Hashing-Based User Modeling for Dynamic Behavior
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep binary hashing techniques are not suitable for sequential behavior data, as they fail to capture the evolving preferences and dynamic nature of user behavior across various time scales, leading to poor performance in nearest neighbor search applications such as lookalike expansion and user segmentation.
Innovation Solution
The development of novel deep learning-based architectures, including Categorical Attention and Relation and Attention CNN (RACNN)/Time-Aware Attention CNN (TAACNN), which transform raw feature data into high-level metadata embeddings and represent user behavior in a 2D space at different time-window sizes, effectively capturing long-term and short-term preferences and behavior patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If deep binary hashing techniques are used for nearest neighbor search, then computational efficiency is improved, but accuracy deteriorates because they fail to capture evolving preferences and dynamic user behavior patterns
Solution Approach 1:
The patent segments user behavior data into multiple time windows (short-term, medium-term, long-term) and generates separate hash codes for each window. This segmentation allows the system to capture dynamic user preferences at different time scales while maintaining computational efficiency through parallel processing of segmented data.
Solution Approach 2:
The patent implements dynamic hashing by generating multiple hash codes that reflect user behavior at different time scales. The system dynamically adapts to evolving user preferences by weighting different time window hash codes differently, allowing the most recent user behavior to have greater influence on the final hash code while preserving historical patterns.
2Speed
If traditional hashing methods are used, then computational speed is improved, but the ability to capture contextual features and patterns in sequential behavior data deteriorates
Solution Approach 1:
The patent extends traditional hashing from a single static hash code to multiple hash codes organized in a temporal dimension. By adding the time window dimension, the system preserves contextual information about when behaviors occurred while maintaining the computational speed benefits of hashing through efficient bit-wise operations on each time window's hash code.
3Measurement precision
If multiple time-window hash codes are generated to capture dynamic behavior, then accuracy of user modeling is improved, but device complexity increases
Solution Approach 1:
The patent merges multiple hash codes from different time windows into a single comprehensive user representation by concatenating or weighting them. This merging strategy maintains high modeling accuracy by preserving information from all time scales while simplifying the system structure compared to maintaining completely separate hashing systems for each time window.
Data Source
AI summary
In one embodiment, a method includes receiving user behavior data and contextual information associated with the user behavior data, the contextual information including a first data portion associated with a first context type. The method includes generating, from the user behavior data and the contextual information using a hashing algorithm, a first heterogeneous hash code including a first portion representing the user behavior data and a second hash code portion representing the first data portion associated with the first context type. The method includes accessing a second heterogeneous hash code including a third hash code portion representing a second data portion associated with the first context type. The method includes comparing the first heterogeneous hash code with the second heterogeneous hash code including determining similarity between the second hash code portion of the first heterogeneous hash code and the third hash code portion of the second heterogenous hash code.