4.4.1: Implement a neuron for linear regression - Training data and outliers

Key Takeaways

foreign you'll put your knowledge of neurons into practice to create a tensorflow.js web app that will train a custom model live in your web browser so first let's define a problem you'll be trying to solve imagine a customer comes to you and says I'd like to create a web app that could estimate the house price of properties in a certain City all right well this sounds like something you should probably try and make with your existing knowledge in your mind you know that property prices might be linearly correlated with certain features like the size of the house but first you'll need some data to see if this works well now after asking the customer for more information you find they've gathered some example data of houses in the area here's an example of the data provided now the full size of the table is over 2 000 rows of data so it's shortened here but essentially it contains size in square feet number of bedrooms and the value of the house in dollars seems like there's enough data to give it a try at this step you may want to check for data to ensure there are no potential issues with how it's been recorded like missing values there are many ways to do this of course if there's a small amount of data like you have here you can visually check the columns that a value is in each or write a small bit of code to show the min max values and check for null values as a starting point there isn't third-party libraries that exists that can do much of his work for you and much more but those libraries are outside the scope of this course however if you're interested you might want to check out another Library called danfo.js that makes working of data easier in JavaScript that you could use as part of your tensorflow.js pipeline for processing data it tries to replicate a library known as pandas which python developers quite often use in data science in order to understand their input data features better for this exercise the data is already in good shape so let's start analyzing what you have now you can start by visualizing the correlation between each column of data that you have and the target value you're trying to predict here you want to predict house price so let's plot all the other values versus the house price to see how they relate first you plot size versus the value of the house as shown here now how did I draw this graph well you can use any graphing library of your choice to plot a chart like this I made these Scatter Plots using a library called plotly which is well suited for JavaScript and is built upon the very popular and Powerful d3.js library for data visualization it should be noted that tensorflow.js has a simple visualization number you can use too that you can import via a script tag check for links for more information so back to the data bin it seems that generally speaking ignoring some of the outliers highlighted the larger the house the more its value which is somewhat as expected you can also see that these features could be correlated with a straight line so it's well suited for linear regression which is what we're going to be doing next you can plot bedrooms versus Cell price which also shows a linear correlation between value and number of bedrooms it should be noted that you can see the variance quite nicely here you can see that for each bedroom level the house values can differ with a range of about a hundred thousand dollars in value again there are some fairly clear outliers which you may want to consider dropping from your data set to achieve better results now there are accurate mathematical ways of finding which points constitute to being outliers which you can then safely remove but again that is beyond the scope of this course and we'll be diving deeper into the world of data science instead however a quick search will provide many resources on this subject if you're interested that you can then replicate in JavaScript as needed intuitively then if your chosen line is going through the central points of those distributions the maximum error you would expect would be around fifty thousand dollars either above or below what the line itself predicts so hopefully your train model will have an average error lower than this okay so given the strong correlation in both chosen features along with the fact that they both seem to be linear in nature they should be well suited for a single neuron to learn from so now head on to the next section where it's time to start coding [Music] [Music]