节点 n17

Looking at the results so far:
- Round 0 (local-k=10): 59.137 (+0.007) - negligible change
- Round 1 (alpha=2.0): 57.961 (-1.169) - worse

The alpha=2.0 hurt significantly, mainly through cell_state dropping from 83.16 to 79.17. So alpha=1.5 is better.

For round 2, I need to find something that gives >1 point improvement. Let me think about what parameters haven't been explored:

1. **local-eta=-2.5**: Between -2 (57.95) and -3 (58.87). The optimum seems near -3, but maybe -2.5 with local-k=10 could be different.

2. **proj-k=20 or 30**: proj-k=25 is current, proj-k=40 was worse (57.70). Maybe proj-k=30 could help.

3. **proj-hvg**: Currently 8000. Maybe 6000 or 10000 could help.

4. **proj-cells**: Currently 3000. Maybe 5000 could give a better PC basis.

5. **Combining local-k=10 with a slightly different eta**: Since local-k=10 slightly improved covariation (45.33 vs 44.98) but slightly hurt cell_state (82.90 vs 83.16), maybe adjusting eta to -3.5 with k=10 could recover cell_state while keeping the covariation gain.

Actually, let me think about this differently. The main issue is that the score is 59.13 and I need >60.13. That's a big jump.

Looking at the components again:
- cell_state: 82.90 (with k=10) - this is already high
- covariation: 45.33 - this is the weakest
- de_recovery: 50.96 - moderate
- direction: 49.84 - moderate

To get a >1 point improvement, I need to find a parameter that meaningfully improves multiple components or significantly improves one.

Let me think about proj-k. Currently 25. The PC subspace captures the covariation structure. A slightly different k might better capture the relevant structure. proj-k=40 was worse, but what about proj-k=30 or proj-k=20?

Actually, let me reconsider the eta parameter with local-k=10. The interaction between k and eta might be different. With k=10, the local directions are different from k=15, so the optimal eta might shift. Let me try eta=-3.5 with k=10. With fewer neighbors, the local directions might be noisier, so a slightly larger |eta| could be needed.

Wait, actually let me think about what could give a substantial improvement. The score formula seems to be a weighted average. If I look at the components:
- de_recovery: 50.96
- direction: 49.84  
- cell_state: 82.90
- covariation: 45.33

If weights are roughly equal (say 25% each), the average is about 57.26, but the actual score is 59.137. So the weights aren't equal. Maybe cell_state has a higher weight?

（只摘前 40 行）