RecommendationGraph

Context : TheAlgorithm, KnowledgeGraph

https://towardsdatascience.com/introduction-to-knowledge-graph-based-recommender-systems-34254efd1960

AutoGraph (Using LanguageModels to make recommendation graphs :

• https://arxiv.org/abs/2412.18241 via
• https://www.marktechpost.com/2025/01/05/autograph-an-automatic-graph-construction-framework-based-on-llms-for-recommendation

Compare : RecommendationArchitecture

Here's NotebookLM's comparison :

Recommendation Graphs

Recommendation Graphs are a modern approach to recommendation systems that uses graph neural networks (GNNs) to analyse relationships between users and items. These relationships can be simple, like a user clicking on an item, or more complex, like the semantic similarity between items. By representing these relationships in a graph structure, GNNs can effectively learn patterns and make accurate recommendations.

However, most GNN-based recommendation methods focus on optimising the model structure and learning strategies based on pre-defined graphs. They often overlook the importance of the graph construction stage [1, 2]. Earlier works for graph construction relied on specific rules or crowdsourcing, which are either too simplistic or too labour-intensive [3]. More recently, researchers are starting to use large language models (LLMs) to automate graph construction, leveraging their vast knowledge and reasoning capabilities [4, 5].

A key challenge in using LLMs for recommendation graphs is incorporating a global view of the data. LLMs have a limited context window, making it difficult to consider all relevant information when inferring relationships between nodes [6]. This often leads to inferior graph structures. Another challenge is the computational cost of using LLMs to infer relationships between millions or billions of user-item pairs [6].

Coward's Recommendation Architecture

Coward's Recommendation Architecture, proposed in the early 2000s, is a cognitive architecture designed to handle complex functionalities by learning from experience [7]. It emphasises a hierarchical and modular structure inspired by the human brain [7]. The architecture is divided into two main subsystems:

• Clustering: This subsystem forms a hierarchy of modules that detect and record partially ambiguous repetitions in the system's inputs. These repetitions are analogous to features or patterns but are not explicitly defined as such [8, 9].
• Competition: This subsystem uses consequence feedback to learn to interpret the outputs of the clustering subsystem into system behaviours. It selects the most appropriate behaviour based on the perceived consequences of previous actions [8, 10].

The key idea behind Coward's architecture is to manage the complexity of learning by using ambiguous representations within the clustering subsystem and resolving this ambiguity through the competition subsystem. This approach is inspired by the need for systems to be constructible, repairable, and modifiable, which requires a modular and hierarchical design [11, 12]. Comparison

While Recommendation Graphs and Coward's Recommendation Architecture are from different eras and domains, there are some interesting parallels:

• Hierarchical and Modular Structure: Both approaches emphasize the importance of a hierarchical and modular structure for handling complex functionalities. This is evident in the layered structure of GNNs and the hierarchical clustering in Coward's architecture.
• Focus on Relationships: Both approaches focus on capturing and exploiting relationships between elements in the data. This is central to the concept of a graph in Recommendation Graphs and the detection of repetitions in Coward's architecture.
• Learning from Experience: Both approaches are fundamentally about learning from experience. Recommendation Graphs use GNNs to learn patterns from interaction data, while Coward's architecture uses consequence feedback to refine its behaviour.

However, there are also significant differences:

• Data Representation: Recommendation Graphs explicitly represent relationships in a graph structure, while Coward's architecture uses a more abstract notion of partially ambiguous repetitions.
• Ambiguity and Context: Coward's architecture explicitly addresses the challenges of ambiguity and context by using ambiguous representations within the clustering subsystem and resolving this ambiguity through the competition subsystem. This is not a primary focus in Recommendation Graphs, which rely on the ability of GNNs to learn from potentially noisy and ambiguous data.
• Role of LLMs: Modern Recommendation Graphs are incorporating LLMs to enhance their ability to capture and exploit semantic relationships between items. This is not a feature of Coward's architecture, which predates the rise of LLMs.

In essence, Recommendation Graphs represent a modern, data-driven approach to recommendation systems, leveraging the power of graph neural networks and large language models. Coward's Recommendation Architecture, on the other hand, offers a more theoretical and cognitive perspective on how systems can learn complex functionalities by managing ambiguity and context.

It's interesting to see how these two approaches, separated by time and domain, converge on certain principles like hierarchical structure and the importance of relationships, while also highlighting the evolution of technology and thinking in the field of artificial intelligence.