Tal Korem

Tal Korem @tkorem.bsky.social [ View ] Jun 11, 2024 at 13:51 UTC

Importantly - we'd love to hear your comments, feedback, and GitHub issues! In particular if there’s additional prior work on this topic that we should note.

0 replies 0 reposts 0 likes

Tal Korem @tkorem.bsky.social [ View ] Jun 11, 2024 at 13:51 UTC

But CV is used not just for evaluation but also for hyperparameter tuning, and distributional bias impacts HPs that affect regression to the mean. For example, we show that it biases for weaker model regularization, which might affect generalization and downstream deployment.

1 replies 0 reposts 0 likes

Tal Korem @tkorem.bsky.social [ View ] Jun 11, 2024 at 13:51 UTC

With RebalancedCV we could see the "real-life" impact of distributional bias. We reproduced 3 recently published analyses that used LOOCV, and showed that it under-evaluated performance in all of them. While the effect isn't major, it is consistent.

1 replies 0 reposts 1 likes

Tal Korem @tkorem.bsky.social [ View ] Jun 11, 2024 at 13:50 UTC

With this in mind, we developed RebalancedCV, an sklearn-compatible package which drops the minimal amount of samples from the training set to maintain the same class balance in the training sets of all folds, thus resolving distributional bias. github.com/korem-lab/Re...

1 replies 0 reposts 1 likes

Tal Korem @tkorem.bsky.social [ View ] Jun 11, 2024 at 13:50 UTC

As the issue is caused by a shift in the class balance of the training set, distributional bias can be addressed with stratified CV - but only if your dataset allows it to happen precisely. The less exact the stratification - the more bias you have (in this plot, closer to 0).

1 replies 0 reposts 0 likes

Tal Korem @tkorem.bsky.social [ View ] Jun 11, 2024 at 13:49 UTC

Does this mean that past work with LOOCV is overinflated? Not quite. Most machine learning algorithms regress to the mean - not to its negative - and so they are actually _under_evaluated. That's the negative bias we started with!

1 replies 0 reposts 1 likes

Tal Korem @tkorem.bsky.social [ View ] Jun 11, 2024 at 13:49 UTC

Distributional bias is a severe information leakage - so severe that we designed a dummy model that can achieve perfect auROC/auPR in ANY binary classification task evaluated via LOOCV (even without features). How? it just outputs the negative mean of the training set labels!

1 replies 0 reposts 0 likes

Tal Korem @tkorem.bsky.social [ View ] Jun 11, 2024 at 13:48 UTC

The issue is that every time one holds out a sample as a test set in LOOCV, the mean label average of the training set shifts slightly, creating a perfect negative correlation across the folds between that mean and the test labels. We call this phenomenon distributional bias:

1 replies 0 reposts 1 likes

Tal Korem @tkorem.bsky.social [ View ] Jun 11, 2024 at 13:48 UTC

This story begins with benchmarking we did for some of our machine learning pipelines. We used random data, so we expected to see random classification accuracy (auROC=0.5). Instead, we found a clear negative bias, that got worse with more imbalanced datasets:

1 replies 0 reposts 1 likes

Tal Korem @tkorem.bsky.social [ View ] Jun 11, 2024 at 13:47 UTC

A bit of background: when training models on small datasets it’s common to use LOOCV, as it maximizes the N of samples for training. It also leaves a single sample for testing, meaning that many performance metrics (e.g., area under ROC curve) require aggregation across folds/iterations.

1 replies 0 reposts 0 likes

Tal Korem @tkorem.bsky.social [ View ] Jun 11, 2024 at 13:46 UTC

New preprint!

In a few words: we don't think you should use leave-one-out cross-validation (LOOCV).

In a lot of words (+RebalancedCV, a LOOCV alternative): arxiv.org/abs/2406.01652

In a thread:

1 replies 3 reposts 4 likes

Tal Korem @tkorem.bsky.social [ View ] May 20, 2024 at 14:44 UTC

So, ISO an alternative search engine to Google, I guess. (It's 9:43 am CDT)

0 replies 0 reposts 0 likes

Tal Korem @tkorem.bsky.social [ View ] May 15, 2024 at 00:13 UTC

That's a cool idea and there's probably public data to do it -- thanks. What we did n that direction was to take paired long/short reads from the same sample, assume that the long reads are a real genome from the sample, and calculate bias with respect to them. It's not gone but it's much reduced.

0 replies 0 reposts 1 likes

Tal Korem @tkorem.bsky.social [ View ] May 15, 2024 at 00:09 UTC

Very apt sequence from the other place
( @baym.lol )

0 replies 0 reposts 1 likes

Tal Korem @tkorem.bsky.social [ View ] May 14, 2024 at 03:11 UTC

When we look at known strain mixtures - the bias goes away! (Fig 5)

1 replies 0 reposts 2 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 23, 2024 at 01:44 UTC

🖥️ 🧬 #microbiome

0 replies 0 reposts 0 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 22, 2024 at 15:16 UTC

Note:  We don't prove there is no leakage in Poore et al., and we don't discuss database errors. We tackle the general question of whether that specific observation is sufficient to indicate a problematic analysis.

0 replies 0 reposts 2 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 22, 2024 at 15:16 UTC

Bottom line: If phenotype-associated originally empty features invalidate downstream analyses, then CLR is also invalid. Gihawi et al. did not present sufficient evidence to claim there is information leakage or error in machine learning (they do show that interpretation is complicated).

1 replies 0 reposts 1 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 22, 2024 at 15:15 UTC

This is the same as predicting preterm birth with "Blautia (CLR)" or "empty feature (CLR)"  - creating a legitimate microbiome predictor - just not one that’s easy to interpret.

1 replies 0 reposts 1 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 22, 2024 at 15:14 UTC

Wait, but didn't Gihawi et al. run Poore et al.'s code on a matrix of zeros and get an accurate classifier? Many got this impression, but the text is clear on what was done - they took a subset of the processed Voom-SNM matrix.

1 replies 0 reposts 1 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 22, 2024 at 15:13 UTC

We can actually see this in three of the four examples that Gihawi et al highlight: a simple CLR transform (sample-wise - so no leakage) recreates the same observation of values associated with a tumor type. Here it is for a weird virus and adrenocortical carcinoma

1 replies 0 reposts 0 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 22, 2024 at 15:12 UTC

Once more - this Blautia OTU is not really there, and it is definitely not related to preterm birth - but it is a real microbiome signature: it represents the (inverse) alpha diversity of the samples.

1 replies 1 reposts 1 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 22, 2024 at 15:12 UTC

But what's biologically "real" about the geometric mean? so, for example, it's related to alpha diversity. To show this, we analyze a real vaginal microbiome dataset. We take the sparsest feature - probably not really there - and once again, after CLR, it's associated with preterm birth.

1 replies 0 reposts 2 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 22, 2024 at 15:11 UTC

Is this feature biologically related to the phenotype? no. Is it really there? no. But is it leakage? no. This is a real microbiome signature! It represents the inverse of the sample geometric mean (and not, however, whatever is in the column title of the feature).

1 replies 0 reposts 0 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 22, 2024 at 15:10 UTC

First, we simulate a 50:50 case:control study in which case samples have a higher geometric mean. We then add an all-zero feature. After CLR? That feature has values and they are perfectly associated with the phenotype.

1 replies 0 reposts 1 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 22, 2024 at 15:09 UTC

We use counter-examples to show that it isn't. We use the CLR transform (with pseudocount), which is widely used in microbiome analysis. CLR uses only the taxa abundances within the samples - no phenotypes or other samples - so no information leakage or artificial tags.

1 replies 0 reposts 0 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 22, 2024 at 15:09 UTC

Gihawi et al. found that taxa with high importance in tumor type classifiers by Poore et al. (2020) were almost all zero in the raw data. They claim this indicates an artificial tag, information leakage, and artifactual models. Is this a necessary conclusion?
journals.asm.org/doi/10.1128/...

1 replies 0 reposts 1 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 22, 2024 at 15:08 UTC

Is finding taxa that were originally zero but have phenotype-associated values following batch correction a "major data analysis error" that invalidates downstream classifiers (in the 2020 cancer-microbiome paper or in general)?
We think it's not.
📰+🧵
www.biorxiv.org/content/10.1...

3 replies 3 reposts 1 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 21, 2024 at 23:17 UTC

🟦🟦🟦🟦 🟩🟪🟪🟩 🟩🟩🟩🟩 🟨🟨🟨🟨 🟪🟪🟪🟪 זה היה מאוד כיף

1 replies 0 reposts 2 likes

Tal Korem @tkorem.bsky.social [ View ] Feb 16, 2024 at 01:22 UTC

Cancer research is also metabolism, bioinformatics, genetics, comp bio, cell bio, etc etc, and no one would dare say it's not a discipline

1 replies 0 reposts 1 likes