Transformer Control Recommendation for Personalized Device Actions

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing action recommendation systems for controlling external electronic devices fail to accurately reflect complex user context correlations and personalized intentions, particularly in sequential and context-aware recommendations.

Innovation Solution

An electronic device equipped with a learning model that generates control recommendations using embedding vectors, transformer encodings, and position information, minimizing loss between training data and control recommendations, to provide context-aware and personalized suggestions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of information

If sequential recommendation is used to provide action recommendations based on user's past control history, then the recommendation system can leverage historical data, but it fails to reflect complex user context correlations

Engineering Contradiction:

Improveuser context correlationVSAvoidrecommendation system complexity

Core Design Contradiction:

Loss of informationVSDevice complexity

Solution Approach 1:

The patent implements a multi-layer encoder architecture where a first encoder processes basic control history sequences, and a second encoder further processes the outputs along with additional context information. This nested structure allows complex context correlations to be captured by layering processing stages, where each encoder builds upon the previous one's output while adding new contextual dimensions.

Inventive Principle:

Principle #7Nested doll (Nesting)

Solution Approach 2:

The patent introduces position encoding to transform the sequential data into a multi-dimensional representation space. By adding positional information as an additional dimension to the control history sequences, the model can capture not only what actions were taken but also their temporal order and contextual relationships, thereby reflecting complex user context correlations.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

2Adaptability or versatility

If context-aware recommendation is used to determine user action recommendation based on user context, then contextual information can be incorporated, but the context cannot be personalized to the user

Engineering Contradiction:

Improvecontext personalizationVSAvoidrecommendation accuracy

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The patent applies different processing mechanisms to different parts of the input data. The first encoder specifically processes control history sequences while the second encoder handles additional context information and position encodings. This localized processing allows the system to maintain specialized handling for personalization-relevant data while incorporating general context, thereby achieving both personalization and reliability.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent divides the recommendation system into distinct encoding modules: a first encoder for control history and a second encoder for additional context with position information. This segmentation allows each module to specialize in processing specific types of data, with the first encoder capturing user-specific patterns from control history and the second encoder integrating broader contextual information, thereby achieving personalized yet reliable recommendations.

Inventive Principle:

Principle #1Segmentation

3Ease of operation

If traditional recommendation methods are used, then implementation is simpler, but capricious user intentions cannot be reflected in the action recommendation

Engineering Contradiction:

Improveuser intention captureVSAvoidlearning model complexity

Core Design Contradiction:

Ease of operationVSDevice complexity

Solution Approach 1:

The patent applies pre-trained language model weights to the encoder components, particularly the second encoder that processes additional context information. This preliminary initialization with pre-trained weights enables the model to capture complex user intentions including capricious behaviors from the start, as the pre-trained weights already encode sophisticated pattern recognition capabilities that can handle unpredictable user intents.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent combines multiple types of encoded information into a unified representation: control history embeddings from the first encoder, additional context embeddings, and position encodings are all integrated in the second encoder. This composite approach allows the model to capture diverse aspects of user behavior including capricious intentions by merging different information sources that collectively represent complex user intents.

Inventive Principle:

Principle #40Composite materials

Data Source

PatentUS20240036527A1Electronic device and computer readable storage medium for control recommendation

Publication Date: 2024.02.01 SAMSUNG ELECTRONICS CO LTD

AI summary

An electronic device is provided. The electronic device includes an interface, a memory, and a processor and configured to provide a control recommendation of an external electronic device using a learning model. The learning model is configured to generate a first output vector by encoding sequential control information about a user using a transformer and summarizing the encoded sequential control information using a query vector, and output a second output vector by encoding the first output vector using a transformer and summarizing the encoded first output vector using time information.