Eco-aware GNN Recommenders

August 30, 2025·45 min·3 clips

The most eco-friendly AI model isn't always the least accurate — Antonio's results challenge the assumed accuracy-efficiency trade-off.

Training is where the cost shows up first. The episode opens with machine learning expense in plain terms: GPU-heavy training, cheaper inference, and CO2 measurement. Code Carbon comes in as the measuring tool. From there, the host asks the sharper recommender systems question: do graph neural network models always need larger embeddings? The guest starts with graph scale. Even a graph with roughly a hundred users can grow enough that the system needs lower dimensional embeddings to pull out relevant information. That makes embedding size the tradeoff. A larger embedding may keep more useful signal for a while, but it also increases computational complexity. That complexity has a real bill. In the guest's account, bigger embeddings usually mean longer training, and training time tracks closely with environmental impact. The host pushes for prediction instead of after the fact explanation. Could a heuristic or regression estimate the cost before training, including the weird cases where a run is much cheaper or more expensive than expected? That question keeps the conversation grounded. Sustainability here is not treated like a slogan. It is tied to variables a team can measure: training time, embedding size, and CO2 impact. The recommender setting gives the discussion its bite. Graph recommenders carry structure into a representation, so embedding choices affect both the information kept and the energy spent. The guest does not sell a universal rule. Larger representations can help, but the benefit has limits and needs measured results. The close turns practical, with publications and code for listeners who want to inspect the work, report bugs, or try follow-up ideas.

As heard by us

A practical look at when graph recommender models justify their energy cost, and when smaller choices may be smarter.

Data Skeptic puts recommender systems on an energy budget, using Eco-Aware Graph Neural Networks for Sustainable Recommendations to ask whether larger graph models are worth the compute they consume.

Read the full review in PlayNext →

Why you'd press play

Wondering how embedding size and training time turn into energy use and carbon?

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Data Skeptic