Huawei Ireland Tech Arena 2026
2nd overall at the onsite final, Dublin, November 2026
A two-phase machine learning competition run by Huawei Ireland Research Centre. In the online phase we predicted which recommended item a user would click. In the onsite final, our model had to compete live against every other team's for the same users. I was my team's technical lead.
- When
- October – November 2026
- Team
- Learn2RecSmartly: six first-year students and one second-year
- My role
- Technical lead
- Stack
- Python, LightGBM, PyTorch, NumPy, SciPy
- Result
- Phase 1: 12th, inside the top-15 cut. Final rounds: 8th, 8th, joint 6th, 5th. Overall after the jury presentations: 2nd of 15
The problem
A recommender shows a user a slate of ten items. Predict which one, if any, they click. Behind the data was a simulated user population, and submissions were scored by ROC AUC on a hidden test set: 30% of it decided the public leaderboard during the competition, and the other 70% decided the final ranking. The top 15 teams went through to an onsite final in Dublin.
Phase 1: click prediction
Validation first. Before modelling anything, we built a validation split that behaved like the hidden test set: train on everything up to a cutoff day and score each user's first events after it. Our early split ran about 0.006 AUC above the leaderboard, and only about two thirds of the gains we measured on taste features carried over to it. Knowing that number shaped every later decision.
User taste from geometry. The biggest gains came from how we represented taste rather than from the model on top:
- Item vectors started as a low-rank SVD of the click matrix, then were fine-tuned with a softmax over each slate, so they learned which item won among the ten actually shown.
- Projecting those vectors onto an 8-dimensional unit sphere (keep the top eight principal components and normalise the length) was the single largest jump on the leaderboard. Users turned out to be spread evenly over that sphere, with no archetypes, and their taste didn't drift over time.
- Skipped items carry information too. A user's taste was the mean of the items they clicked minus a shrunk mean of the items they were shown and skipped (a Rocchio-style update).
A Bayesian layer. On the sphere, we modelled each user with a von Mises–Fisher distribution and used its posterior predictive as features: how concentrated a user's taste is, how uncertain each item is, and a per-user "no click" utility fitted under the slate likelihood. Skips were weighted by their predicted click probability, which gives the exact likelihood gradient.
Models. A LightGBM ranker took those features, in bagged configurations across seeds and SVD sizes. Alongside it ran SlateNet, a small transformer over the whole slate, written in PyTorch. The submission was a rank blend of seven tree models plus SlateNet, and every uploaded file was rebuilt from a recipe with a manifest, so any score could be traced back to exactly what produced it.
The last day. Near the top, gains shrink to a few hundred-thousandths of AUC, smaller than the noise between leaderboard splits. So we treated each public score as a measurement:
- We estimated how much of a validation gain transferred to the leaderboard, by type of evidence.
- We measured the cost of adding pure noise to the blend, and only shipped changes that beat it by at least two times.
- Every change had its pass rule written down before we looked at the result.
- Every shipped file was rebuilt byte for byte by two or three independent runs.
| Stage | Main change | Public AUC |
|---|---|---|
| First models | SVD taste + LightGBM | 0.744 → 0.798 |
| Fine-tuned taste | Slate-softmax fine-tuning, skip-aware taste | 0.814 → 0.816 |
| Sphere projection | Taste vectors on a unit sphere | 0.821 |
| Bayesian features | von Mises–Fisher posterior, weighted skips | 0.823 → 0.824 |
| Blends | Rank blend of 7 models + SlateNet | 0.825 |
| Final day | Measured-transfer stack | 0.833 |
We were 17th on the public leaderboard (30% of the test set) and finished 12th on the private leaderboard (the other 70%), which decided the ranking. That put us inside the top-15 cut for the final.
Phase 2: the onsite final
The final changed the question. It was no longer "will this user click?" but "out of every team's offer, will they pick ours?" In each of four rounds over two days, every team's model offered one item to each user, all fifteen offers formed one shared slate, and the simulated user picked one item or none. We trained on a provided machine and submitted a packaged model, which ran once on new data under a strict time limit. AI tools were allowed on our laptops, but no autonomous agent could operate on the competition machine: a person typed every command there.
Reading the labels. We brought our Phase 1 model to round 1 and placed 8th. The released results showed why almost at once: users went back to items they had already clicked, and our model was busy predicting taste. From then on, our work between rounds was to read the labels, find the largest, best-supported pattern, test it, and ship it.
- Round 2: re-offering past clicks took our score from about 27 to about 48. We placed 8th again.
- Round 3: a new dataset about five times larger put the time limit at risk. We submitted a safe entry first and replaced it once something measured better.
- Round 4: the first round whose data included the game's own history. We scored each past click by how often the user picked it, how often they ignored it, and how recently. That rule had been tested blind on an earlier round and replicated independently three times.
The gain never came from a bigger model. On the last round's data, a gradient-boosted model trained on the same signals did slightly worse than the hand-written rule. After the four rounds, each team presented its approach to the jury, and the final ranking put us 2nd overall.
My role
I led the technical side: I set the direction, decided which evidence was strong enough to act on, made the final call on each submission, and typed the commands on the competition machine during the final. Much of the execution was done by AI coding agents (Claude Code) on laptops and in the cloud, coordinated through a shared log of claims and results. Teammates ran their own agents overnight to search for improvements and re-check each other's results. The team and the agents did the work. My job was making sure the work pointed at the right thing.
What I learned
- Measure how your validation transfers before trusting it. Ours overstated some kinds of gain by a third, and others entirely.
- Representation beats model size. Phase 1's biggest jumps came from how taste was represented, not from the model on top.
- Test on data shaped like what's scored. An idea we rejected before the final's first round, tested on the wrong kind of data, turned out to be the most important idea of the event.
- Coordination is a technical problem. A shared, append-only claims log stopped two people (or agents) from building the same thing twice.
- Agents scale effort, not judgement. They multiplied what we could test overnight, but a person still had to decide what counted as evidence.
Thanks
To my teammates Adarsh Arun, Nihal Mala, Mithil Mudholkar, Tanmai Hanumanthu, Shaurya Kapoor and Anish Achar, and to Huawei Ireland Research Centre and the organising team for a well-designed challenge.