Building Recommendation Systems Using Graph Neural Networks

Summary

Srihar Shah from Condonas Technology Labs presents on constructing recommendation systems using Graph Neural Networks (GNNs). He explains the basics of GNNs, their application within the company, and how they process graph data with nodes and edges. Shah delves into message passing, a core GNN concept, and its functions: message, aggregate, and update. He showcases a GNN example using the Deep Graph Library and discusses 'Recon', a project for personalized recommendations, utilizing heterogeneous graphs to represent user interactions. The talk concludes with a model architecture overview, highlighting a two-layer encoder and a simple decoder for predicting user-article links.

Takeaways

• 🧠 Sreeharsha Shah from Conde Nast Technology Labs discusses building recommendation systems using Graph Neural Networks (GNNs).
• 🌐 Graph Machine Learning is a branch of machine learning that deals with graph data, which consists of nodes and edges, and can include feature vectors.
• 🔍 Graphs can be bi-directional, directed, have self-edges, and be classified as homogeneous or heterogeneous, and even multi-graphs.
• 🤖 GNNs are a type of neural network that learn embeddings for nodes, edges, or entire graphs by considering the neighborhood of a node.
• 📨 The concept of 'message passing' is central to GNNs, where nodes collect messages from neighbors, aggregate them, and update their representation.
• 🔢 Mathematically, GNNs involve message functions, aggregate functions, and update functions to transform node representations.
• 🛠️ The Deep Graph Library (DGL) is used for implementing GNNs, providing APIs for message passing, aggregation, and updates.
• 🔄 'ReckON' is a machine learning project at Conde Nast that uses GNNs to improve user and content-based intelligence for personalized recommendations.
• 📊 The system represents user interactions as a heterogeneous graph and uses node features and relations to make recommendations, even for new users and articles.
• 📧 ReckON is used in a newsletter personalization use case, predicting which recently published articles a user might be interested in based on their past interactions.
• 🔄 The recommendation problem is posed as a link prediction problem, identifying potential future links between users and articles based on past interactions.

Q & A

What is the main topic of the talk by Sridhar Shah?

The main topic of the talk is building recommendation systems using Graph Neural Networks.

What is Graph Machine Learning according to the transcript?

Graph Machine Learning is a branch of machine learning that deals with graph data, which typically consists of nodes and edges, where nodes can be connected to each other and may have associated feature vectors.

What are the different types of graphs mentioned in the talk?

The different types of graphs mentioned are bi-directional, directed, graphs with self-edges, homogeneous, and heterogeneous graphs. A heterogeneous graph contains nodes and relations of different types, while a homogeneous graph has all nodes and relations of the same type.

What is a multi-graph as explained in the transcript?

A multi-graph is a graph that allows multiple edges between the same pair of nodes. An example given is the relationship between employees and locations, where an employee living in a location and an employee working at a location are considered as multiple relations.

What is the concept of message passing in Graph Neural Networks?

Message passing in Graph Neural Networks is a concept where nodes send messages to each other via edges, aggregate these messages, and update their representations based on the aggregated information from their neighbors.

What are the three key functions involved in message passing as described in the talk?

The three key functions involved in message passing are the message function, the aggregate function, and the update function. The message function defines how messages are created, the aggregate function determines how messages are combined, and the update function decides how the aggregated messages are used to update node representations.

How does the graph neural network model handle nodes with varying neighborhood sizes?

The graph neural network model handles nodes with varying neighborhood sizes by using an aggregation function that produces a fixed-length representation for each node's neighborhood, ensuring that the node's representation is consistent regardless of the number of neighbors.

What is the purpose of the 'update_all' function in the Deep Graph Library example?

The 'update_all' function in the Deep Graph Library example is used to apply message passing for the entire graph. It defines the message function, aggregation function, and update function to be applied to all nodes in the graph.

What is the RECON project mentioned in the transcript?

RECON, which stands for Recommendations Using Knowledge Networks, is a machine learning project aimed at improving user and content-based intelligence to provide better personalized recommendations and entity representations for downstream tasks.

How does the RECON project represent user demographics in the graph?

In the RECON project, user demographics are represented as separate entities in the graph rather than as node features. This design choice helps address the cold start problem, allowing the model to make recommendations for new users or users with no interactions by leveraging demographic information.

What is the scoring module's role in the RECON project?

The scoring module in the RECON project takes the graph, the trained model, and the candidates as inputs to predict the links between users and articles. It uses the graph for neural message passing to get representations and the trained model to make inferences, ultimately scoring each candidate pair to determine recommendations.