Feature Set Embedding for Incomplete Data Classification

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Classification techniques face challenges when dealing with incomplete data due to missing or conflicting features, as existing methods often require modeling the missing feature distribution and considering instances as vectors from a given feature space, which can be inflexible and inefficient.

Innovation Solution

A method and system that generate pairs of features and values from incomplete data, apply a transformation function to map these pairs into an embedding space, and use a prediction function to generate confidence assessments for classification, allowing for the classification of incomplete data without explicitly handling missing feature distributions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If existing classification methods are used to handle incomplete data, then the classification can be performed, but the methods require modeling the missing feature distribution and considering instances as vectors from a given feature space, which reduces flexibility and efficiency

Engineering Contradiction:

Improveflexibility in handling incomplete dataVSAvoidcomplexity of modeling missing feature distribution

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent segments the feature space by treating each feature independently rather than as a complete vector. Missing features are handled individually through separate embedding layers, allowing the system to process incomplete data without requiring a complete feature vector. This segmentation approach eliminates the need to model the joint distribution of missing features while maintaining classification effectiveness.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent transforms the classification problem from operating in the original feature space to operating in an embedding space. By mapping features to embedding vectors and aggregating them through summation, the system changes the dimensional representation and operational framework, allowing flexible handling of incomplete data without constraints from the original feature space geometry or distribution assumptions.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

2Productivity

If traditional classification approaches are applied to incomplete data, then classification results can be obtained, but the approaches require full model re-training when new features are introduced

Engineering Contradiction:

Improveefficiency of model adaptationVSAvoidtime required for full model re-training

Core Design Contradiction:

ProductivityVSLoss of time

Solution Approach 1:

The patent creates a universal classification framework where the aggregation function (summation of embedding vectors) and prediction layer remain constant regardless of which features are present. This universal structure allows the model to handle different feature subsets and incorporate new features without requiring complete re-training, as the core architecture remains applicable to any combination of features.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent pre-defines the aggregation mechanism (summing embedding vectors) and prediction structure in advance, creating a flexible framework that can accommodate new features. By establishing this preliminary structure, the system avoids the need for complex re-training when new features are introduced, as the framework is already designed to handle variable feature sets through its inherent flexibility.

Inventive Principle:

Principle #10Preliminary action

3Reliability

If complete feature vectors are required for classification, then traditional classification algorithms can be applied, but this requirement reduces the system's ability to handle real-world incomplete data effectively

Engineering Contradiction:

Improveeffectiveness in handling real-world incomplete dataVSAvoidsimplicity of data processing requirements

Core Design Contradiction:

ReliabilityVSEase of operation

Solution Approach 1:

Instead of requiring complete feature vectors and then handling missing data as an exception, the patent inverts the approach by designing the system to naturally accept and process incomplete feature sets from the start. The embedding and aggregation architecture is constructed to work with partial information, making completeness unnecessary and simplifying the operational requirements for real-world data processing.

Inventive Principle:

Principle #13The other way round (Inversion)

Data Source

PatentUS8706668B2Feature set embedding for incomplete data

Publication Date: 2014.04.22 NEC CORP

AI summary

Methods and systems for classifying incomplete data are disclosed. In accordance with one method, pairs of features and values are generated based upon feature measurements on the incomplete data. In addition, a transformation function is applied on the pairs of features and values to generate a set of vectors by mapping each of the pairs to a corresponding vector in an embedding space. Further, a hardware processor applies a prediction function to the set of vectors to generate at least one confidence assessment for at least one class that indicates whether the incomplete data is of the at least one class. The method further includes outputting the at least one confidence assessment.