Thomas Vladeck’s Post

If you were to design a statistical problem in a lab, I doubt you could come up with anything more challenging than media mix modeling. So, how do we tackle it here at Recast? Statistical research is at the core of what we do, and it’s reflected across our organization: over 30% of our team has a PhD in either math or statistics, and I wouldn't be surprised if we had the most complex Bayesian model in the world. We take a lot of pride in our model, being research-focused, compute-heavy, and having a great team of scientists. But one of the challenges in media mix modeling is that you’re communicating to marketers, and they have to use the model to make day-to-day decisions about where they're going to allocate spend. And not just that – they also need to be able to communicate their rationale to their own stakeholders and executive teams. So, while we’re scientists, they’re often not. And so the results of our model need to be communicated in an accessible way that is easy to understand, use, and relay within their organization. Not only do we have to be the best in the world at creating a statistical computing engine to estimate models of this complexity, but we also have to break it down into digestible insights and recommendations. We work very closely with our customers every day to make sure that the information that we're learning about their marketing mix is being communicated effectively and that they can understand the work that we're doing. That’s why we say Recast is built *by* scientists *for* marketers. To learn more about Recast, check us out here: https://lnkd.in/eUpbB2BR

This is an important article to measure the central tendency of data in a distribution. Use cases: 1. Training vs OOT dataset 2. Champion vs Challenger datasets in A/B testing 3. Supervised (using look alikes) vs Unsupervised (clustering)

I am excited to share my latest article, "A Comprehensive Introduction to Generalized Linear Models"! In this deep dive, we unravel the versatility and power of GLMs, shedding light on their significance in statistical analysis across various fields. From choosing the right link function to overcoming common challenges, this guide has it all. Dive in to enhance your analytical toolkit! #datascience #statistics #statisticalmodelling

A strong recommendation to read: "Statistical Modeling: The Two Cultures" by Leo Breiman (https://lnkd.in/gC5ggiS9) It is an inspiring discussion about the difference between a top-down approach, where we assume to know the model that has given rise to our data (e.g., assuming the data follow a normal distribution or a relationship between two variables can be expressed by linear regression), and a bottom-up approach, where the data generating mechanism is assumed unknown and an algorithm learns the model. It can also be perceived as a comparison between statistics and machine learning. What are the similarities and the differences, where is the boundary? I have put quite some thoughts into this issue, inspired inter alia by this paper, and would be intrigued to get your opinions on the comparison. Stéphane P.A. Bordas (FLSW) Florian Felice Stijn Vansteelandt Georgi Baklicharov Andreas Groll Christoph Molnar #statistics #machinelearning #data #datascience #datamodeling #algorithms

This is actually quite an idea. Viable and controllable way of predictive modelling.

Histogram as a machine learning model: ·  A histogram is a graph that visualizes the frequency of numerical data and is used extensively in data science and statistics. But what does it do mathematically?   ·   The histogram of a sample (or dataset) is supposed to mimic the shape of the PDF of the population from which the sample is drawn. Hence, we can think of a histogram as a machine learning model that learns the PDF of the population based on the sample that we have drawn from that population. This is a very simple model since it only has one hyperparameter which is the bin width (h). This hyperparameter controls the complexity of this model.   ·   Like any other machine learning model, we can calculate the bias and variance of a histogram. The mean squared error of the model (MSE) is the sum of variance and the square of bias in the absence of any noise in the sample (MSE = Bias^2 + Variance). As h goes to zero, the model becomes more complicated. Hence the absolute value of bias decreases and the variance increases (overfitting). On the other hand, when h increases, the model becomes simpler, so the variance decreases, but the absolute value of bias increases (underfitting). In both cases, the MSE increases.   ·   The right value for (h*) is the value at which the model is neither too simple nor too complex, so the average MSE is minimized. Since we don’t have the population from which the sample is drawn, we can only estimate h*. Scott’s rule estimates h* assuming that the population has a normal distribution. There are some other rules like Sturges' formula, Rice rule and Doane's formula.   ·   My recent article in Towards Data Science discusses the math behind histograms and kernel density estimation. It shows how a histogram is mathematically is mathematically constructed, and how the bias and variance of a histogram can be calculated. Here is the link: https://lnkd.in/gY_zxnDV #machinelearning  #data #analytics #datascience #artificialintelligence  #probability #statistics #mathematics #statisticalanalysis

🚀 Day 12 of #100DaysOfDataScience 🚀 While still working to improve my prediction model for house prices, I began breaking down columns with categorical data into numerical columns using the One-hot encoding method, the number of columns in my dataset grew and I started thinking, wouldn't this be a problem for the model? is there a limit to the number of columns I can select as features to train the model? Did some research and found out about the "Curse of Dimentionality" : This means that model might become computationally expensive and prone to overfitting. How do I go around this? Did some more research and found out a few considerations and potential strategies to manage high-dimensional data after one-hot encoding: 1. Feature Selection: This technique suggest that only the most informative features are used while the least relevant features are discarded. Techniques like mutual information, recursive feature elimination, or L1 regularization can be used to select a subset of important features. 2. Feature Extraction: Involves using dimensionality reduction techniques like Principal Component Analysis (PCA) or t-distributed Stochastic Neighbor Embedding (t-SNE) to reduce the dimensionality of the dataset while retaining most of the information. 3. Domain Knowledge: Domain expertise definately help to spot features that might not be that neccesary. 4. Tree-Based Models: Tree-based algorithms, such as decision trees and random forests, can handle high-dimensional data effectively and automatically select the most important features during training. 5. Gradient Boosting: Gradient boosting models like XGBoost and LightGBM can handle high-dimensional data effectively while avoiding overfitting. By combining Feature selection, Domain knowledge, and a Tree-based model (Random forest), I was able to break the "Curse of Dimentionality". I hope you found this helpful. I'd be happy to hear your thoughts 😄

I made a new post on medium where I use a pulsar classification dataset to talk about the concepts of ranking and lift and how they can be useful in problems where classification is difficult and/or you have limited resources to allocate. Ranking can also be useful when you want to produce a 'score' to use in a different model (persuasion analytics), and while I mention that in the article I don't really expound on it. Constructive comments and feedback are always welcome. https://lnkd.in/ewpfumXj

🚀 **Predicting Calories Burnt using Machine Learning 🤖🔥** Excited to share my latest machine learning project where I delved into predicting calories burnt based on user exercise data! 🏋️♂️📊 Here's a brief journey through the key steps and findings: 1. Data Collection & Processing: 📈🔍 - Imported necessary libraries for analysis and ML. - Merged 'calories.csv' and 'exercise.csv' to create a unified dataframe. - Checked data shape, info, and handled null values. 2. Data Analysis & Visualization: 📊👀 - Explored statistical measures and visualized gender distribution. - Analyzed skewness and plotted PDF and QQ plots for features like Age, Height, Weight, Duration, Heart Rate, Body Temp, and Calories. 3. Correlation Analysis: 🌐🔗 - Constructed a heatmap, revealing strong correlations between Calories and features like Duration, Heart Rate, and Body Temperature. 4. Data Preprocessing: 🧹🔄 - Converted categorical 'Gender' values to numerical. - Employed IQR to eliminate outliers. 5. Model Building & Hyperparameter Tuning: 🤖🎯 - Defined features and target variable. - Split data, applied Standard Scaling. - Utilized XGBoost Regressorwith Randomized Search CV for hyperparameter tuning. 6. Model Evaluation: 📊🔄 - Predicted calories on the test set. - Evaluated the model using MAE, R-squared, and RMSE. - Visualized results with Actual vs Predicted plots, confirming model accuracy. 7. Conclusion & Recommendations: 🤝🔍 - The XG-Boost Regressor model, post hyperparameter tuning, displayed strong performance. - Utilized visual aids for robust data understanding and correlation identification. - Recommendations for further exploration, feature engineering, and potential improvements. 8. Future Work: 🚀🔮 - Explore alternative ML models for comparison. - Incorporate user feedback and domain knowledge for enhanced accuracy. Thrilled about the results and the potential for further refinement! 💡💪 If you're passionate about data science or have insights to share, I'd love to connect and hear your thoughts. Let's keep pushing the boundaries of what's possible in the world of machine learning! 🚀🌐 #MachineLearning #DataScience #CaloriesPrediction #XGBoost #DataAnalysis #LinkedInPost #DataDrivenDecisions

🚀 Exploring Linear Regression: OLS vs. Gradient Descent 📊 In the dynamic world of data science and machine learning, choosing the right algorithm is crucial. Today, let's dive into a debate that often arises in the realm of linear regression: Ordinary Least Squares (OLS) vs. Gradient Descent. 🔍 OLS - The Classic Approach Ordinary Least Squares is like the seasoned veteran of linear regression. It minimizes the sum of squared differences between actual and predicted values, providing a closed-form solution. OLS is elegant, easy to implement, and ideal for small to moderately sized datasets. It's a go-to choice when assumptions like linearity and homoscedasticity are met. 🔄 Gradient Descent - The Versatile Challenger Enter Gradient Descent, a more dynamic and flexible approach. It iteratively adjusts parameters to minimize the cost function, making it suitable for large datasets and non-linear problems. Gradient Descent excels in scenarios where OLS may struggle, offering efficiency and scalability. 🌟 The Verdict? It Depends! There's no one-size-fits-all answer. OLS might be the champion in certain scenarios, providing analytical solutions efficiently. On the other hand, Gradient Descent shines when dealing with massive datasets and complex models. 🤔 So, which to choose? Consider the size and nature of your dataset. If it's relatively small and meets OLS assumptions, it might be your go-to. For larger datasets or complex models, where efficiency and scalability matter, Gradient Descent could be the hero. 💡 Key Takeaway: Understanding the strengths and weaknesses of OLS and Gradient Descent is crucial. In the ever-evolving landscape of data science, adaptability, and a nuanced approach often lead to the best outcomes. #DataScience #MachineLearning #LinearRegression #Algorithm #GradientDescent #OLS #Analytics #TechTalk 🚀