Data drift and machine learning
Download as PPTX, PDF0 likes645 views
This document discusses machine learning methods and analysis. It provides an overview of machine learning, including that it allows computer programs to teach themselves from new data. The main machine learning techniques are described as supervised learning, unsupervised learning, and reinforcement learning. Popular applications of these techniques are also listed. The document then outlines the typical steps involved in applying machine learning, including data curation, processing, resampling, variable selection, building a predictive model, and generating predictions. It stresses that while data is important, the right analysis is also needed to apply machine learning effectively. The document concludes by discussing issues like data drift and how to implement validation and quality checks to safeguard automated predictions from such problems.
Data drift and machine learning
- 1. Machine Learning Methods and Analysis Smita Agrawal Data Scientist – Infosys
- 3. What is Machine Learning? Machine learning is simply a set of computer programs that can teach themselves to grow and adopt when exposed to new data Machine learning is broadly categorised into Supervised, Unsupervised and Reinforcement Learning techniques Supervised Learning Unsupervised Learning Reinforcement Learning Machine Learning Classification Churn Prediction Fraud Detection Image Classification Regression Market Mix Modelling ARPU Forecasting Life/Age Expectancy Advertising Popularity Prediction Dimensionality Reduction Meaningful Compression Feature Selection Clustering Customer Profiling Targeted Marketing Recommender Systems Real-time Decisions Robot Navigations Self-Learning Models The crux of Machine Learning lies in “History Repeats itself!”
- 4. How to be Machine Learning enabled? A Superficial View of steps 1. Data Curation Data curation involves gathering data across relevant attributes such that they can distinguish and thereby help us learn more about why something is happening 2. Processing Data 3. Resampling Resampling data so that data is free of any biased or over-powered characteristics of data are selected such that all the characteristics are balanced Leveraging measures like Mean, Median or Mode to process curated data so that any outliers or anomalies can be addressed 4. Variable Selection Not all the variables in the curated data really distinguish the characteristics. Variable selection is a critical step which enables in retaining these variables 5. Predictive Model A suitable Machine Learning can now be applied to the data obtained from step 4 to store the patterns observed 6. Generate Predictions This is the most exciting step as it is the stage where we can predict with certain confidence whether the data points belong to a characteristic Data is an oil to our Machine Learning’s engine. However, only oil will not ignite the engine! Data Oil with the right levers will ignite the engine and enable us in achieving our imperfectly perfect predictions!
- 5. Emergence of big data has created tremendous opportunities for businesses to gain real-time insights Make more informed decisions by leveraging data from the exploding number of digital systems However, as often is the case with disruptive technologies, the innovations behind big data have created a critical problem – one that we call Data Drift Data drift creates serious challenges to fully harness the insights available from big data Data drift is defined as: The unpredictable, unannounced and unending mutation of data characteristics caused by the operation, maintenance and modernization of the systems that produce the data Some Severe Impacts - Erodes data fidelity Operational reliability Ultimately the productivity of your data scientists and engineers It increases your costs Delays time to analysis Decreases the productivity and agility of your data engineers Leads to poor decision-making by data scientists and the line of business Source: Streamsets Data Drift
- 6. Fail Safe Mechanism - Prepping Validation for Automated Predictions 1. Ensure variable data types match in validation data with that in the historical data 2. Match all the unique values in categorical variables with the historical data to check for any new values that have come in the daily data 3. Measure the mean and most frequent value for numerical variables in historical and validation data 4. Any new values that are observed in categorical variables in the validation data will be handled as part of fail safe mechanism 5. There would be no predictions generated for these records as your predictive model does not know that new value as there were no instances of this value in the history when the model was built and causes a fatal error leading to no predictions 6. Generate a report for all the excluded records by variables to further analyse the trend drifts and enhance the model 7. Ongoing Data Quality Check: Check whether there is a behavioural shift of more than ± 5% in the modal value across variables to ensure that none of the data curation stage was broken which could have a consequent impact on the predictions 8. The variation should be compared to the historical data’s modal values 9. This should be followed as a real-time(daily) practice even while automatically updating the model
- 7. Some Common Data Drift Measures Population Stability Index Kolmogorov - Smirnov Statistic Kullback - Leibler Divergence Histogram Intersection 01 02 03 04
- 8. Smita Agrawal Data Scientist – Infosys Thank you!!!