Neural Network for User Data Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting user data are limited as they either aggregate user behavior data separately from user actions or focus solely on the sequence of actions, leading to restricted model training capabilities due to isolated error backpropagation processes.
Innovation Solution
The proposed method trains neural network models using combined latent state vectors from both user action and feature data, applying error backpropagation simultaneously to enhance prediction capabilities by dynamically selecting architectures based on vector dimensions and feature distributions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If separate prediction models are used for user vectors and user actions, then each model can be trained independently, but the error backpropagation occurs in isolation which limits the set of metrics and data that can be correctly trained
Solution Approach 1:
The patent combines separate prediction models for user vectors and user actions into a single unified predictive model. This unified model processes both user feature data and user action sequence data through integrated neural network architectures, allowing error backpropagation to occur simultaneously across all components. The combination enables the model to learn synergistic patterns between user attributes and behavior sequences, thereby expanding the set of metrics that can be correctly trained while maintaining independent training capabilities through modular architecture design.
2Device complexity
If user behavior data is aggregated separately from user actions, then data processing becomes simpler, but the model's ability to predict user metrics is limited
Solution Approach 1:
The patent segments the input data into distinct components: user feature vectors (static attributes) and user action sequences (dynamic behaviors). Each segment is processed by specialized neural network modules designed for its specific characteristics. The user feature module processes static attributes through fully connected layers, while the user action module processes sequential data through recurrent or transformer architectures. The segmented processing maintains data integrity and enables targeted feature extraction, which is then integrated to produce accurate user metric predictions without overwhelming computational complexity.
3Loss of time
If only the sequence of user actions is considered, then temporal patterns are captured, but the ability to predict user metrics is similarly limited
Solution Approach 1:
The patent introduces an intermediary integration layer that combines the outputs of the user feature module and user action module. This intermediary component receives processed representations from both static user attributes and dynamic action sequences, then fuses them through attention mechanisms or feature concatenation. The intermediary layer enables temporal patterns from action sequences to interact with user profile features, creating a comprehensive user representation that enhances metric prediction accuracy while preserving the temporal information captured from action sequences.
Data Source
AI summary
The group of inventions relates to technologies for automated prediction of user data: actions and features using a predictive model. The technical result is to increase the efficiency of predicting user data. A method for automated prediction of user data is proposed. The method comprises the step of obtaining, by at least one processing unit, user action features, represented as an array of first vectors. Further, the method comprises obtaining user features represented as an array of second user feature vectors. Also, training neural network model on said arrays of first vectors and second vectors of features using an error backpropagation method to obtain trained model the output of which generates first and second latent state feature vectors, wherein said trainable neural network model is configured to dynamically select an architecture depending on said first and second feature vectors.