Masked Sub-Model Training with Local Gradients for Video Recommendations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The training of sub-models within recommendation systems for video delivery systems is challenging due to the complexity of the recommendation system, leading to inadequate training of parameters and issues like gradient attenuation, which affects the quality and speed of recommendation generation.
Innovation Solution
A masked model training process is employed to train sub-models by using outputs closer to the sub-models, masking certain inputs during the training phase to improve parameter adjustment and gradient updates, thereby enhancing the training effectiveness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If traditional model training is used in complex recommendation systems, then the system can process diverse inputs, but the sub-models suffer from inadequate parameter training and gradient attenuation
Solution Approach 1:
The patent divides the recommendation system into multiple sub-models, each responsible for processing specific types of inputs (e.g., user behavior sequences, item features). This segmentation allows each sub-model to be trained independently with targeted loss functions, improving parameter training quality while managing overall system complexity through modular architecture.
Solution Approach 2:
The patent introduces intermediate loss functions that act as mediators between the sub-models and the final recommendation output. These intermediate losses provide gradient feedback closer to each sub-model, preventing gradient attenuation and enabling better parameter adjustment without requiring complete retraining of the entire complex system.
2Measurement precision
If the recommendation system processes comprehensive user behavior data, then recommendation quality improves, but training time and computational resources increase
Solution Approach 1:
The patent applies preliminary actions by pre-processing user behavior sequences into fixed-length representations before feeding them to sub-models. This preprocessing step organizes comprehensive data into structured formats that can be efficiently trained, reducing computational overhead and training time while preserving the quality of recommendation signals.
Solution Approach 2:
The patent uses partial action by focusing training on specific subsequences of user behavior rather than requiring complete sequence processing for each training example. This allows the system to learn from representative portions of user behavior patterns, achieving good recommendation accuracy with reduced computational resources and faster training.
3Ease of operation
If sub-models are trained with outputs from the overall model, then gradient attenuation occurs, but training the sub-models with local outputs improves parameter adjustment
Solution Approach 1:
The patent implements feedback mechanisms where each sub-model receives gradient signals from its own local output and an intermediate loss function, rather than only from the final overall model output. This feedback loop preserves gradient signal strength and enables effective parameter adjustment by providing timely and relevant gradient information directly to each sub-model's parameters.
Data Source
AI summary
In some embodiments, a method receives a first sequence of inputs for processing via a sub-model of a plurality of sub-model. The plurality of sub-models are part of a main model. An input in the sequence of inputs is masked with a masked value to generate a second sequence of inputs. The method processes the second sequence of inputs using the sub-model to generate a sequence of features that correspond to the second sequence of inputs and processes the sequence of features to generate a first output. The first output is processed to generate a second output of the main model. The sub-model is trained based on a feature in the sequence of features that corresponds to the masked input and the second output.