Dynamic Utility Embeddings for Multi-Label Model Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learned models struggle to reliably predict various types of user interactions with the same result or recommendation, such as clicks, shares, and likes, due to the processor-intensive nature of extending two-tower deep neural networks to handle multiple interaction types.
Innovation Solution
A modified two-tower deep neural network model that regresses multiple outputs for different interaction types, sharing most parameters across label types and using a utility function to concatenate embeddings, allowing for efficient approximate k-nearest neighbor searches without retraining or reindexing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If separate machine learned models are used for each interaction type, then prediction reliability for multiple interaction types improves, but computational overhead and device complexity increase significantly
Solution Approach 1:
The patent combines multiple separate machine learned models into a single unified model that predicts multiple interaction types simultaneously. The model uses shared embedding layers for queries and items, with separate output heads for different interaction types (clicks, shares, likes), allowing one model to replace multiple separate models while maintaining prediction reliability across all interaction types
Solution Approach 2:
The unified model achieves multi-functionality by handling multiple interaction type predictions within a single model architecture. The model processes the same query-item pairs and generates predictions for different interaction types through shared feature extraction layers, enabling one model to perform the work of multiple specialized models
2Reliability
If separate machine learned models are used for each interaction type, then prediction reliability for multiple interaction types improves, but processing speed and productivity decrease due to multiple inference passes
Solution Approach 1:
By merging multiple models into one unified model, the system performs all interaction type predictions in a single inference pass rather than requiring separate inference passes for each model. This combines the reliability of multiple specialized models while achieving the processing speed of a single model
Solution Approach 2:
The unified model enables continuous prediction across all interaction types in one uninterrupted inference process. The shared embedding layers generate features once, and all interaction type predictions are generated continuously from these shared features without restarting the inference process
3Reliability
If separate machine learned models are used for each interaction type, then prediction reliability for multiple interaction types improves, but energy consumption increases due to multiple model inferences
Solution Approach 1:
The patent merges multiple energy-consuming model inferences into a single unified model inference. By sharing computation across different interaction type predictions, the system maintains the predictive reliability of multiple models while significantly reducing the total computational energy required
4Device complexity
If a single machine learned model predicts multiple interaction types, then device complexity and computational overhead decrease, but measurement precision and reliability of predictions worsen
Solution Approach 1:
The unified model applies segmentation by dividing the prediction task into distinct output heads for different interaction types, while maintaining shared feature extraction layers. This segmentation allows the model to maintain specialized prediction capabilities for each interaction type while benefiting from shared computational resources
Solution Approach 2:
The model implements local quality by having specialized output layers for each interaction type that are optimized for their specific prediction tasks, while sharing the general feature extraction capabilities. Each interaction type prediction maintains its own precision through dedicated output processing
Data Source
AI summary
In an example embodiment, a technique is presented that accesses training data that includes information about items, queries for items, and labels for the combinations of items and queries. The labels may correspond to different events, and there may be multiple different labels for the same combination of item and query. A machine learned model is then trained to learn a function for embedding each item to which a label pertains and a function for embedding each query to which a label pertains. Then, for each item in the training data, the items are embedded using the machine learned model, and the item embeddings for the item are concatenated into a single item embedding. At inference time, a similar concatenation is performed for multiple query embeddings. The concatenated embeddings are then used as input to an approximate k-nearest neighbor search function.