Welcome to the PythonQL web demo

Collecting the full information about a customer journey through your services, and then analyzing the business from those journeys, is a powerful way to improve customer experience. A customer journey is a list of heterogeneous events with timestamps, so running analytics on it is challenging. Modeling the journeys as a table and using SQL does not quite work, and neither do typical data analysis frameworks. PythonQL handles this because it combines queries over nested data with the analytics that data science needs.

• q0Display dataset
• q1Compute statistics of loan defaults per state
• q2Count all new customers who hold a balance above 300 dollars for each state
• q3What fraction of former customers were denied a loan within one month of closing their account
• q4What percentage of late customers who received more than one call per month left the bank
• q5Compute the average burn rate across all customers
• q6Compare the average burn rate with the burn rate of customers who closed their accounts

Even though data is supposed to be clean and of top quality in the 21st century, that is not always the case. Cleaning and integrating dirty data is a common problem that falls onto data scientists, and it is especially painful when you need a quick assessment of a data source or want to know whether integrating it would be worth the trouble. Here we have a customer database and an order database we are considering integrating. The order data is questionable, and our task is to find out quickly whether it is useful to us.

• q0Display the datasets
• q1Count the exact number of distinct customers and of similar sounding names using soundex matching
• q2Check how many records from the master dataset will join with the order database
• q3Check what percentage of order amounts in the order dataset are malformed
• q4Check that each customer placed an order in a city where one of their addresses is registered
• q5Check what percentage of customers can be joined with valid, local order histories
• q6Measure the percentage of total revenue from matched local orders, grouped by city

Getting the first results from your models is exciting, but a lot of time will be spent analyzing model performance and improving the results. Basic measures like AUC or MRSE are easy to calculate, but what happens when you need to compare two models qualitatively and figure out which cases are better covered by each model? The ad hoc analytical abilities of PythonQL come to the rescue. Here we are not querying a database, we are querying the results of our machine learning models.

• q0View the datasets
• q1Group errors by zip code and sort zip codes by error percentage
• q2Analyse errors through the matrix of city and crime rate, sorted by error percentage
• q3Analyse errors by distance from the nearest school, building percentile groups
• q4Sort cities by the correlation of the two models with coefficient and mean squared error
• q5Rank the matrix of cities and crime rates by the difference in mean root squared error
• q6Create rules for which model to pick depending on the city and crime rate combination

In this scenario we query JSON data that comes from a NoSQL database, join it to more structured relational data, and then construct nested JSON objects that use hierarchy to organize data. We have a JSON product catalog of cars with reviews, where the manufacturer, price and review fields can each be a plain value or a nested object, plus a small relational dataset with dealer information.

• q0Display the datasets
• q1How many cars of each manufacturer are there in the catalog
• q2Order products by the difference between suggested retail price and average selling price
• q3Order manufacturers by the average review rating
• q4Add reviews from the product database to the dealer offerings
• q5Order dealer offers by the difference between advertisement price and average selling price
• q6Build a nested JSON dataset combining dealer data with product reviews