PREDICTING ANIMAL EMOTIONS USING ANIMAL EMOTION KNOWLEDGE GRAPH AND GRAPH NEURAL NETWORK

Smart overview of the Invention

A novel method and system are introduced for predicting animal emotions using an animal emotion knowledge graph (AEKG) alongside a graph neural network-transformer model. Unlike existing methods that rely heavily on visual and language data, this approach integrates a broader range of inputs, including neurobiological data and behavioral studies. Real-time graphs are generated from multimodal inputs such as audio, video, physiological, and environmental data, enabling the prediction of complex emotional states in animals. This system aims to enhance the understanding of animal emotions, thereby improving their care and management.

Technological Background

Traditional animal emotion detection technologies have limitations, primarily focusing on basic emotions and relying on single data sources, which reduces accuracy. The scarcity of generalized datasets and the absence of standardized models further complicate the interpretation of animal emotions. Current methodologies also face challenges in real-time analysis and the provision of actionable recommendations. The new method addresses these issues by using a dynamic, adaptive model that can incorporate new data relationships and update its representations in real-time.

System Components

The system comprises a memory storing instructions, communication interfaces, and hardware processors configured to generate real-time graphs represented as an AEKG. These graphs are derived from multimodal input data captured by sensors, including audio, video, physiological, and environmental data. The trained graph neural network-transformer model determines primary, secondary, and tertiary emotions from these graphs. Additionally, an emotion intensity function calculates the intensity of these emotions, which aids in predicting future emotional states using temporal analysis.

Model Training and Functionality

The GNN-transformer model is trained using a labeled set of scenario-specific subgraphs extracted from the AEKG. These subgraphs represent species-specific, context-specific, and rare emotional scenarios. The model's GNN layers learn relationships between features, while transformer layers capture temporal dependencies. A fusion layer combines these learnings to predict emotions in real-time. The model's multi-level detection process identifies emotions using subsets of features from different data types, ensuring a comprehensive understanding of animal emotional states.

Real-Time Recommendations

The system generates real-time recommendations using a large language model that queries a recommendation database based on the determined emotions, predicted future emotions, emotion intensity, and environmental context. This approach facilitates informed decision-making regarding animal care and management, supporting applications in various fields such as pet care, livestock management, and wildlife conservation. By providing actionable insights, the system enhances human-animal interactions and contributes to the welfare and well-being of animals.