In the contemporary era of big data intelligence, knowledge has evolved into the fundamental driving force behind technological development and innovation. However, the sheer volume of data generated today is often fragmented, multi-source, and heterogeneous, existing in isolated silos that limit its utility. The primary significance of big data knowledge fusion lies in its ability to bridge these gaps. By systematically associating and integrating these fragmented data points, knowledge fusion constructs a unified, comprehensive knowledge system. This process is not merely a technical exercise in data aggregation; it serves as a critical support mechanism for enhancing knowledge completeness. Without effective fusion, AI systems operate on partial truths. By unifying data, it enables a transition from simple information processing to advanced cognitive intelligence. This allows machines not just to perceive data, but to understand relationships, reason through complexities, and maintain a holistic view of the world. Consequently, knowledge fusion acts as the bedrock for the next generation of intelligent systems, ensuring that decision-making is based on a reliable, cohesive, and complete representation of reality.

The research landscape of knowledge fusion has advanced significantly, evolving from basic methodologies to complex, high-dimensional integration suitable for the big data era. This progress can be categorized into foundational links and emerging frontiers:

1. Foundational Links (CLASSIC METHods):

The systematic review identifies three core pillars that constitute the traditional framework of knowledge fusion:

Knowledge Representation Learning: This involves transforming symbolic knowledge into computable vector spaces, allowing machines to process complex relationships mathematically.

Knowledge Alignment Matching: This link focuses on identifying and linking equivalent entities or concepts across different, disparate datasets.

Knowledge Conflict Resolution: When multi-source data disagrees (e.g., conflicting birth dates for the same person), this process utilizes truth discovery algorithms to resolve contradictions and ensure the veracity of the fused knowledge.

2. New Progress in the Context of Big Data:

Beyond these basics, the field has made deep explorations into frontier directions to handle the complexity of modern data:

Temporal Knowledge Fusion: Recognizing that facts change over time, recent research focuses on incorporating time-series data to capture the evolution of knowledge, rather than treating it as static.

Cross-Modal Knowledge Fusion: With the rise of multimedia data, progress has been made in fusing text, images, and video, breaking down the barriers between different sensory modalities to create a comprehensive understanding.

LLMs (large language models) Knowledge Fusion: Leveraging the generative and reasoning capabilities of LLMs (like GPT series), researchers are developing new methods to automate knowledge extraction and integration, utilizing the parametric knowledge embedded within these massive models.

The panoramic review provides a crucial roadmap for the theoretical expansion and technological evolution of big data knowledge fusion. It concludes that while traditional methods provide a necessary scaffolding, the future of the field requires addressing increasingly complex, dynamic, and abstract challenges.

Looking forward, two primary research trends highlight the future direction of the field:

The Fusion of Symbolic and Parametric Knowledge: This represents the convergence of explicit, structured knowledge (Knowledge Graphs) with the implicit, probabilistic knowledge found in deep neural networks (LLMs). This hybrid approach aims to combine the interpretability and accuracy of symbolic systems with the generalization and flexibility of parametric models, effectively bridging the gap between "logical" and "intuitive" AI.

Cross-Modal Temporal Knowledge Fusion: Future research will likely focus on the intersection of modality and time. It is no longer sufficient to fuse static images and text; systems must understand how multi-modal relationships evolve dynamically. This is essential for applications like digital twins and autonomous agents that must navigate a changing, multi-sensory world.

In summary, the field is moving toward a more integrated, dynamic, and cognitive-aware methodology, providing the essential guidance needed to build the intelligent infrastructures of the future.