Beyond Percentiles: Advanced Statistical Methods for Defining System Thresholds

GOKE ADEKUNLE; #Wolfwords
11 min readJan 25, 2024

--

As I delved into Emmanuel Bakare’s insightful blog post, “Exploring The Math of Thresholds,” I found myself captivated by the intricacies of defining thresholds in system monitoring. The journey through percentiles, curve fitting, and the law of large numbers left me hungry for a deeper understanding. It sparked a curiosity to explore beyond the familiar territory of percentiles and venture into the realm of advanced statistical methods for defining system thresholds.

Emmanuel’s careful investigation into percentiles, which involves grouping data into different categories based on sorting methods, provided a solid starting point. However, considering the constantly changing nature of data analytics, there’s a call to delve into more advanced statistical techniques. These sophisticated methods can better handle the various challenges posed by monitoring different types of systems.

Please understand that OP’s exploration of thresholds serves as a guiding light, laying the groundwork for this expedition into advanced statistical methods. So as I journey beyond percentiles, I draw inspiration from his insights, acknowledging the significance of a nuanced understanding of system metrics.

In the upcoming segments, I intend to unravel the complexities of Bayesian thinking, decipher the language of hypothesis testing, and explore the symbiosis of machine learning and threshold definition.

Unveiling Advanced Statistical Techniques

1. Bayesian Approaches to Threshold Definition

Embracing Bayesian statistics for threshold determination represents a departure from traditional methods, introducing a probabilistic perspective that adapts to the dynamic nature of system behaviour. Let’s critically explore the merits and intricacies of Bayesian approaches, drawing on Emmanuel Bakare’s foundational insights into percentiles.

Incorporating Prior Knowledge

Strengths: Bayesian methods stand out in their ability to seamlessly integrate prior knowledge into the analysis. Unlike frequentist approaches, which rely solely on current data, Bayesian statistics allow us to incorporate existing insights or beliefs about the system. This proves invaluable when setting thresholds, especially in situations where historical context significantly influences system behaviour.

# Bayesian Threshold Definition
import pymc3 as pm

# Prior belief about threshold distribution
prior_distribution = pm.Normal("prior_distribution", mu=60, sigma=10)

# Observed data
observed_data = [65, 70, 55, 58, 62]

# Bayesian model
with pm.Model() as model:
threshold = pm.Normal("threshold", mu=prior_distribution, sigma=5)
likelihood = pm.Normal("likelihood", mu=threshold, observed=observed_data)

trace = pm.sample(1000, tune=1000)

By incorporating a prior distribution (prior belief) about the threshold, the Bayesian model combines historical knowledge with current observations, enabling a more informed determination of system thresholds.

Iterative Understanding of System Behaviour

Strengths: Bayesian approaches offer an iterative learning process. As new data becomes available, the model can be updated to refine our understanding of system behaviour. This adaptability is particularly advantageous in scenarios where the system’s dynamics evolve, ensuring that thresholds stay relevant.

# Updating Bayesian Model with New Data
new_observation = 67

with model:
new_likelihood = pm.Normal("new_likelihood", mu=threshold, observed=new_observation)
new_trace = pm.sample(1000, tune=1000, start=pm.sample_prior_predictive(trace))

The model can be easily updated with new observations, allowing for a continuous learning process. This iterative understanding ensures that thresholds remain accurate as the system evolves.

Challenges and Considerations

Challenges:

  • Computational Intensity: Bayesian methods can be computationally intensive, especially as the complexity of the model and the volume of data increase.
  • Subjectivity in Priors: The choice of prior distributions can introduce subjectivity. Sensitivity to the selection of priors requires careful consideration.

Case Study: Consider a scenario where the system’s behaviour is influenced by external factors. Bayesian methods might struggle if the prior beliefs do not align with the true dynamics of these external influences.

Conclusion

Bayesian approaches bring notable advantages to the table, offering a flexible framework for threshold determination. By seamlessly incorporating prior knowledge and enabling iterative updates, these methods address the challenges posed by evolving system behaviours. However, we all should be mindful of computational demands and the subjective nature of prior choices.

2. Statistical Hypothesis Testing: Unmasking Hidden Patterns

In data analytics, the pursuit of actionable insights and meaningful interpretations propels analysts towards the adoption of sophisticated or complex statistical methodologies. This section serves as a means to understand the latent potential of statistical hypothesis testing, a tool that empowers data analysts to unearth concealed patterns within complex datasets. I intend to try and dissect its strengths, delve into practical illustrations through code, and critically discuss its application in refining system thresholds. Drawing on Bakare’s insights into deterministic and non-deterministic systems, I’ll explore how hypothesis testing becomes a pivotal asset, especially in non-uniform and dynamic environments.

Unveiling Hidden Patterns

Strengths: Hypothesis testing serves as a powerful tool for revealing patterns that might otherwise remain obscured. By formulating and evaluating hypotheses about the system's behaviour, we systematically uncover underlying trends and relationships within the data.

Code Illustration:

# Hypothesis Testing Example
import scipy.stats as stats

# Sample data from a non-uniform system
data_sample = [25, 30, 35, 40, 45, 60, 65, 70, 75, 80]

# Perform a t-test to assess if the mean is significantly different from a reference value
reference_mean = 50
t_stat, p_value = stats.ttest_1samp(data_sample, reference_mean)

# Check the p-value to make a decision
if p_value < 0.05:
print("Reject the null hypothesis: The mean is significantly different.")
else:
print("Fail to reject the null hypothesis: The mean is not significantly different.")

In this example, a t-test is employed to assess whether the mean of the data sample significantly differs from a reference value. This demonstrates how hypothesis testing can reveal hidden patterns by evaluating statistical significance.

Application to Non-Uniform and Dynamic Environments

Strengths: Hypothesis testing proves particularly valuable in non-uniform and dynamic systems. These environments, characterised by varying patterns and behaviours, necessitate robust statistical tools. Hypothesis testing enables analysts to adaptively refine thresholds based on observed data, accommodating the inherent complexity of non-uniform systems.

Code Illustration:

# Hypothesis Testing in Dynamic Environment
import pandas as pd
from statsmodels.stats.multicomp import pairwise_tukeyhsd

# Simulated data for a dynamic system with multiple groups
data_dynamic = pd.DataFrame({
'group': ['A']*20 + ['B']*20 + ['C']*20,
'value': [30, 35, 40, 45, 50, 55, 60, 65, 70, 75]*6
})

# Perform Tukey's HSD test for pairwise comparisons
tukey_results = pairwise_tukeyhsd(data_dynamic['value'], data_dynamic['group'])

# Visualize the results
tukey_results.plot_simultaneous()

In a dynamic environment with multiple groups, Tukey’s HSD test is applied for pairwise comparisons. This showcases how hypothesis testing can adapt to non-uniform systems, providing insights into group differences.

Challenges and Considerations

Challenges:

  • Assumption Dependencies: Hypothesis testing often relies on assumptions such as normality and independence, which may not always hold in real-world scenarios.
  • P-value Misinterpretation: Improper interpretation of p-values can lead to erroneous conclusions. Caution is required to avoid the “significance threshold” trap.

Case Study: Consider a scenario where assumptions of normality are violated. Hypothesis testing might provide inaccurate results, emphasising the importance of understanding the underlying data distribution.

Conclusion

Statistical hypothesis testing emerges as a formidable ally in the quest to define system thresholds. Its ability to unveil hidden patterns and adapt to non-uniform, dynamic environments positions it as a versatile tool for data analysts. However, practitioners should remain vigilant about assumption dependencies and exercise prudence in interpreting p-values. Emmanuel Bakare’s insights into deterministic and non-deterministic systems aptly set the stage for appreciating the necessity of robust statistical methodologies in the nuanced realm of threshold definition.

3. Machine Learning Integration: A Fusion of Predictive Power

Here, I’m just trying to see how machine learning models can redefine the process of defining system thresholds. From regression models to ensemble methods, I’m trying to unravel the diverse applications and implications of incorporating machine learning prowess into this critical analytical domain.

The advantages of machine learning are apparent: the capability to discern intricate patterns, adapt to changing conditions, and forecast system behaviour with a degree of accuracy. However, the integration of machine learning into threshold-setting processes also presents challenges, notably in terms of model interpretability. Understanding the “why” behind a model’s prediction becomes crucial, especially when setting thresholds that impact critical decision-making processes.

Model Interpretability

The emphasis on model interpretability will be a recurring theme in this discussion. While machine learning models excel in predictive accuracy, the black-box nature of certain algorithms raises concerns about the interpretability of their decisions. I try to address this challenge by exploring techniques and approaches that strike a balance between predictive power and interpretability.

Reference to Bakman’s Insights: This exploration builds upon Emmanuel Bakare’s discussions on prediction intervals, drawing parallels between the predictive capabilities of machine learning models and the dynamic threshold determination discussed in Bakare’s work. By referencing Bakare’s insights, we aim to bridge the conceptual gap and showcase how machine learning extends beyond prediction intervals to offer a holistic and adaptive approach to defining system thresholds.

Code Illustrations and Case Studies:

To substantiate these discussions, I will be using practical implementations with code illustrations and case studies. Through hands-on examples, I should be able to demonstrate how different machine learning models can be applied to real-world scenarios, showcasing their effectiveness in capturing and adapting to the nuances of system behaviour.

Regression Models:

A major part of machine learning’s predictive power lies in its ability to leverage regression models. These models, such as linear regression or more complex variants like polynomial regression, excel at identifying linear and non-linear relationships within data. Through code illustrations, I intend to showcase how regression models can be trained to predict system metrics, offering insights into potential threshold values. This application becomes particularly valuable in deterministic systems where understanding the internal relationships between metrics is essential.

# Sample code for linear regression
from sklearn.linear_model import LinearRegression
from sklearn.model_selection import train_test_split
from sklearn.metrics import mean_squared_error

# Assuming 'features' and 'targets' are your dataset columns
X_train, X_test, y_train, y_test = train_test_split(features, targets, test_size=0.2, random_state=42)

# Initialize and train the linear regression model
regression_model = LinearRegression()
regression_model.fit(X_train, y_train)

# Make predictions on the test set
predictions = regression_model.predict(X_test)

# Evaluate the model performance
mse = mean_squared_error(y_test, predictions)
print(f"Mean Squared Error: {mse}")

Objective: Predict system metrics using a linear relationship with input features.

Code Explanation:

  • The code starts by splitting the dataset into training and testing sets.
  • It initializes a linear regression model, trains it on the training set, and predicts values for the test set.
  • Mean Squared Error (MSE) is calculated to evaluate the model’s performance.

Relevance: Linear regression is suitable for systems where a linear relationship between input metrics and the target metric exists.

Ensemble Methods for Robust Predictions

Ensemble methods, such as Random Forests or Gradient Boosting, bring robustness to predictive analysis. By combining multiple models, ensemble methods mitigate overfitting and enhance predictive accuracy. We’ll explore how ensemble methods can be employed to predict system metrics and discuss their role in refining threshold definitions.

# Sample code for Random Forest
from sklearn.ensemble import RandomForestRegressor

# Assuming 'features' and 'targets' are your dataset columns
X_train, X_test, y_train, y_test = train_test_split(features, targets, test_size=0.2, random_state=42)

# Initialize and train the Random Forest model
random_forest_model = RandomForestRegressor()
random_forest_model.fit(X_train, y_train)

# Make predictions on the test set
predictions = random_forest_model.predict(X_test)

# Evaluate the model performance
mse = mean_squared_error(y_test, predictions)
print(f"Mean Squared Error: {mse}")

Objective: Enhance predictive accuracy using an ensemble method (Random Forest).

Code Explanation:

  • Similar to linear regression, the dataset is split, and a Random Forest model is initialised, trained, and used for predictions.
  • MSE is calculated to assess the model’s accuracy.

Relevance: Ensemble methods like Random Forest are robust and suitable for capturing non-linear relationships within data.

Interpretable Machine Learning

The challenge of interpretability in machine learning models is met with techniques like SHAP (SHapley Additive exPlanations) values or LIME (Local Interpretable Model-agnostic Explanations). These methods provide insights into the features driving model predictions, aiding in understanding the factors influencing threshold determinations.

# Sample code for SHAP values
import shap

# Assuming 'model' is your trained machine learning model
explainer = shap.Explainer(model)
shap_values = explainer.shap_values(X_test)

# Summary plot to visualize feature importance
shap.summary_plot(shap_values, X_test)

Objective: Understand feature importance in machine learning predictions.

Code Explanation:

  • SHAP values are computed using an explainer on the trained model and input data.
  • A summary plot visualises the importance of features in driving model predictions.

Relevance: Interpretability is crucial for understanding the factors influencing threshold determinations in real-world scenarios.

Case Studies: Unraveling the Impact of Machine Learning on System Thresholds

Case Study 1: Non-Uniform System Behavior

Context: Consider a scenario where a web application experiences fluctuating user activity throughout the day. Traditional static thresholds struggle to adapt to these non-uniform patterns, often triggering false alarms or missing genuine anomalies.

Machine Learning Approach: By leveraging machine learning, specifically time-series forecasting models, the system can predict expected user activity at different time intervals. In this case study, we employ a Long Short-Term Memory (LSTM) neural network for its ability to capture sequential dependencies in time-series data.

# Sample code for LSTM time-series forecasting
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense

# Assuming 'features' and 'targets' are your time-series data
model = Sequential()
model.add(LSTM(50, activation='relu', input_shape=(n_steps, n_features)))
model.add(Dense(1))
model.compile(optimizer='adam', loss='mse')

# Train the LSTM model
model.fit(X_train, y_train, epochs=50, validation_data=(X_val, y_val))

Results: The LSTM model learns the temporal patterns in user activity and provides predictions with adaptability to non-uniform behaviour. By comparing predicted values with actual observations, dynamic thresholds can be established, preventing unnecessary alerts during periods of heightened or reduced activity.

Case Study 2: Adaptive Thresholds in Dynamic Environments

Context: Imagine a cloud-based service with varying workloads throughout the day. Determining thresholds in such dynamic environments demands continuous adaptation to ensure optimal performance and resource utilisation.

Machine Learning Approach: Ensemble methods, particularly Random Forests, prove invaluable in this context. By training the model on historical data encompassing diverse workload scenarios, the Random Forest learns intricate relationships between system metrics and workload variations.

# Sample code for Random Forest in dynamic workload scenarios
from sklearn.ensemble import RandomForestRegressor

# Assuming 'features' and 'targets' are your dataset columns
model = RandomForestRegressor()
model.fit(X_train, y_train)

# Make predictions for current workload
predicted_workload = model.predict(current_features)

Results: The adaptability of Random Forests shines as workload patterns evolve. The model, having learned from historical data, accurately predicts system metrics under varying workloads. This predictive capability enables the establishment of dynamic thresholds that align with the current workload, optimizing resource allocation and enhancing system performance.

Discussion: These case studies illuminate the dynamic and adaptive nature of machine learning in threshold definition. The ability to capture nuanced patterns, whether in non-uniform user activity or dynamic workload scenarios, empowers data analysts and scientists to move beyond static thresholds. Machine learning becomes a strategic ally, offering precision and adaptability crucial for modern system monitoring challenges.

Conclusion

Learning from Percentiles:
I started by understanding how percentiles work, thanks to Bakare’s detailed explanation. While this gave me a good foundation, I realised that in the dynamic field of data analytics, I need more advanced methods to handle diverse patterns in the systems I monitor.

Bayesian Adventure:
I then entered the world of Bayesian statistics, where probability and prior knowledge come together. Bayesian methods offer a way to learn iteratively, but I also acknowledged the challenges, such as the need for careful computation and dealing with subjective beliefs.

Power of Hypothesis Testing:
Hypothesis testing became a valuable tool, capable of uncovering hidden patterns in different types of systems. Bakare’s insights into deterministic and non-deterministic systems showed how adaptable and versatile hypothesis testing can be, even in non-uniform and dynamic environments.

Machine Learning Unleashed:
The article reached its peak with the integration of machine learning. From simple regression models that highlight relationships to ensemble methods that create robust predictions, machine learning marks a new era. I emphasised the importance of understanding how these models make predictions, echoing Bakare’s exploration of prediction intervals.

Real-World Stories:
I shared two case studies illustrating how machine learning impacts system thresholds. One showed the adaptability of an LSTM neural network in handling non-uniform user activity, while the other demonstrated how Random Forests adjust thresholds based on dynamic workload variations.

Precision in Complexity:
As I wrap up, the message is clear: precision in defining thresholds is not just a goal but a necessity. The traditional approach of relying solely on percentiles is no longer sufficient in the ever-changing world of data.

A thank-you note goes to Emmanuel Bakare for laying the groundwork. Fueled by curiosity, everyone should be on a journey beyond the familiar. The evolving field of analytics awaits more discoveries and stories. It’s not just for progress; it’s transformation into becoming a master.

--

--

GOKE ADEKUNLE; #Wolfwords

At the intersection of Payments, Data Science, Finance, Psychology, Artificial Intelligence, Arts, and Business.