穆再排尔·穆合塔尔
← All work

04 2026 / 6th Jittor AI Challenge · Track 1

Cold Start

Link prediction when more than half the graph has never been seen.

Role
Core member
Period
2026
Stack
Jittor · graph learning · MRR · Val-Hot
Status
Competition GitHub
A sparse constellation of pale dots and faint lines on black, a few nodes marked in vermilion.
PL. 04 / Cold Start A sparse constellation of pale dots and faint lines on black, a few nodes marked in vermilion. 2026

The leaderboard said 0.9357. The task was temporal link prediction, and 54% of the candidate pairs involved a node with no history. A model that had learned the warm part of the graph could look competent on average and fail exactly where the task was hard.

The task

Given a temporal graph, predict whether an edge will appear between two nodes in the next window. The official baseline was a strong GNN. Our first job was not to beat it; it was to understand what the metric was rewarding.

Three traps

Temporal leakage. A random split lets the model see future edges while predicting past ones. The score goes up, the task gets easier, and the result stops meaning anything. We split by time, not by row.

Candidate-pool mismatch. The official evaluation samples negatives differently from the training distribution. A model tuned on training negatives can look better locally than it will on the leaderboard. Val-Hot rebuilds the candidate pool to match the evaluation protocol.

Cold-start blindness. Aggregate MRR hides the population that matters. We reported warm and cold subsets separately and optimized for cold even when the aggregate moved slowly.

Val-Hot

Val-Hot is a small, open local benchmark. It is not a leaderboard replica; it is a stable signal. The point is to rank two candidate changes correctly, so the team spends submissions on ideas that are likely to survive.

The benchmark became more valuable than any single feature. It turned arguments into experiments. When someone proposed a change, we could ask: does it improve cold MRR on Val-Hot, or only the average?

The heat feature

The largest single gain came from the least sophisticated idea: a zero-parameter heat feature. Count how often a node appears in the recent graph, normalize it, add it to the score. No training, no embedding, no learned weight.

It contributed +0.047 MRR. When cold nodes are the difficulty, recent frequency is a strong prior: a node that has been active lately is more likely to appear again, even if the model has never learned anything about it. The feature is not elegant. It is useful.

The result

The team moved from 0.9357 to 1.2443. The gain came from a better validation loop, a repaired evaluation, and a handful of simple features. The lesson is not that handcrafted features beat deep learning. The lesson is that on a noisy benchmark, the ability to measure is part of the model.

In most machine learning stories, the cold-start nodes are an edge case. Here they were the majority.

Measurements

+0.31 total MRR gain
54% cold-start candidates
+0.047 from heat feature
1.2443 final score

Next — 01 / 2026

Autonomous Landing

A three-tier vision pipeline that brings a drone down without GPS.