Digital Content Session Prediction With Feature Attribution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems struggle to accurately determine user experiences and root causes during digital content system sessions due to the wide range of connectivity levels, device capabilities, and user expectations, leading to inefficient and computationally intensive analytical processes.
Innovation Solution
A system that generates device-specific, geographic, and application-level features using a multiclass prediction deep neural network to predict disruptions and delights, determining contribution levels of these features, and generating attribution reports to identify root causes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If brute force analytical techniques are used to analyze user experience data, then comprehensive analysis coverage is achieved, but computational resources and analysis time are excessively consumed
Solution Approach 1:
The patent segments user experience analysis into multiple independent feature dimensions (device features, geographic features, application features, session features). Each dimension is processed separately through dedicated neural network layers, enabling parallel computation and reducing overall analysis time while maintaining comprehensive coverage.
Solution Approach 2:
The patent transforms raw user experience data into standardized feature vectors with specific parameter transformations. Device characteristics, geographic information, and application metrics are converted into normalized numerical representations that can be efficiently processed by the neural network, improving computational speed without losing analytical precision.
2Measurement precision
If multiple feature types are analyzed to improve prediction accuracy, then user experience prediction precision is improved, but system complexity increases
Solution Approach 1:
The system divides complex feature analysis into distinct modules: device feature extraction, geographic feature extraction, application feature extraction, and session feature extraction. Each module handles specific feature types independently, reducing the complexity of managing all features simultaneously while maintaining comprehensive analysis capability.
Solution Approach 2:
The patent introduces feature extraction layers as intermediary components between raw data and the prediction model. These intermediary layers process and transform diverse input features into standardized representations, simplifying the overall system architecture by creating clear separation between data sources and prediction logic.
3Measurement precision
If detailed feature extraction is performed for device, geographic, and application characteristics, then root cause identification accuracy is improved, but computational resources are increased
Solution Approach 1:
The patent performs preliminary feature extraction and transformation before the main prediction process. Device characteristics, geographic information, and application features are pre-processed into standardized vectors in advance, reducing the computational burden during real-time prediction while maintaining high root cause identification accuracy.
Solution Approach 2:
Computational resources are allocated efficiently by segmenting the analysis into separate feature extraction streams. Each stream (device, geographic, application) processes its specific features independently, allowing for optimized resource utilization and avoiding redundant computations across all feature types.
Data Source
AI summary
The disclosed computer-implemented methods and systems leverage machine learning techniques to generate disruption and delight predictions associated with digital media sessions. For example, the methods and systems discussed herein generate input features for a deep neural network that represent various characteristics associated with a session. By applying the deep neural network to the generated input features, the methods and systems described herein generate accurate disruption and delight predictions in addition to an attribution report detailing which of the characteristics represented among the input features had the greatest impact on the generated predictions. Various other methods, systems, and computer-readable media are also disclosed.