Machine Learning Model for Entity Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods face challenges in determining the similarity of entities across various features due to the large number of possible features and combinations, leading to computational inefficiencies and the need for extensive processing resources, as well as the lack of available data for training models.
Innovation Solution
A machine-learning model is trained to encode selected feature values as vector representations, allowing for the comparison of entities and enabling self-supervised learning by using noise models to distort feature values, thus eliminating the need for labeled datasets and reducing the number of models required.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If individual models are created for each feature and combination of features to determine entity similarity, then the accuracy of similarity determination is improved, but the device complexity and processing resources required increase significantly
Solution Approach 1:
The patent applies universality by creating a single machine learning model that can handle multiple features and feature combinations simultaneously. Instead of training separate models for each feature pair, one universal model learns to process various feature combinations (e.g., image and text, text and audio) through self-supervised learning, reducing the number of models from potentially dozens to just one while maintaining similarity determination accuracy.
Solution Approach 2:
The patent employs parameter changes by dynamically selecting which features to use for similarity determination based on the specific input data. The system can adaptively choose relevant features (image, text, audio, or combinations) without requiring pre-configured models for each possible combination. This parameter-based approach allows the single model to efficiently handle variable feature sets.
2Device complexity
If a single machine-learning model is trained to handle all feature combinations for entity similarity, then the processing resources and memory required are reduced, but the model training becomes infeasible due to lack of labeled data for all feature combinations
Solution Approach 1:
The patent applies self-service through self-supervised learning, where the model generates its own training data by creating synthetic feature combinations from existing single-feature data. The system automatically synthesizes training examples (e.g., combining image features with text features) without requiring manual labeling or pre-existing multi-feature labeled datasets, enabling the model to train itself on all feature combinations.
Solution Approach 2:
The patent uses preliminary action by pre-processing existing single-feature data to prepare it for multi-feature model training. The system extracts and stores individual feature representations (image embeddings, text embeddings, audio embeddings) in advance, which are then combined during the self-supervised training phase to create comprehensive training examples for the universal model.
3Adaptability or versatility
If multiple individual models are maintained for different feature combinations, then the similarity assessment can be customized for different features, but the processing time and memory usage increase significantly
Solution Approach 1:
The patent applies universality by designing a single model that natively supports multiple feature types and combinations. The model accepts variable input configurations (image only, text only, audio only, or any combination) and processes them efficiently in one pass, eliminating the need to sequentially query multiple specialized models. This universal architecture maintains feature selection flexibility while reducing processing time.
4Measurement precision
If previously collected and categorized data is used to train models for entity similarity, then the training process can be supervised and accurate, but such data is often unavailable for all possible feature combinations
Solution Approach 1:
The patent applies self-service by implementing self-supervised learning, where the model creates its own training data from available single-feature examples. Instead of requiring pre-collected labeled multi-feature data, the system automatically generates synthetic training pairs by combining existing feature representations, thereby overcoming the scarcity of labeled training data for all feature combinations.
Solution Approach 2:
The patent uses an intermediary approach by introducing a feature combination synthesis mechanism that bridges the gap between available single-feature data and required multi-feature training examples. This intermediary process automatically generates plausible feature combinations (e.g., pairing image features with related text features) to create sufficient training data without manual annotation.
Data Source
AI summary
There are many situations in which it is useful to identify similar entities or to quantify how similar one entity is to another. However, similarity is often subjective, making it challenging to determine the similarity of two entities computationally. In order to allow for determining the similarity of entities according to a customizable definition of similarity, a selection of features for determining similarity is received from a user device and a machine-learning model is trained to encode the values of the selected features for a particular entity as a representation. The representations of two different entities may be compared to determine the similarity of the two entities according to the selected features. The trained machine-learning model may, additionally or alternatively, be used for cluster analysis to identify similar groups of entities.