Odor mixtures: a classical model can hold its own against deep networks
A preprint on arXiv shows that XGBoost, with explicit comparison features, competes with deep models on odor mixture similarity prediction.

In predicting the similarity of odor mixtures, a classical machine learning method can come close to deep learning models, deep neural networks, provided it receives the data in the right form. This is the thesis of a preprint published on arXiv, that is, a study distributed by its authors in its public version.1 According to the authors, XGBoost, a classical algorithm, reaches competitive performance on the main benchmark with specialized deep models for mixtures, when it is given explicit points of comparison between the two mixtures.1
Why predicting similarity between mixtures is difficult
The task requires reasoning about relationships between sets of odorant molecules rather than individual molecules. The Principal Odor Map (POM) provides good numerical representations of single molecules, but an effective representation for one molecule is not automatically useful for comparing mixtures.1 An earlier work, POMMIX, showed that taking the structure of mixtures into account within a deep learning model, with attention-based interactions, substantially improves the results. The authors of the preprint read that result this way: what matters is a model aware of the structure of mixtures, not the superiority of deep learning in itself.1
What they changed in classical models
In earlier comparisons, classical models started from weak representations: two descriptors calculated separately, one per mixture, then concatenated and given to a regressor. According to the authors, these data say what each mixture is but not how the two differ, and with little data available this can penalize classical models.1 The new work instead builds symmetric features for unordered pairs of mixtures, using sums, products and absolute differences calculated element by element on the aggregated descriptors. To these it adds a smaller block of statistics on the molecular interactions between the two mixtures.1 The representations were tested with XGBoost and with kernel ridge regression (KRR), a regression that uses a kernel function to measure the similarity between data.1
What they found
In the ablation studies, in which components are removed to see how much they weigh, the absolute differences alone already give good predictive ability. Adding sums and products improves it further.1 KRR also benefits substantially from explicit comparison features, with further improvements thanks to the choice of kernel and data preprocessing.1
What it means
Their main claim is that what may settle the outcome is not the divide between classical and deep methods but whether the input spells out the comparison between the two mixtures.1 The result concerns the main benchmark and a context with little data, such as artificial olfaction. The authors conclude that, for classical models in this type of prediction, explicitly representing how two mixtures differ is a key element.1
This article was written by artificial intelligence from the sources cited, without human review. Automatic newsroom by Redazia.


