With the Right Representation, Classical Machine Learning Rivals Deep Models for Olfactory Mixture Similarity
Organizations: Department of Statistics, University of Michigan, Ann Arbor, MI, United States
Abstract
Modeling olfactory mixtures remains a critical challenge in machine olfaction because the task requires reasoning over relationships between sets of odorants rather than over individual molecules in isolation. However, prior comparisons between deep and classical approaches also varied how the relationship between the two mixtures was represented, since the classical baselines primarily relied on weak representations that described each mixture independently and provided little direct signal about how the mixtures differed. In this work, we revisit classical learning using explicit symmetric comparison features on unordered pairs of mixtures, evaluating these representations with XGBoost and kernel ridge regression (KRR). We find that XGBoost with explicit comparison features achieves performance competitive with specialized deep mixture models on the main benchmark. Ablation studies show that coordinate-wise absolute differences alone provide strong predictive performance, while combining them with sums and products yields further improvements. KRR also benefits substantially from explicit comparison features, with further gains from appropriate kernel choice and preprocessing. Overall, these results suggest that explicitly encoding how two mixtures differ is an important ingredient for classical machine learning in low-data olfactory mixture prediction.
Figures & tables
| Method | Pearson | RMSE | Kendall |
|---|---|---|---|
| Snitz baseline | |||
| XGBoost RDKit | |||
| XGBoost POM embed | |||
| CheMix POM embed | |||
| POMMix | |||
| XGBoost mixture-level |
| Representation | Maximum depth | Pearson | RMSE |
|---|---|---|---|
| Independent POM | 6 | ||
| Full mixture-level | 6 | ||
| Independent POM | 1000 | ||
| Full mixture-level | 1000 |
| Model | Pearson | RMSE | Validation Pearson |
|---|---|---|---|
| Mixture-level only | |||
| Cross-mixture only | |||
| Combined |
| Features | Pearson | RMSE |
|---|---|---|
| Absolute differences only | ||
| Sums and products only | ||
| Full mixture-level |
| Feature family | Gain share |
|---|---|
| sym_absdiff_* | |
| sym_prod_* | |
| sym_sum_* | |
| cross_* |
| Model | Pearson | RMSE | Validation Pearson |
|---|---|---|---|
| KRR sample-product kernel | |||
| KRR explicit absdiff kernel | |||
| KRR explicit mixture-level kernel | |||
| KRR explicit combined kernel | |||
| XGBoost mixture-level only | |||
| XGBoost combined |
Appendix figures & tables24 assets
Supplementary material from the paper’s appendix.
Appendix
| Training seed | Mixture-level | Combined |
|---|---|---|
| Pearson correlation | ||
| 0 | ||
| 1 | ||
| 2 | ||
| 3 | ||
| 4 | ||
| Depth | Features | Pearson | RMSE |
|---|---|---|---|
| 6 | Absolute differences | ||
| 6 | Sums and products | ||
| 6 | Full mixture-level | ||
| 8 | Absolute differences | ||
| 8 | Sums and products | ||
| 8 | Full mixture-level |
| Depth | Pearson gap | RMSE gap | Gap change |
|---|---|---|---|
| 6 | |||
| 8 | |||
| 10 | |||
| 20 | |||
| 50 | |||
| 100 |
| Depth | cv0 | cv1 | cv2 | cv3 | cv4 |
|---|---|---|---|---|---|
| 6 | |||||
| 8 | |||||
| 10 | |||||
| 20 | |||||
| 50 | |||||
| 100 |
| Depth | cv0 | cv1 | cv2 | cv3 | cv4 |
|---|---|---|---|---|---|
| 6 | |||||
| 8 | |||||
| 10 | |||||
| 20 | |||||
| 50 | |||||
| 100 |
| Limit | Features | Train RMSE | Actual depth | Selected trees |
|---|---|---|---|---|
| 6 | Absolute differences | 6–6 | ||
| 6 | Sums and products | 6–6 | ||
| 6 | Full mixture-level | 6–6 | ||
| 8 | Absolute differences | 8–8 | ||
| 8 | Sums and products | 8–8 | ||
| 8 | Full mixture-level | 8–8 |
| Seed | Features | Pearson | RMSE |
|---|---|---|---|
| 0 | Absolute differences | ||
| 0 | Sums and products | ||
| 0 | Full mixture-level | ||
| 1 | Absolute differences | ||
| 1 | Sums and products | ||
| 1 | Full mixture-level |
| Seed | Features | Pearson | RMSE |
|---|---|---|---|
| 0 | Absolute differences | ||
| 0 | Sums and products | ||
| 0 | Full mixture-level | ||
| 1 | Absolute differences | ||
| 1 | Sums and products | ||
| 1 | Full mixture-level |
| Seed | Features | Pearson | RMSE |
|---|---|---|---|
| 0 | Absolute differences | ||
| 0 | Sums and products | ||
| 0 | Full mixture-level | ||
| 1 | Absolute differences | ||
| 1 | Sums and products | ||
| 1 | Full mixture-level |
| Seed | Features | Pearson | RMSE |
|---|---|---|---|
| 0 | Absolute differences | ||
| 0 | Sums and products | ||
| 0 | Full mixture-level | ||
| 1 | Absolute differences | ||
| 1 | Sums and products | ||
| 1 | Full mixture-level |
| Seed | Features | Pearson | RMSE |
|---|---|---|---|
| 0 | Absolute differences | ||
| 0 | Sums and products | ||
| 0 | Full mixture-level | ||
| 1 | Absolute differences | ||
| 1 | Sums and products | ||
| 1 | Full mixture-level |
| Depth | Features | Pearson | RMSE |
|---|---|---|---|
| 6 | Absolute differences | ||
| 6 | Sums and products | ||
| 6 | Full mixture-level | ||
| 8 | Absolute differences | ||
| 8 | Sums and products | ||
| 8 | Full mixture-level |
| Depth | Features | Pearson | RMSE |
|---|---|---|---|
| 20 | Absolute differences | ||
| 20 | Sums and products | ||
| 20 | Full mixture-level | ||
| 50 | Absolute differences | ||
| 50 | Sums and products | ||
| 50 | Full mixture-level |
| Depth | Pearson gap | RMSE gap | Gap change |
|---|---|---|---|
| 6 | |||
| 8 | |||
| 10 | |||
| 12 | |||
| 16 | |||
| 24 |
| Depth | cv0 | cv1 | cv2 | cv3 | cv4 |
|---|---|---|---|---|---|
| 6 | |||||
| 8 | |||||
| 10 | |||||
| 12 | |||||
| 16 | |||||
| 24 |
| Depth | cv0 | cv1 | cv2 | cv3 | cv4 |
|---|---|---|---|---|---|
| 6 | |||||
| 8 | |||||
| 10 | |||||
| 12 | |||||
| 16 | |||||
| 24 |
| Limit | Features | Train RMSE | Actual depth | Selected trees |
|---|---|---|---|---|
| 6 | Absolute differences | 6–6 | ||
| 6 | Sums and products | 6–6 | ||
| 6 | Full mixture-level | 6–6 | ||
| 8 | Absolute differences | 8–8 | ||
| 8 | Sums and products | 8–8 | ||
| 8 | Full mixture-level | 8–8 |
| Seed | Features | Pearson | RMSE |
|---|---|---|---|
| 0 | Absolute differences | ||
| 0 | Sums and products | ||
| 0 | Full mixture-level | ||
| 1 | Absolute differences | ||
| 1 | Sums and products | ||
| 1 | Full mixture-level |
| Seed | Features | Pearson | RMSE |
|---|---|---|---|
| 0 | Absolute differences | ||
| 0 | Sums and products | ||
| 0 | Full mixture-level | ||
| 1 | Absolute differences | ||
| 1 | Sums and products | ||
| 1 | Full mixture-level |
| Seed | Features | Pearson | RMSE |
|---|---|---|---|
| 0 | Absolute differences | ||
| 0 | Sums and products | ||
| 0 | Full mixture-level | ||
| 1 | Absolute differences | ||
| 1 | Sums and products | ||
| 1 | Full mixture-level |
| Seed | Features | Pearson | RMSE |
|---|---|---|---|
| 0 | Absolute differences | ||
| 0 | Sums and products | ||
| 0 | Full mixture-level | ||
| 1 | Absolute differences | ||
| 1 | Sums and products | ||
| 1 | Full mixture-level |
| Model / preprocessing | Pearson | RMSE |
|---|---|---|
| XGBoost, primary inputs | ||
| XGBoost, standardized | ||
| XGBoost, standardized and rotated | ||
| KRR, linear | ||
| KRR, original RBF grid | ||
| KRR, median-distance RBF grid |
Pairwise Comparison
Pairwise comparison uses judgments between two items to infer rankings, preferences, quality scores, or decisions from limited and noisy data.
See all 81 papers