Causal Dilated Convolutional Embedding for Customer Behavior Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional data storage systems fail to acquire and analyze user interactions with sufficient granularity, limiting their ability to predict customer behavior effectively in online environments.
Innovation Solution
A system that collects and stores detailed user interactions, builds unique models for each client based on user types and event activities, and uses an embedding component, causal dilated convolutional elements, and dense neural networks to predict customer behavior by generating a distribution of times for desired events and likelihood estimations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional data storage systems track and monitor user interactions, then user behavior data is collected, but the data granularity is insufficient to properly analyze user interactions
Solution Approach 1:
The patent segments user interaction data into multiple hierarchical levels including session-level interactions, event-level details, and attribute-level information. This segmentation allows the system to capture fine-grained data details while organizing them in a structured manner that enables comprehensive analysis without information loss.
Solution Approach 2:
The patent introduces additional dimensions for data capture by incorporating temporal sequences, spatial contexts, and multi-modal interaction types. This dimensional expansion enables the system to record user interactions with sufficient granularity across multiple axes, transforming insufficient 1D data into rich multi-dimensional data structures.
2Reliability
If detailed user interaction data is collected and stored, then prediction accuracy improves, but system complexity increases
Solution Approach 1:
The predictive system is segmented into distinct functional modules including data collection components, processing engines, modeling modules, and output generators. Each module handles specific aspects of the prediction pipeline, reducing overall system complexity while maintaining high prediction accuracy through specialized processing at each stage.
Solution Approach 2:
The patent introduces intermediary data structures and processing layers that bridge raw detailed user interaction data and prediction models. These intermediaries include aggregated event streams, feature vectors, and contextual representations that simplify the transformation process while preserving critical information for accurate predictions.
3Quantity of substance
If conventional systems monitor user interactions, then some behavior data is obtained, but sufficient data at proper granularity is not acquired
Solution Approach 1:
The system segments data collection into multiple layers capturing different quantities and granularities simultaneously. Event-level segmentation captures fine-grained interaction details, while session-level segmentation aggregates broader behavioral patterns, ensuring both sufficient quantity and proper granularity are achieved across different data dimensions.
Solution Approach 2:
The patent merges multiple data sources and interaction types into a unified detailed event stream. This combination consolidates quantitative data from various channels while maintaining the granular precision of individual interaction events, achieving both data quantity and measurement precision through integrated collection.
Data Source
AI summary
Various implementations of the invention for predicting customer behavior are described. Various implementations of the invention comprise an embedding component configured to receive and embed sequential inputs regarding a plurality of customer interactions with an online presence of a client; a plurality of causal dilated convolutional “CDC” elements configured to receive the embedded sequential inputs and to output a feature vector, where each CDC element comprises two causal dilated convolutions with regularization that is bypassed with a skip connection; a plurality of dense neural network elements configured to receive the feature vector and non-sequential inputs regarding a plurality of other customer interactions with the client, where each of the plurality of dense neural network elements comprises two dense neural networks with regularization that is bypassed with a skip connection; and an output generator configured to receive the output from the plurality of dense neural network elements and to generate a distribution of times over which a particular customer event will occur and/or a likelihood estimation that the particular customer event will occur within a particular time period.