Stanford CS224W Lecture 19: Ranking 315K GNN Designs with Anchor Models

🌏 中文版

This is Lecture 19 of Stanford CS224W: Machine Learning with Graphs, Fall 2025, dated 2025-12-04. It follows the course schedule and the official Lecture 19 Conclusion deck; speaker attribution follows the slides.

Materials and gaps

Public materials include the official slides and optional readings on the schedule. Canvas video, live Q&A, board work, and Ed discussions are unavailable and are not reconstructed. The public 2021 videos are not evidence for a 2025 lecture.

Complete lecture agenda

The actual conclusion question

The official Lecture 19 deck studies GNN model selection over roughly 315K designs and 32 tasks. The best design varies by task, so one fixed leaderboard is insufficient.

Step 1: randomly evaluate 100 designs on a small dataset

The official procedure first samples and trains 100 random designs on a small dataset to obtain a performance spectrum. It does not predict rankings for unevaluated designs.

Step 2: uniformly select 12 anchors across the spectrum

It then selects 12 anchor models uniformly across that performance spectrum. Covering weak, middle, and strong performance makes the anchor set informative about task behavior.

Step 3: represent each task by its anchor ranking

The same anchors are evaluated on each task. Their relative ranking becomes a behavior-based task representation rather than a similarity guessed from dataset names.

Step 4: define task similarity from anchor rankings

Task similarity is determined by anchor-model rankings. Similar anchor orderings indicate a potentially useful source task.

Step 5: transfer the best designs from a similar task

The best designs already observed on the similar source task are transferred and evaluated on the target. The deck does not propose predicting a complete ranking of all unevaluated designs or a top-k ranking model.

Controlled random search as prior methodology

The deck first presents controlled random search (CRS) as an earlier design-space search method, then develops anchor-based task similarity and best-design transfer. CRS belongs to the methodological context; the 0.771 in the OGB result is explicitly labeled Previous SOTA and must not be attributed to CRS.

OGB benchmark results

In the deck's OGB example, transferring from a similar task reaches 0.785, compared with the deck's Previous SOTA of 0.771. Transfer from a dissimilar task falls to 0.736. The 0.771 result is not a CRS score.

Reading the three numbers together

The 0.785-versus-Previous-SOTA-0.771 comparison and the 0.736 dissimilar-task negative control are distinct evidence. Calling 0.771 a CRS result, or quoting only 0.785, would misstate the deck's mechanism.

Boundaries of 315K and 32

315K is the official design-space size and 32 is the task count. This article does not claim to have rerun that complete experiment or turn it into a production guarantee.

A faithful miniature reproduction

Randomly evaluate 100 designs, uniformly select 12 anchors over their performance range, run those anchors across tasks, choose a source by ranking similarity, transfer its best designs, and record both similar-task and dissimilar-task transfer results.

Artifacts required for reproduction

Save the sampled 100 designs, their performance, the 12 anchor positions, per-task anchor rankings, task similarities, transferred source designs, and both similar-task and dissimilar-task target results.

The final workflow

Use a small-dataset performance spectrum to select anchors, use anchor rankings to identify similar tasks, and transfer known best designs from those tasks. This is a budgeted model-transfer procedure, not a universal ranking predictor.

References

• CS224W Fall 2025 schedule
• Lecture 19 official slides