We Used 5 Outlier Detection Strategies on a Actual Dataset: They Disagreed on 96% of Flagged Samples

# Introduction

All tutorials on information science make detecting outliers seem like fairly straightforward. Take away all values higher than three normal deviations; that is all there’s to it. However when you begin working with an precise dataset the place the distribution is skewed and a stakeholder asks, “Why did you take away that information level?” you out of the blue notice you do not have a superb reply.

So we ran an experiment. We examined 5 of essentially the most generally used outlier detection strategies on an actual dataset (6,497 Portuguese wines) to search out out: do these strategies produce constant outcomes?

They did not. What we realized from the disagreement turned out to be extra beneficial than something we might have picked up from a textbook.

Picture by Writer

We constructed this evaluation as an interactive Strata pocket book, a format you should utilize in your personal experiments utilizing the Information Undertaking on StrataScratch. You may view and run the complete code right here.

# Setting Up

 
Our information comes from the Wine High quality Dataset, publicly accessible by UCI’s Machine Studying Repository. It comprises physicochemical measurements from 6,497 Portuguese “Vinho Verde” wines (1,599 pink, 4,898 white), together with high quality scores from professional tasters.

We chosen it for a number of causes. It is manufacturing information, not one thing generated artificially. The distributions are skewed (6 of 11 options have skewness ( > 1 )), so the information don’t meet textbook assumptions. And the standard scores allow us to test if the detected “outliers” present up extra amongst wines with uncommon scores.

Beneath are the 5 strategies we examined:

# Discovering the First Shock: Inflated Outcomes From A number of Testing

 
Earlier than we might examine strategies, we hit a wall. With 11 options, the naive method (flagging a pattern based mostly on an excessive worth in at the very least one characteristic) produced extraordinarily inflated outcomes.

IQR flagged about 23% of wines as outliers. Z-Rating flagged about 26%.

When almost 1 in 4 wines get flagged as outliers, one thing is off. Actual datasets don’t have 25% outliers. The issue was that we had been testing 11 options independently, and that inflates the outcomes.

The maths is simple. If every characteristic has lower than a 5% chance of getting a “random” excessive worth, then with 11 unbiased options:
[ P(text{at least one extreme}) = 1 – (0.95)^{11} approx 43% ]

In plain phrases: even when each characteristic is completely regular, you’d count on almost half your samples to have at the very least one excessive worth someplace simply by random probability.

To repair this, we modified the requirement: flag a pattern solely when at the very least 2 options are concurrently excessive.

 

 
Altering min_features from 1 to 2 modified the definition from “any characteristic of the pattern is excessive” to “the pattern is excessive throughout multiple characteristic.”

Here is the repair in code:

# Rely excessive options per pattern
outlier_counts = (np.abs(z_scores) > 3.5).sum(axis=1)
outliers = outlier_counts >= 2

# Evaluating 5 Strategies on 1 Dataset

 
As soon as the multiple-testing repair was in place, we counted what number of samples every technique flagged:

 

 
Here is how we arrange the ML strategies:

from sklearn.ensemble import IsolationForest
from sklearn.neighbors import LocalOutlierFactor
 
iforest = IsolationForest(contamination=0.05, random_state=42)
lof = LocalOutlierFactor(n_neighbors=20, contamination=0.05)

Why do the ML strategies all present precisely 5%? Due to the contamination parameter. It requires them to flag precisely that share. It is a quota, not a threshold. In different phrases, Isolation Forest will flag 5% no matter whether or not your information comprises 1% true outliers or 20%.

# Discovering the Actual Distinction: They Establish Totally different Issues

 
Here is what shocked us most. After we examined how a lot the strategies agreed, the Jaccard similarity ranged from 0.10 to 0.30. That is poor settlement.

Out of 6,497 wines:

• Solely 32 samples (0.5%) had been flagged by all 4 main strategies
• 143 samples (2.2%) had been flagged by 3+ strategies
• The remaining “outliers” had been flagged by only one or 2 strategies

You would possibly suppose it is a bug, nevertheless it’s the purpose. Every technique has its personal definition of “uncommon”:

 

 
If a wine has residual sugar ranges considerably increased than common, it is a univariate outlier (Z-Rating/IQR will catch it). But when it is surrounded by different wines with comparable sugar ranges, LOF will not flag it. It is regular throughout the native context.

So the true query is not “which technique is finest?” It is “what sort of uncommon am I looking for?”

# Checking Sanity: Do Outliers Correlate With Wine High quality?

 
The dataset contains professional high quality scores (3-9). We needed to know: do detected outliers seem extra steadily amongst wines with excessive high quality scores?

 

 
Excessive-quality wines had been twice as prone to be consensus outliers. That is a superb sanity test. In some instances, the connection is evident: a wine with manner an excessive amount of risky acidity tastes vinegary, will get rated poorly, and will get flagged as an outlier. The chemistry drives each outcomes. However we will not assume this explains each case. There may be patterns we’re not seeing, or confounding components we have not accounted for.

# Making Three Selections That Formed Our Outcomes

// 1. Utilizing Sturdy Z-Rating Slightly Than Normal Z-Rating

A Normal Z-Rating makes use of the imply and normal deviation of the information, each of that are affected by the outliers current in our dataset. A Sturdy Z-Rating as a substitute makes use of the median and Median Absolute Deviation (MAD), neither of which is affected by outliers.

Consequently, the Normal Z-Rating recognized 0.8% of the information as outliers, whereas the Sturdy Z-Rating recognized 3.5%.

# Sturdy Z-Rating utilizing median and MAD
median = np.median(information, axis=0)
mad = np.median(np.abs(information - median), axis=0)
robust_z = 0.6745 * (information - median) / mad

// 2. Scaling Crimson And White Wines Individually

Crimson and white wines have completely different baseline ranges of chemical substances. For instance, when combining pink and white wines right into a single dataset, a pink wine that has completely common chemistry relative to different pink wines could also be recognized as an outlier based mostly solely on its sulfur content material in comparison with the mixed imply of pink and white wines. Due to this fact, we scaled every wine sort individually utilizing the median and Interquartile Vary (IQR) of every wine sort, after which mixed the 2.

# Scale every wine sort individually
from sklearn.preprocessing import RobustScaler
scaled_parts = []
for wine_type in ['red', 'white']:
    subset = df[df['type'] == wine_type][features]
    scaled_parts.append(RobustScaler().fit_transform(subset))

// 3. Realizing When To Exclude A Methodology

Elliptic Envelope assumes your information follows a multivariate regular distribution. Ours did not. Six of 11 options had skewness above 1, and one characteristic hit 5.4. We saved the Elliptic Envelope within the comparability for completeness, however left it out of the consensus vote.

# Figuring out Which Methodology Performs Finest For This Wine Dataset

Picture by Writer

Can we choose a “winner” given the traits of our information (heavy skewness, blended inhabitants, no identified floor reality)?

Sturdy Z-Rating, IQR, Isolation Forest, and LOF all deal with skewed information moderately effectively. If pressured to select one, we might go together with Isolation Forest: no distribution assumptions, considers all options directly, and offers with blended populations gracefully.

However no single technique does every little thing:

• Isolation Forest can miss outliers which can be solely excessive on one characteristic (Z-Rating/IQR catches these)
• Z-Rating/IQR can miss outliers which can be uncommon throughout a number of options (multidimensional outliers)

The higher method: use a number of strategies and belief the consensus. The 143 wines flagged by 3 or extra strategies are much more dependable than something flagged by a single technique alone.

Here is how we calculated consensus:

# Rely what number of strategies flagged every pattern
consensus = zscore_out + iqr_out + iforest_out + lof_out
high_confidence = df[consensus >= 3]  # Recognized by 3+ strategies

With out floor reality (as in most real-world initiatives), technique settlement is the closest measure of confidence.

# Understanding What All This Means For Your Personal Initiatives

 
Outline your downside earlier than selecting your technique. What sort of “uncommon” are you really searching for? Information entry errors look completely different from measurement anomalies, and each look completely different from real uncommon instances. The kind of downside factors to completely different strategies.

Verify your assumptions. In case your information is closely skewed, the Normal Z-Rating and Elliptic Envelope will steer you unsuitable. Have a look at your distributions earlier than committing to a technique.

Use a number of strategies. Samples flagged by three or extra strategies with completely different definitions of “outlier” are extra reliable than samples flagged by only one.

Do not assume all outliers ought to be eliminated. An outlier could possibly be an error. It may be your most attention-grabbing information level. Area information makes that decision, not algorithms.

# Concluding Remarks

 
The purpose right here is not that outlier detection is damaged. It is that “outlier” means various things relying on who’s asking. Z-Rating and IQR catch values which can be excessive on a single dimension. Isolation Forest and LOF discover samples that stand out of their general sample. Elliptic Envelope works effectively when your information is definitely Gaussian (ours wasn’t).

Work out what you are actually searching for earlier than you choose a technique. And in case you’re unsure? Run a number of strategies and go together with the consensus.

# FAQs

// 1. Figuring out Which Approach I Ought to Begin With

A great place to start is with the Isolation Forest method. It doesn’t assume how your information is distributed and makes use of all your options on the identical time. Nonetheless, if you wish to determine excessive values for a selected measurement (reminiscent of very hypertension readings), then Z-Rating or IQR could also be extra appropriate for that.

// 2. Selecting a Contamination Fee For Scikit-learn Strategies

It relies on the issue you are attempting to unravel. A generally used worth is 5% (or 0.05). However understand that contamination is a quota. Which means that 5% of your samples shall be labeled as outliers, no matter whether or not there really are 1% or 20% true outliers in your information. Use a contamination price based mostly in your information of the proportion of outliers in your information.

// 3. Eradicating Outliers Earlier than Splitting Prepare/take a look at Information

No. You must match an outlier-detection mannequin to your coaching dataset, after which apply the educated mannequin to your testing dataset. If you happen to do in any other case, your take a look at information is influencing your preprocessing, which introduces leakage.

// 4. Dealing with Categorical Options

The strategies lined right here work on numerical information. There are three potential options for categorical options:

• encode your categorical variables and proceed;
• use a way designed for mixed-type information (e.g. HBOS);
• run outlier detection on numeric columns individually and use frequency-based strategies for categorical ones.

// 5. Realizing If A Flagged Outlier Is An Error Or Simply Uncommon

You can’t decide from the algorithm alone when an recognized outlier represents an error versus when it’s merely uncommon. It flags what’s uncommon, not what’s unsuitable. For instance, a wine that has an especially excessive residual sugar content material may be a knowledge entry error, or it may be a dessert wine that’s meant to be that candy. Finally, solely your area experience can present a solution. If you happen to’re uncertain, mark it for evaluation somewhat than eradicating it mechanically.
 
 

Nate Rosidi is a knowledge scientist and in product technique. He is additionally an adjunct professor educating analytics, and is the founding father of StrataScratch, a platform serving to information scientists put together for his or her interviews with actual interview questions from prime firms. Nate writes on the newest developments within the profession market, offers interview recommendation, shares information science initiatives, and covers every little thing SQL.