Tutorial: Building regression models

You must create three tables to load the three public datasets into Amazon Redshift. The datasets are Toronto Bike Ridership Data, historical weather data, and historical holidays data. Run the following query in the Amazon Redshift query editor to create tables named ridership, weather, and holiday.

CREATE TABLE IF NOT EXISTS ridership (
    trip_id INT,
    trip_duration_seconds INT,
    trip_start_time timestamp,
    trip_stop_time timestamp,
    from_station_name VARCHAR(50),
    to_station_name VARCHAR(50),
    from_station_id SMALLINT,
    to_station_id SMALLINT,
    user_type VARCHAR(20)
);

CREATE TABLE IF NOT EXISTS weather (
    longitude_x DECIMAL(5, 2),
    latitude_y DECIMAL(5, 2),
    station_name VARCHAR(20),
    climate_id BIGINT,
    datetime_utc TIMESTAMP,
    weather_year SMALLINT,
    weather_month SMALLINT,
    weather_day SMALLINT,
    time_utc VARCHAR(5),
    temp_c DECIMAL(5, 2),
    temp_flag VARCHAR(1),
    dew_point_temp_c DECIMAL(5, 2),
    dew_point_temp_flag VARCHAR(1),
    rel_hum SMALLINT,
    rel_hum_flag VARCHAR(1),
    precip_amount_mm DECIMAL(5, 2),
    precip_amount_flag VARCHAR(1),
    wind_dir_10s_deg VARCHAR(10),
    wind_dir_flag VARCHAR(1),
    wind_spd_kmh VARCHAR(10),
    wind_spd_flag VARCHAR(1),
    visibility_km VARCHAR(10),
    visibility_flag VARCHAR(1),
    stn_press_kpa DECIMAL(5, 2),
    stn_press_flag VARCHAR(1),
    hmdx SMALLINT,
    hmdx_flag VARCHAR(1),
    wind_chill VARCHAR(10),
    wind_chill_flag VARCHAR(1),
    weather VARCHAR(10)
);

CREATE TABLE IF NOT EXISTS holiday (holiday_date DATE, description VARCHAR(100));

The following query loads the sample data into the tables that you created in the previous step.

COPY ridership
FROM
    's3://redshift-ml-bikesharing-data/bike-sharing-data/ridership/' 
    IAM_ROLE default 
    FORMAT CSV 
    IGNOREHEADER 1 
    DATEFORMAT 'auto' 
    TIMEFORMAT 'auto' 
    REGION 'us-west-2' 
    gzip;

COPY weather
FROM
    's3://redshift-ml-bikesharing-data/bike-sharing-data/weather/' 
    IAM_ROLE default 
    FORMAT csv 
    IGNOREHEADER 1 
    DATEFORMAT 'auto' 
    TIMEFORMAT 'auto' 
    REGION 'us-west-2' 
    gzip;

COPY holiday
FROM
    's3://redshift-ml-bikesharing-data/bike-sharing-data/holiday/' 
    IAM_ROLE default 
    FORMAT csv 
    IGNOREHEADER 1 
    DATEFORMAT 'auto' 
    TIMEFORMAT 'auto' 
    REGION 'us-west-2' 
    gzip;

The following query performs transformations on the ridership and weather datasets to remove bias or anomalies. Removing bias and anomalies results in improved model accuracy. The query simplifies the tables by creating two new views called ridership_view and weather_view.

CREATE
OR REPLACE VIEW ridership_view AS
SELECT
    trip_time,
    trip_count,
    TO_CHAR(trip_time, 'hh24') :: INT trip_hour,
    TO_CHAR(trip_time, 'dd') :: INT trip_day,
    TO_CHAR(trip_time, 'mm') :: INT trip_month,
    TO_CHAR(trip_time, 'yy') :: INT trip_year,
    TO_CHAR(trip_time, 'q') :: INT trip_quarter,
    TO_CHAR(trip_time, 'w') :: INT trip_month_week,
    TO_CHAR(trip_time, 'd') :: INT trip_week_day
FROM
    (
        SELECT
            CASE
                WHEN TRUNC(r.trip_start_time) < '2017-07-01' :: DATE THEN CONVERT_TIMEZONE(
                    'US/Eastern',
                    DATE_TRUNC('hour', r.trip_start_time)
                )
                ELSE DATE_TRUNC('hour', r.trip_start_time)
            END trip_time,
            COUNT(1) trip_count
        FROM
            ridership r
        WHERE
            r.trip_duration_seconds BETWEEN 60
            AND 60 * 60 * 24
        GROUP BY
            1
    );

CREATE
OR REPLACE VIEW weather_view AS
SELECT
    CONVERT_TIMEZONE(
        'US/Eastern',
        DATE_TRUNC('hour', datetime_utc)
    ) daytime,
    ROUND(AVG(temp_c)) temp_c,
    ROUND(AVG(precip_amount_mm)) precip_amount_mm
FROM
    weather
GROUP BY
    1;

The following query creates a table that combines all the relevant input attributes from ridership_view and weather_view into the trip_data table.

CREATE TABLE trip_data AS
SELECT
    r.trip_time,
    r.trip_count,
    r.trip_hour,
    r.trip_day,
    r.trip_month,
    r.trip_year,
    r.trip_quarter,
    r.trip_month_week,
    r.trip_week_day,
    w.temp_c,
    w.precip_amount_mm,CASE
        WHEN h.holiday_date IS NOT NULL THEN 1
        WHEN TO_CHAR(r.trip_time, 'D') :: INT IN (1, 7) THEN 1
        ELSE 0
    END is_holiday,
    ROW_NUMBER() OVER (
        ORDER BY
            RANDOM()
    ) serial_number
FROM
    ridership_view r
    JOIN weather_view w ON (r.trip_time = w.daytime)
    LEFT OUTER JOIN holiday h ON (TRUNC(r.trip_time) = h.holiday_date);

The following query splits your data into a training set and a validation set by designating 80% of the dataset for training and 20% for validation. The training set is the input for the ML model to identify the best possible algorithm for the model. After the model is created, you use the validation set to validate the model accuracy.

CREATE TABLE training_data AS
SELECT
    trip_count,
    trip_hour,
    trip_day,
    trip_month,
    trip_year,
    trip_quarter,
    trip_month_week,
    trip_week_day,
    temp_c,
    precip_amount_mm,
    is_holiday
FROM
    trip_data
WHERE
    serial_number > (
        SELECT
            COUNT(1) * 0.2
        FROM
            trip_data
    );

CREATE TABLE validation_data AS
SELECT
    trip_count,
    trip_hour,
    trip_day,
    trip_month,
    trip_year,
    trip_quarter,
    trip_month_week,
    trip_week_day,
    temp_c,
    precip_amount_mm,
    is_holiday,
    trip_time
FROM
    trip_data
WHERE
    serial_number <= (
        SELECT
            COUNT(1) * 0.2
        FROM
            trip_data
    );

The following query creates a regression model to predict the trip_count value for any input date and time. In the following example, replace DOC-EXAMPLE-BUCKET with your own S3 bucket.

CREATE MODEL predict_rental_count
FROM
    training_data TARGET trip_count FUNCTION predict_rental_count 
    IAM_ROLE default 
    PROBLEM_TYPE regression 
    OBJECTIVE 'mse' 
    SETTINGS (
        s3_bucket 'DOC-EXAMPLE-BUCKET',
        s3_garbage_collect off,
        max_runtime 5000
    );

Use the following query to output aspects of the model, and find the mean square error metric in the output. Mean square error is a typical accuracy metric for regression problems.

show model predict_rental_count;

Run the following prediction queries against the validation data to compare the predicted trip count to the actual trip count.

SELECT
    trip_time,
    actual_count,
    predicted_count,
    (actual_count - predicted_count) difference
FROM
    (
        SELECT
            trip_time,
            trip_count AS actual_count,
            PREDICT_RENTAL_COUNT (
                trip_hour,
                trip_day,
                trip_month,
                trip_year,
                trip_quarter,
                trip_month_week,
                trip_week_day,
                temp_c,
                precip_amount_mm,
                is_holiday
            ) predicted_count
        FROM
            validation_data
    )
LIMIT
    5;

The following query calculates the mean square error and root mean square error based on your validation data. You use mean square error and root mean square error to measure the distance between the predicted numeric target and the actual numeric answer. A good model has a low score in both metrics. The following query returns the value of both metrics.

SELECT
    ROUND(
        AVG(POWER((actual_count - predicted_count), 2)),
        2
    ) mse,
    ROUND(
        SQRT(AVG(POWER((actual_count - predicted_count), 2))),
        2
    ) rmse
FROM
    (
        SELECT
            trip_time,
            trip_count AS actual_count,
            PREDICT_RENTAL_COUNT (
                trip_hour,
                trip_day,
                trip_month,
                trip_year,
                trip_quarter,
                trip_month_week,
                trip_week_day,
                temp_c,
                precip_amount_mm,
                is_holiday
            ) predicted_count
        FROM
            validation_data
    );

The following query calculates the percent error in trip count for each trip time on 2017-01-01. The query orders the trip times from the time with the lowest percent error to the time with the highest percent error.

SELECT
    trip_time,
    CAST(ABS(((actual_count - predicted_count) / actual_count)) * 100 AS DECIMAL (7,2)) AS pct_error
FROM
    (
        SELECT
            trip_time,
            trip_count AS actual_count,
            PREDICT_RENTAL_COUNT (
                trip_hour,
                trip_day,
                trip_month,
                trip_year,
                trip_quarter,
                trip_month_week,
                trip_week_day,
                temp_c,
                precip_amount_mm,
                is_holiday
            ) predicted_count
        FROM
            validation_data
    )
WHERE
   trip_time LIKE '2017-01-01 %%:%%:%%'
ORDER BY
   2 ASC;