User Representation Matching via Event Data Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The online travel industry faces challenges in analyzing user behavior due to sparse and dispersed data across multiple devices, accounts, and platforms, making it difficult to build accurate traveler profiles without personal identifiable information (PII) due to GDPR restrictions and the decline of third-party cookies.
Innovation Solution
A method using a data matching server that generates merged user representations based on event search data, determines pairwise features, and employs machine learning algorithms to identify users without PII, leveraging natural language processing and information fusion techniques to enhance traveler representation learning and improve click-through rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If third-party cookies and PII data are used for user identification, then user profiling accuracy is improved, but privacy protection and GDPR compliance deteriorate
Solution Approach 1:
The patent extracts and removes PII data and third-party cookies from the user identification process. Instead of using traditional identifiers, the system processes only anonymized event data such as search queries, booking patterns, and interaction behaviors. This extraction principle directly resolves the contradiction by eliminating privacy-harmful data elements while maintaining profiling functionality through alternative anonymized features.
Solution Approach 2:
The patent introduces an intermediary layer of anonymized event processing between raw user data and profiling algorithms. Event data undergoes transformation through natural language processing and feature extraction to create intermediate representations that preserve behavioral patterns without containing personally identifiable information. This intermediary mechanism enables accurate profiling while preventing direct access to sensitive user data.
2Adaptability or versatility
If user data is dispersed across multiple devices and platforms, then device versatility and user flexibility are improved, but data sparsity and matching difficulty worsen
Solution Approach 1:
The patent merges event data from multiple devices and platforms into unified user profiles through probabilistic matching algorithms. The system combines search events, booking events, and interaction events across different devices, accounts, and platforms to create consolidated user representations. This merging principle directly addresses data sparsity by aggregating dispersed information while preserving multi-device versatility.
Solution Approach 2:
The patent performs preliminary data enrichment and feature extraction before user matching occurs. Event data is pre-processed to extract meaningful features such as search patterns, booking preferences, and interaction behaviors. This preliminary action prepares the data in advance, making it easier to match users across devices despite data dispersion, thereby reducing the impact of sparsity while maintaining flexibility.
3Device complexity
If traditional advertising models are used, then implementation simplicity is improved, but click-through rates and advertising effectiveness deteriorate
Solution Approach 1:
The patent changes the parameters used for advertising targeting from demographic and PII-based features to behavior-based features derived from event data. The system uses search patterns, booking behaviors, and interaction metrics as targeting parameters instead of traditional demographic information. This parameter transformation maintains relative system simplicity while significantly improving advertising relevance and click-through rates by targeting users based on actual behavior rather than broad demographics.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Methods, systems, and computer program products for determining user representations based on matching. A matching request associated with a user is received. Event search data for a plurality of events for the plurality of users is obtained. A merged user representation for a plurality of candidates associated with the plurality of users is generated based on the event search data. A subset of candidates from the plurality of candidates is selected based on the merged user representation. Pairwise features are determined based on similarities between the subset of the candidates. A learned user representation is determined by identifying, using a machine learning algorithm, at least one user of the plurality of users from the subset of the candidates based on the pairwise features. The learned user representation associated with the at least one identified user of the plurality of users is provided.