Learned Table Representations for Spreadsheet Formula Recommendation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Non-technical users face challenges in authoring complex formulas in spreadsheets due to the need to understand function syntax and parameters, leading to low accuracy in existing formula prediction methods.

Innovation Solution

A deep learning-based approach that represents spreadsheets and regions as images, using machine learning models to identify similar spreadsheets and regions, and adapt formulas from reference cells to the target cell, enhancing accuracy and efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If non-technical users attempt to author complex formulas manually, then they can create custom calculations, but they face challenges due to needing to understand function syntax and parameters, resulting in low accuracy

Engineering Contradiction:
Improveformula prediction accuracyVSAvoiduser effort to understand syntax
Core Design Contradiction:
Measurement precisionVSEase of operation

Solution Approach 1:

The system enables formulas to recommend themselves automatically by analyzing spreadsheet structure, data types, and contextual patterns. The formula recommendation engine autonomously identifies appropriate formulas without requiring users to manually search or understand complex syntax, with users simply needing to review and accept suggestions.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent replaces the mechanical process of manual formula construction with an AI-based recommendation system that uses machine learning models to predict appropriate formulas. This substitution transforms the workflow from active formula authoring to passive formula selection, significantly reducing the cognitive load and syntax knowledge required.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

If existing formula prediction methods are used, then some formula suggestions can be provided, but accuracy remains low due to limited training examples and inability to capture spreadsheet context

Engineering Contradiction:
Improveformula prediction accuracyVSAvoidtraining examples
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent transforms spreadsheet data into multi-dimensional vector representations that capture structural, contextual, and semantic information across multiple dimensions. This dimensional transformation enables the model to process and learn from complex spreadsheet patterns that traditional flat training approaches cannot capture, significantly improving prediction accuracy with the same training data volume.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The system performs preliminary analysis of spreadsheet structure, data types, and contextual relationships before formula recommendation. By pre-processing and encoding spreadsheet characteristics into comprehensive feature vectors, the system prepares rich training examples that capture essential contextual information, enabling more accurate predictions without requiring additional training data.

Inventive Principle:
Principle #10Preliminary action

3Productivity

If manual formula creation is required, then precise calculations can be achieved, but productivity decreases due to time-consuming syntax lookup and parameter understanding

Engineering Contradiction:
Improveformula creation speedVSAvoidtime for syntax lookup
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The formula recommendation system automatically analyzes the current spreadsheet context and generates relevant formula suggestions without requiring user intervention for syntax lookup or parameter research. Users simply review the AI-generated suggestions and select the appropriate one, dramatically reducing the time investment required for formula creation.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system pre-computes and caches formula recommendations based on spreadsheet patterns and contextual analysis. When users need formulas, pre-generated suggestions are immediately available, eliminating the need for real-time syntax lookup and parameter research, thus significantly improving formula creation speed.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4614385A1Recommended formulas in spreadsheets using learned table representations
Publication Date: 2025.09.10 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP4614385A1 patent drawingFigure 1
  • EP4614385A1 patent drawingFigure 2
  • EP4614385A1 patent drawingFigure 3A

AI summary

The methods and systems provided herein automatically identify similar spreadsheets to target spreadsheets. The methods and systems automatically predict formulas that users want to author in a target cell of a target spreadsheet by identifying a reference formula from a similar region in a similar reference sheet that is similar to the target cell. The methods and systems generate a predicted formula by adapting parameters of a reference formula to a context of the target cell. The methods and systems provide an output with the predicted formula in the target cell of the target spreadsheet.