Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LiDDA: Data Driven Attribution at LinkedIn

John Bencina, Erkut Aykutlug, Yue Chen, Zerui Zhang, Stephanie Sorenson, Shao Tang, Changshuai Wei

TL;DR

LiDDA presents a transformer-based data-driven attribution framework deployed at LinkedIn that unifies member-level signals with aggregate MMM factors. The approach uses a self-attentive, time-aware architecture with positional encodings, privacy-preserving imputation of external touchpoints, and training-time calibration to align with macro-level MMM outputs. Extensive offline and online validations demonstrate high predictive power, attribution stability, and alignment with experimental lift, and the system is applied to both GTM marketing and LinkedIn Ads measurement. The work provides practical deployment lessons and a scalable path for integrating DDA with MMM in large-scale ad-tech environments.

Abstract

Data Driven Attribution, which assigns conversion credits to marketing interactions based on causal patterns learned from data, is the foundation of modern marketing intelligence and vital to any marketing businesses and advertising platform. In this paper, we introduce a unified transformer-based attribution approach that can handle member-level data, aggregate-level data, and integration of external macro factors. We detail the large scale implementation of the approach at LinkedIn, showcasing significant impact. We also share learning and insights that are broadly applicable to the marketing and ad tech fields.

LiDDA: Data Driven Attribution at LinkedIn

TL;DR

LiDDA presents a transformer-based data-driven attribution framework deployed at LinkedIn that unifies member-level signals with aggregate MMM factors. The approach uses a self-attentive, time-aware architecture with positional encodings, privacy-preserving imputation of external touchpoints, and training-time calibration to align with macro-level MMM outputs. Extensive offline and online validations demonstrate high predictive power, attribution stability, and alignment with experimental lift, and the system is applied to both GTM marketing and LinkedIn Ads measurement. The work provides practical deployment lessons and a scalable path for integrating DDA with MMM in large-scale ad-tech environments.

Abstract

Data Driven Attribution, which assigns conversion credits to marketing interactions based on causal patterns learned from data, is the foundation of modern marketing intelligence and vital to any marketing businesses and advertising platform. In this paper, we introduce a unified transformer-based attribution approach that can handle member-level data, aggregate-level data, and integration of external macro factors. We detail the large scale implementation of the approach at LinkedIn, showcasing significant impact. We also share learning and insights that are broadly applicable to the marketing and ad tech fields.
Paper Structure (55 sections, 35 equations, 13 figures, 5 tables, 5 algorithms)

This paper contains 55 sections, 35 equations, 13 figures, 5 tables, 5 algorithms.

Table of Contents

  1. Introduction
  2. LinkedIn Use Case
  3. Overview of Related Work
  4. Media-Mix Modeling
  5. Rules Based Attribution
  6. Data-Driven Attribution
  7. Exploration on Existing Approaches
  8. Our contributions
  9. LiDDA System
  10. System Overview
  11. Path Creation
  12. Sequence Representation
  13. Touchpoint Representation
  14. Offline Processing
  15. Positional Representations
  16. ...and 40 more sections

Figures (13)

  • Figure 1: High level diagram showing the end-to-end components of our attribution modeling framework
  • Figure 2: We group touchpoints from some owned channels, e.g., Linkedin Platform, within a session into a single representative touchpoint. The attribution given to this one touchpoint is distributed evenly among the grouped campaigns.
  • Figure 3: Distribution of bias in the bootstrap means of the action weights, calculated as the difference between each bootstrap mean and the actual sample mean. The bias distributions center around 0.
  • Figure 4: The online validation refers to the comparison between measured lift from experiment data and the simulated lift output by LiDDA and the aforementioned calibration
  • Figure 5: Attribution weights for LinkedIn lead conversions show a time decay effect in aggregate; ad touches closer to the conversion tend to get more credit.
  • ...and 8 more figures