requiring manual feature engineering, capturing temporal and behavioral nuances that traditional methods miss.
Data Source
AI summary
A method for predicting life time value (LTV) of a user playing a mobile F2P game and a computing system for performing the same are disclosed. According to an embodiment of the present disclosure, it is possible to accurately predict LTV of a user using various types of features, while addressing the curse of dimensionality problem that may occur when using a very large number of features.