Day #144- Uber Engineering - NLP, Machine Learning Notes

Uber engineering blog provides great insights on ML / Big Data / NLP Initiatives and guidelines. Sharing Some of Key Learning's from blog posts.

Link #1 - Forecasting at Uber

Need for Forecasting

• Identify High Demand Areas
• Predict User Supply Demand
• Identify Trends, Seasonality, Competition, Pricing
• Data Driven marketing decisions

Visualization Key Points

• Daily Trends
• Hourly Trends
• Weekly Trends
• Weekend Trends

Forecasting Techniques

• Statistical - ARIMA, Holt-Winter
• ML Based - RNN, Quantile Regression Forest (QRF), Gradient Boosting Trees (GBM), Support Vector Regression (SVR), Gaussian process regression (GP)

Summary

• Understand Trend, External Factors like Weather, Concerts
• Quantile Regression Forest (QRF) -  provide predictions at percentiles
• Gradient Boosting Trees (GBM) - a prediction model in the form of an ensemble of weak prediction models, typically decision trees 
• Support Vector Regression (SVR)
• Gaussian process regression (GP) - The prediction is probabilistic (Gaussian)
• Central Limit Theorem - Almost all measurable "random" variables in real world follow some kind of normal distribution
• Binomial - Only two outcomes of trail
• Poisson - Events that take place over and over again. Rate of Event denoted by lambda

Link #2 - COTA: Improving Uber Customer Care with NLP & Machine Learning

NLP Models built to interpret

• Phonology
• Morphology
• Grammar
• Syntax
• Semantics

Character level, word-level, pharase-level or sentence level or document level language modelling

Key Concepts in Implementation

• Fetch Trip Data
• Fetch Ticket Text
• Preprocessing (Stemming, Lowercasing, Stop-word removal, Lemmatization)
• Feature Engineering (LSI / TF_IDF)
• Cosine Similarity
• ML Algorithm for Ranking (Predictions - Issues / Solutions)
• Topic modeling transformation is carried out on the bag-of-word representation

Link #3 - Applying Customer Feedback: How NLP & Deep Learning Improve Uber’s Maps

Key Concepts

• Tickets encoded using One Hot Encoder
• Classification for type of ticket
• Alternatively used word2vec based representation
• WordCNN and LSTM based models

More Reads - Uber’s Big Data Platform: 100+ Petabytes with Minute Latency
Databook: Turning Big Data into Knowledge with Metadata at Uber

Happy Learning!!!