Simple example using House Pricing Scenario

[TylerBrinkley]

This is the first time I've looked into machine learning but I have a use case I'd like to test with it.

To get started I've created a simple example from the house pricing scenario which somewhat closely matches my use case but the results I'm getting are not at all close to what I expected. The data I'm providing is simply linear in terms of just the SqftLiving input parameter to the Price where Price = SqftLiving * 100. The SqftLot is held constant for training and prediction so it should be a non-factor.

I'm just trying to predict the Price when the SqftLiving is 1500 which with the linear model created by the provided data should make it about $150,000.

However, the results I get vary wildly from the negative to the postive 10's of millions every time I run the program which is unexpected. Could someone look into this simple example and let me know what if anything I'm doing is causing these poor results?

class Program
{
    static void Main(string[] args)
    {
        var filePath = "C://Temp/kc_house_data.csv";

        File.WriteAllText(filePath, @"100000,1000,8000
200000,2000,8000
400000,4000,8000");

        var pipeline = new LearningPipeline
        {
            new TextLoader<HousePriceData>(filePath, separator: ","),
            new ColumnConcatenator("Features", "SqftLiving", "SqftLot"),
            new StochasticDualCoordinateAscentRegressor()
        };

        var model = pipeline.Train<HousePriceData, HousePricePrediction>();

        var prediction = model.Predict(new HousePriceData { SqftLiving = 1500, SqftLot = 8000 });

        Console.WriteLine(prediction.Price);
        Console.ReadLine();
    }
}

public class HousePriceData
{
    [Column(ordinal: "0", name: "Label")]
    public float Price;

    [Column(ordinal: "1")]
    public float SqftLiving;

    [Column(ordinal: "2")]
    public float SqftLot;
}

public class HousePricePrediction
{
    [ColumnName("Score")]
    public float Price;
}

[Ivanidzo4ka]

Hello Tyler.
Let me point on the fact what you have only 3 data points, and our algorithms expected to be worked on millions of data points, with huge feature vector. For you as a person. it's obvious what formula states behind this 3 points, but from machine point of view, 3 points is not enough.

I've made following change in your code:

        Random random = new Random();
        StringBuilder sb = new StringBuilder();
        for ( int i = 0; i < n; i++)
        {
            var sqft = random.NextDouble();
            sb.AppendLine(string.Format("{0},{1},8000", sqft*10, sqft ));
        }
        File.WriteAllText(filePath, sb.ToString());

And I played with different values for n.
n=100 gives me prediction for your example around 141 000 which is about 6% error.
n=1000 gives me prediction 14910 which 0.6% error
n=100000 gives me 14992 which is 0.06% error
So the more data you give to algorithm, the better it performs.

You can also play with different learning models. For example if I modify my generating code to make
var sqft = random.NextDouble() *10_000;
to reflect real values for square foots I can use FastTreeRegressor, and it will give me reasonable prediction (2% error) even with 100 examples.

[TylerBrinkley]

@Ivanidzo4ka Thanks for the insight and the follow up with more data point results. I look forward to playing with this some more where I'll have a much larger data set.