Cluster-Based Feature Generation for Sparse Labeled Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models face challenges due to the imbalance between the limited number of labeled samples and abundant unlabeled samples, leading to reduced generalization capabilities and inefficiencies in prediction generation.
Innovation Solution
A method that creates a cluster model comprising both labeled and unlabeled samples, identifies cluster information, and adds new features derived from clustering, forming an expanded feature space for training data sets to enhance the training of machine learning models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If supervised machine learning is used with limited labeled samples, then the model can be trained with clear guidance, but the generalization capability is reduced due to insufficient training data
Solution Approach 1:
The patent combines supervised learning (using labeled samples) with unsupervised learning (using unlabeled samples) into a hybrid training approach. The system creates clusters from unlabeled samples and uses these cluster assignments as additional training signals, effectively merging two learning paradigms to overcome the limitation of scarce labeled data while improving generalization capability.
Solution Approach 2:
The patent introduces cluster assignments as an intermediary mechanism that bridges labeled and unlabeled data. By computing clusters from unlabeled samples and using these cluster labels as intermediate training targets, the system enables the model to learn from unlabeled data in a supervised manner, thereby expanding the effective training dataset without requiring additional manual labeling.
2Reliability
If only labeled samples are used for training, then the training process is straightforward, but the feature space is limited reducing model performance
Solution Approach 1:
The patent adds a new dimension to the training process by incorporating cluster assignment features alongside the original input features. This creates an extended feature space where each sample is represented by both its original features and its cluster membership, allowing the model to leverage structural information from unlabeled data without complicating the fundamental training architecture.
3Reliability
If unlabeled samples are incorporated into training, then the feature space is expanded improving generalization, but the training complexity increases
Solution Approach 1:
The patent performs preliminary clustering of unlabeled samples before the main training process. By pre-computing cluster assignments and using these as fixed training targets, the system separates the complex unsupervised learning task from the supervised training, reducing overall training complexity while still benefiting from the expanded feature space.
Data Source
AI summary
A computer implemented method for machine learning model training. A number of processor units creates a cluster model comprising labeled samples and unlabeled samples. The number of processor units identifies cluster information for the labeled samples from the cluster model. The number of processor units adds a set of new features to a set of original features for the labeled samples using the cluster information to form an extended set of features for the labeled samples, wherein the labeled samples with the set of original features and the set of new features form a training data set for training a machine learning model.