In this notebook, we will explore data corresponding to taxi rides in New York City to build a Machine Learning model in support of a fare-estimation tool. The idea is to suggest a likely fare to taxi riders so that they are not surprised, and so that they can protest if the charge is much higher than expected.
Let's start off with the Python imports that we need.
In [1]:
from google.cloud import bigquery
import seaborn as sns
import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
import shutil
The dataset that we will use is a BigQuery public dataset. Click on the link, and look at the column names. Switch to the Details tab to verify that the number of records is one billion, and then switch to the Preview tab to look at a few rows.
Let's write a SQL query to pick up interesting fields from the dataset.
In [3]:
sql = """
SELECT
pickup_datetime, pickup_longitude, pickup_latitude, dropoff_longitude,
dropoff_latitude, passenger_count, trip_distance, tolls_amount,
fare_amount, total_amount
FROM `nyc-tlc.yellow.trips`
LIMIT 10
"""
In [4]:
client = bigquery.Client()
trips = client.query(sql).to_dataframe()
trips
Out[4]:
Let's increase the number of records so that we can do some neat graphs. There is no guarantee about the order in which records are returned, and so no guarantee about which records get returned if we simply increase the LIMIT. To properly sample the dataset, let's use the HASH of the pickup time and return 1 in 100,000 records -- because there are 1 billion records in the data, we should get back approximately 10,000 records if we do this.
In [5]:
sql = """
SELECT
pickup_datetime,
pickup_longitude, pickup_latitude,
dropoff_longitude, dropoff_latitude,
passenger_count,
trip_distance,
tolls_amount,
fare_amount,
total_amount
FROM
`nyc-tlc.yellow.trips`
WHERE
ABS(MOD(FARM_FINGERPRINT(CAST(pickup_datetime AS STRING)), 100000)) = 1
"""
In [6]:
trips = client.query(sql).to_dataframe()
trips[:10]
Out[6]:
Let's explore this dataset and clean it up as necessary. We'll use the Python Seaborn package to visualize graphs and Pandas to do the slicing and filtering.
In [7]:
ax = sns.regplot(x="trip_distance", y="fare_amount", fit_reg=False, ci=None, truncate=True, data=trips)
ax.figure.set_size_inches(10, 8)
Hmm ... do you see something wrong with the data that needs addressing?
It appears that we have a lot of invalid data that is being coded as zero distance and some fare amounts that are definitely illegitimate. Let's remove them from our analysis. We can do this by modifying the BigQuery query to keep only trips longer than zero miles and fare amounts that are at least the minimum cab fare ($2.50).
Note the extra WHERE clauses.
In [8]:
sql = """
SELECT
pickup_datetime,
pickup_longitude, pickup_latitude,
dropoff_longitude, dropoff_latitude,
passenger_count,
trip_distance,
tolls_amount,
fare_amount,
total_amount
FROM
`nyc-tlc.yellow.trips`
WHERE
ABS(MOD(FARM_FINGERPRINT(CAST(pickup_datetime AS STRING)), 100000)) = 1
AND trip_distance > 0 AND fare_amount >= 2.5
"""
In [9]:
trips = client.query(sql).to_dataframe()
ax = sns.regplot(x="trip_distance", y="fare_amount", fit_reg=False, ci=None, truncate=True, data=trips)
ax.figure.set_size_inches(10, 8)
What's up with the streaks at \$45 and \$50? Those are fixed-amount rides from JFK and La Guardia airports into anywhere in Manhattan, i.e. to be expected. Let's list the data to make sure the values look reasonable.
Let's examine whether the toll amount is captured in the total amount.
In [10]:
tollrides = trips[trips['tolls_amount'] > 0]
tollrides[tollrides['pickup_datetime'] == '2010-04-29 12:28:00']
Out[10]:
Looking a few samples above, it should be clear that the total amount reflects fare amount, toll and tip somewhat arbitrarily -- this is because when customers pay cash, the tip is not known. So, we'll use the sum of fare_amount + tolls_amount as what needs to be predicted. Tips are discretionary and do not have to be included in our fare estimation tool.
Let's also look at the distribution of values within the columns.
In [11]:
trips.describe()
Out[11]:
Hmm ... The min, max of longitude look strange.
Finally, let's actually look at the start and end of a few of the trips.
In [12]:
def showrides(df, numlines):
lats = []
lons = []
for iter, row in df[:numlines].iterrows():
lons.append(row['pickup_longitude'])
lons.append(row['dropoff_longitude'])
lons.append(None)
lats.append(row['pickup_latitude'])
lats.append(row['dropoff_latitude'])
lats.append(None)
sns.set_style("darkgrid")
plt.figure(figsize=(10,8))
plt.plot(lons, lats)
showrides(trips, 10)
In [13]:
showrides(tollrides, 10)
As you'd expect, rides that involve a toll are longer than the typical ride.
We need to some clean-up of the data:
We could do preprocessing in BigQuery, similar to how we removed the zero-distance rides, but just to show you another option, let's do this in Python. In production, we'll have to carry out the same preprocessing on the real-time input data.
This sort of preprocessing of input data is quite common in ML, especially if the quality-control is dynamic.
In [14]:
def preprocess(trips_in):
trips = trips_in.copy(deep=True)
trips.fare_amount = trips.fare_amount + trips.tolls_amount
del trips['tolls_amount']
del trips['total_amount']
del trips['trip_distance']
del trips['pickup_datetime']
qc = np.all([\
trips['pickup_longitude'] > -78, \
trips['pickup_longitude'] < -70, \
trips['dropoff_longitude'] > -78, \
trips['dropoff_longitude'] < -70, \
trips['pickup_latitude'] > 37, \
trips['pickup_latitude'] < 45, \
trips['dropoff_latitude'] > 37, \
trips['dropoff_latitude'] < 45, \
trips['passenger_count'] > 0,
], axis=0)
return trips[qc]
tripsqc = preprocess(trips)
tripsqc.describe()
Out[14]:
The quality control has removed about 300 rows (11400 - 11101) or about 3% of the data. This seems reasonable.
Let's move on to creating the ML datasets.
Let's split the QCed data randomly into training, validation and test sets.
In [15]:
shuffled = tripsqc.sample(frac=1)
trainsize = int(len(shuffled['fare_amount']) * 0.70)
validsize = int(len(shuffled['fare_amount']) * 0.15)
df_train = shuffled.iloc[:trainsize, :]
df_valid = shuffled.iloc[trainsize:(trainsize+validsize), :]
df_test = shuffled.iloc[(trainsize+validsize):, :]
In [16]:
df_train.describe()
Out[16]:
In [17]:
df_valid.describe()
Out[17]:
In [18]:
df_test.describe()
Out[18]:
Let's write out the three dataframes to appropriately named csv files. We can use these csv files for local training (recall that these files represent only 1/100,000 of the full dataset) until we get to point of using Dataflow and Cloud ML.
In [19]:
def to_csv(df, filename):
outdf = df.copy(deep=False)
outdf.loc[:, 'key'] = np.arange(0, len(outdf)) # rownumber as key
# reorder columns so that target is first column
cols = outdf.columns.tolist()
cols.remove('fare_amount')
cols.insert(0, 'fare_amount')
print (cols) # new order of columns
outdf = outdf[cols]
outdf.to_csv(filename, header=False, index_label=False, index=False)
to_csv(df_train, 'taxi-train.csv')
to_csv(df_valid, 'taxi-valid.csv')
to_csv(df_test, 'taxi-test.csv')
In [20]:
!head -10 taxi-valid.csv
In [21]:
!ls -l *.csv
We have 3 .csv files corresponding to train, valid, test. The ratio of file-sizes correspond to our split of the data.
In [22]:
%%bash
head taxi-train.csv
Looks good! We now have our ML datasets and are ready to train ML models, validate them and evaluate them.
Before we start building complex ML models, it is a good idea to come up with a very simple model and use that as a benchmark.
My model is going to be to simply divide the mean fare_amount by the mean trip_distance to come up with a rate and use that to predict. Let's compute the RMSE of such a model.
In [23]:
def distance_between(lat1, lon1, lat2, lon2):
# haversine formula to compute distance "as the crow flies". Taxis can't fly of course.
dist = np.degrees(np.arccos(np.minimum(1,np.sin(np.radians(lat1)) * np.sin(np.radians(lat2)) + np.cos(np.radians(lat1)) * np.cos(np.radians(lat2)) * np.cos(np.radians(lon2 - lon1))))) * 60 * 1.515 * 1.609344
return dist
def estimate_distance(df):
return distance_between(df['pickuplat'], df['pickuplon'], df['dropofflat'], df['dropofflon'])
def compute_rmse(actual, predicted):
return np.sqrt(np.mean((actual-predicted)**2))
def print_rmse(df, rate, name):
print ("{1} RMSE = {0}".format(compute_rmse(df['fare_amount'], rate*estimate_distance(df)), name))
FEATURES = ['pickuplon','pickuplat','dropofflon','dropofflat','passengers']
TARGET = 'fare_amount'
columns = list([TARGET])
columns.extend(FEATURES) # in CSV, target is the first column, after the features
columns.append('key')
df_train = pd.read_csv('taxi-train.csv', header=None, names=columns)
df_valid = pd.read_csv('taxi-valid.csv', header=None, names=columns)
df_test = pd.read_csv('taxi-test.csv', header=None, names=columns)
rate = df_train['fare_amount'].mean() / estimate_distance(df_train).mean()
print ("Rate = ${0}/km".format(rate))
print_rmse(df_train, rate, 'Train')
print_rmse(df_valid, rate, 'Valid')
print_rmse(df_test, rate, 'Test')
The RMSE depends on the dataset, and for comparison, we have to evaluate on the same dataset each time. We'll use this query in later labs:
In [24]:
def create_query(phase, EVERY_N):
"""
phase: 1=train 2=valid
"""
base_query = """
SELECT
(tolls_amount + fare_amount) AS fare_amount,
CONCAT(CAST(pickup_datetime AS STRING), CAST(pickup_longitude AS STRING), CAST(pickup_latitude AS STRING), CAST(dropoff_latitude AS STRING), CAST(dropoff_longitude AS STRING)) AS key,
EXTRACT(DAYOFWEEK FROM pickup_datetime)*1.0 AS dayofweek,
EXTRACT(HOUR FROM pickup_datetime)*1.0 AS hourofday,
pickup_longitude AS pickuplon,
pickup_latitude AS pickuplat,
dropoff_longitude AS dropofflon,
dropoff_latitude AS dropofflat,
passenger_count*1.0 AS passengers
FROM
`nyc-tlc.yellow.trips`
WHERE
trip_distance > 0
AND fare_amount >= 2.5
AND pickup_longitude > -78
AND pickup_longitude < -70
AND dropoff_longitude > -78
AND dropoff_longitude < -70
AND pickup_latitude > 37
AND pickup_latitude < 45
AND dropoff_latitude > 37
AND dropoff_latitude < 45
AND passenger_count > 0
"""
if EVERY_N == None:
if phase < 2:
# training
query = "{0} AND ABS(MOD(FARM_FINGERPRINT(CAST(pickup_datetime AS STRING)), 4)) < 2".format(base_query)
else:
query = "{0} AND ABS(MOD(FARM_FINGERPRINT(CAST(pickup_datetime AS STRING)), 4)) = {1}".format(base_query, phase)
else:
query = "{0} AND ABS(MOD(FARM_FINGERPRINT(CAST(pickup_datetime AS STRING)), {1})) = {2}".format(base_query, EVERY_N, phase)
return query
query = create_query(2, 100000)
df_valid = client.query(query).to_dataframe()
print_rmse(df_valid, 2.56, 'Final Validation Set')
The simple distance-based rule gives us a RMSE of $7.42. We have to beat this, of course, but you will find that simple rules of thumb like this can be surprisingly difficult to beat.
Let's be ambitious, though, and make our goal to build ML models that have a RMSE of less than $6 on the test set.
Copyright 2016 Google Inc. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.