Multi-Touch Attribution Using Temporal Convolution for Causal Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional attribution systems struggle to accurately track and analyze long-term and non-linear causal relationships in data, leading to inaccurate and unreliable attributions.
Innovation Solution
A data-driven approach using a temporal convolutional network (TCN) to generate predictive values for precursor events, combined with an attribution component and logistic regression, to compute accurate attribution values that account for sequential and long-term causal dependencies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional attribution systems are used, then the system complexity remains low, but the measurement precision of causal relationships deteriorates
Solution Approach 1:
The patent replaces conventional rule-based and statistical attribution methods with a deep learning-based temporal convolutional network (TCN) system. The TCN uses causal convolution operations to automatically learn and extract temporal patterns from event sequences, substituting manual attribution rules with automated neural network-based causal inference that can handle non-linear relationships and long-term dependencies.
Solution Approach 2:
The patent introduces an attention mechanism as an intermediary component between the TCN and the attribution calculation. This attention mechanism selectively weights different events in the sequence based on their relevance to the conversion, serving as a mediator that highlights causally significant events while filtering out noise, thereby improving attribution precision without requiring complete system redesign.
2Duration of action of moving object
If conventional attribution systems track long-term data, then the time coverage increases, but the reliability of attribution results deteriorates
Solution Approach 1:
The patent applies preliminary action by using the TCN to pre-process and extract relevant features from long-term event sequences before attribution calculation. The causal convolution operations in the TCN are designed to capture temporal dependencies while preventing information leakage from future to past events, ensuring that long-term patterns are properly encoded before being used for attribution, thus maintaining reliability across extended time periods.
Solution Approach 2:
The patent segments the long-term event sequence into manageable temporal windows or layers within the TCN architecture. By dividing the long sequence into smaller segments that can be processed independently and then combined, the system maintains computational efficiency and attribution reliability even when analyzing extended time periods with multiple user interactions and conversions.
3Adaptability or versatility
If conventional systems analyze non-linear causal relationships, then the adaptability improves, but the measurement precision deteriorates
Solution Approach 1:
The patent substitutes traditional linear statistical models with a non-linear deep learning architecture (TCN). The neural network automatically learns non-linear transformations of input events through multiple convolutional layers, enabling the system to adapt to complex non-linear causal relationships while maintaining high measurement precision through the differentiable nature of the network and its ability to capture intricate interaction patterns.
Data Source
AI summary
An apparatus and method for causal multi-touch attribution are described. One or more aspects of the apparatus and method include a time series component configured to generate an ordered series representing a plurality of precursor events corresponding to a result event, wherein each of the precursor events is associated with an event category from a set of event categories; a temporal convolution network configured to generate a series of predictive values corresponding to the plurality of precursor events by computing a plurality of hidden vector representations for at least one of the precursor events; and an attribution component configured to compute an attribution value for each of the event categories based on the series of predictive values.