In this lab, you apply nearest neighbors technique to help a taxi company predict the length of their rides. Imagine that we are hired to consult for LiftOff, a limo and taxi service that is just opening up in NYC. Liftoff wants it's taxi drivers to target longer rides, as the longer the ride the more money it makes. LiftOff has the following theory:
- The pickup location of a taxi ride can help predict the length of the ride.
LiftOff asks us to do some analysis to write a function that will allow it to predict the length of a taxi ride for any given location.
Our technique will be the following:
- Collect Obtain the data containing all of the taxi information, and only select the attributes of taxi trips that we need
- Explore Examine the attributes of our data, and plot some of our data on a map
- Train Write our nearest neighbors formula, and change the number of nearby trips to predict the length of a new trip
- Predict Use our function to predict trip lengths of new locations
Luckily for us, NYC Open Data collects information about NYC taxi trips and provides this data on its website.
For your reading pleasure, the data has already been downloaded into the trips.json file in this lab which you can find here. We'll use Python's json library to take the data from the trips.json file and store it as a variable in our notebook.
import json
# First, read the file
trips_file = open('trips.json')
# Then, convert contents to list of dictionaries
trips = json.load(trips_file)Press shift + enter
The next step is to explore the data. First, let's see how many trips we have.
len(trips)Not bad at all. Now let's see what each individual trip looks like. Each trip is a dictionary, so we can see the attributes of each trip with the keys function.
trips[0].keys()Ok, now that we have explored some of our data, let's begin to think through what data is relevant for our task.
Remember that our task is to use the trip location to predict the length of a trip. So let's select the pickup_latitude, pickup_longitude, and trip_distance from each trip. That will give us the trip location and related trip_distance for each trip. Then based on these actual trip distances we can use nearest neighbors to predict an expected trip distance for a trip, provided an actual location.
Add in about trip distance
Write a function called parse_trips(trips) that returns a list of the trips with only the following attributes:
trip_distancepickup_latitudepickup_longitude
def parse_trips(trips):
passparsed_trips = parse_trips(trips)
parsed_trips and parsed_trips[0]
# {'trip_distance': '18.379999999999999',
# 'pickup_latitude': '40.64499',
# 'pickup_longitude': '-73.781149999999997'}Now, there's just one change to make. If you look at one of the trips, all of the values are strings. Let's change them to be floats.
def float_values(trips):
passcleaned_trips = float_values(parsed_trips)cleaned_trips[0]
# {'trip_distance': 18.38,
# 'pickup_latitude': 40.64499,
# 'pickup_longitude': -73.78115}Now that we have paired down our data, let's get a sense of our trip data. We can use the folium Python library to plot a map of Manhattan, and our data. First we must import folium, and then use the Map function to pass through a location, and zoom_start. If a map isn't showing up below, copy and paste the command pip install folium into your terminal to install folium then try again.
import folium
manhattan_map = folium.Map(location=[40.7589, -73.9851], zoom_start=11)manhattan_mapOk, now let's see how we could add a dot to mark a specific location. We'll start with Times Square.
marker = folium.CircleMarker(location = [40.7589, -73.9851], radius=10)
marker.add_to(manhattan_map)Above, we first create a marker. Then we add that circle marker to the manhattan_map we created earlier.
manhattan_mapDo you see that blue dot near Time's Square? That is our marker.
So now that we can plot one marker on a map, we should have a sense of how we can plot many markers on a map to display our taxi ride data. We simply plot a map, and then we add a marker for each location of a taxi trip.
Now let's write some functions to allow us to plot maps and add markers a little more easily.
As a first step towards this, note that the functions to create both a marker and map each take in a location as two element list, representing the latitude and longitude values. Take another look:
marker = folium.CircleMarker(location = [40.7589, -73.9851])
manhattan_map = folium.Map(location=[40.7589, -73.9851])So let's write a function to create this two element list from a trip. Write a function called location that takes in a trip as an argument and returns a list where the first element is the latitude and the second is the longitude. Remember that a location looks like the following:
first_trip = {'pickup_latitude': 40.64499, 'pickup_longitude': -73.78115, 'trip_distance': 18.38}
first_tripdef location(trip):
passfirst_location = location(first_trip) # [40.64499, -73.78115]
first_location # [40.64499, -73.78115]Ok, now that we can turn a trip into a location, let's turn a location into a marker. Write a function called to_marker that takes in a location (in the form of a list) as an argument, and returns a folium circleMarker for that location. The radius of the marker should always equal 6.
def to_marker(location):
passimport json
times_square_marker = to_marker([40.7589, -73.9851])
times_square_marker and times_square_marker.location # [40.7589, -73.9851]times_square_marker and json.loads(times_square_marker.options)['radius'] # 6Ok, now that we know how to produce a single marker, let's write a function to produce lots. We can write a function called markers_from_trips that takes in a list of trips, and returns a marker object for each trip.
def markers_from_trips(trips):
passtrip_markers = markers_from_trips(cleaned_trips)cleaned_trips[0:4]trip_markers and len(trip_markers) # 1000list(map(lambda marker: marker.location, trip_markers[0:4]))
# [[40.64499, -73.78115],
# [40.766931, -73.982098],
# [40.77773, -73.951902],
# [40.795678, -73.971049]]Ok, now that we have a function that creates locations, and a function that creates markers, it is time to write a function to plot a map.
Write a function called map_from that, provided the first argument of a list location and second argument an integer representing the zoom_start, returns a folium map the corresponding location and zoom_start attributes.
Hint: The following is how to write a map with folium:
folium.Map(location=location, zoom_start=zoom_amount)
def map_from(location, zoom_amount):
passtimes_square_map = map_from([40.7589, -73.9851], 15)
times_square_map and times_square_map.location # [40.7589, -73.9851]times_square_marker and times_square_marker.add_to(times_square_map)
times_square_mapNow that we have a marker and a map, now let's write a function that adds a lot of markers to a map. This function should add each marker in the list to the map object then return the updated map object.
manhattan_map = map_from([40.7589, -73.9851], 13)def add_markers(markers, map_obj):
passmap_with_markers = add_markers(trip_markers, manhattan_map)map_with_markersOk, let's write a function that given a latitude and longitude will predict the distance for us. We'll do this by first finding the nearest trips given a latitude and longitude.
Here we once again apply the nearest neighbors formula. As a first step, write a function named distance_location that calculates the distance in pickup location between two trips.
import math
def distance_location(selected_trip, neighbor_trip):
passfirst_trip = {'pickup_latitude': 40.64499, 'pickup_longitude': -73.78115, 'trip_distance': 18.38}
second_trip = {'pickup_latitude': 40.766931, 'pickup_longitude': -73.982098, 'trip_distance': 1.3}
distance_first_and_second = distance_location(first_trip, second_trip)
distance_first_and_second and round(distance_first_and_second, 3) # 0.235Ok, next write a function called distance_between_neighbors that adds a new key-value pair, called distance_from_selected, that indicates the distance of the neighbor_trip from the selected_trip.
def distance_between_neighbors(selected_trip, neighbor_trip):
passdistance_between_neighbors(first_trip, second_trip)
# {'pickup_latitude': 40.766931,
# 'pickup_longitude': -73.982098,
# 'trip_distance': 1.3,
# 'distance_from_selected': 0.23505256047318146}Ok, now our neighbor_trip has another attribute called distance_from_selected, that indicates the distance from the neighbor_trip's pickup location from the selected_trip.
Understand the data: Our dictionary now has a few attributes, two of which say distance. Let's make sure we understand the difference.
distance_from_selected: This is our calculation of the distance of the neighbor's pickup location from the selected trip.trip_distance: This is the attribute we were provided initially. It tells us the length of the neighbor's taxi trip from pickup to drop-off.
Next, write a function called distance_all that provided a list of neighbors, returns each of those neighbors with their respective distance_from_selected numbers.
def distance_all(selected_individual, neighbors):
passcleaned_trips and distance_all(first_trip, cleaned_trips[0:4])Now write the nearest neighbors formula to calculate the distance of the selected_trip from all of the cleaned_trips in our dataset. If no number is provided, it should return the top 3 neighbors.
def nearest_neighbors(selected_trip, trips, number = 3):
passnew_trip = {'pickup_latitude': 40.64499,
'pickup_longitude': -73.78115,
'trip_distance': 18.38}
nearest_three_neighbors = nearest_neighbors(new_trip, cleaned_trips or [], number = 3)
nearest_three_neighbors
# [{'trip_distance': 7.78,
# 'pickup_latitude': 40.64483,
# 'pickup_longitude': -73.781578,
# 'distance_from_selected': 0.0004569288784918792},
# {'trip_distance': 12.7,
# 'pickup_latitude': 40.644657,
# 'pickup_longitude': -73.782229,
# 'distance_from_selected': 0.0011292165425673159},
# {'trip_distance': 17.3,
# 'pickup_latitude': 40.648509,
# 'pickup_longitude': -73.783508,
# 'distance_from_selected': 0.0042359798158141185}]Ok great! Now that we can provide a new trip location, and find the distances of the three nearest trips, we can take calculate an estimate of the trip distance for that new trip location.
We do so simply by calculating the average of it's nearest neighbors.
import statistics
def mean_distance(neighbors):
nearest_distances = list(map(lambda neighbor: neighbor['trip_distance'], neighbors))
return round(statistics.mean(nearest_distances), 3)
nearest_three_neighbors = nearest_neighbors(new_trip, cleaned_trips or [], number = 3)
distance_estimate_of_selected_trip = mean_distance(nearest_three_neighbors) # 12.7
distance_estimate_of_selected_tripNow, as we know from the last lesson, one tricky element is to determine how many neighbors to choose, our
- If we choose a
$k$ value too low, our formula will be too heavily influenced by a single neighbor, whereas if our$k$ value is too high, we will be choosing so many neighbors that our nearest neighbors formula will not be adjust enough according to locations.
Ok, let's experiment with this.
First, let's choose a midtown location, to see what the trip distance would be. A Google search reveals the coordinates of 51st and 7th avenue to be the following.
midtown_trip = dict(pickup_latitude=40.761710, pickup_longitude=-73.982760)seven_closest = nearest_neighbors(midtown_trip, cleaned_trips, number = 7)
seven_closest
# [{'trip_distance': 0.58,
# 'pickup_latitude': 40.761372,
# 'pickup_longitude': -73.982602,
# 'distance_from_selected': 0.00037310588309379025},
# {'trip_distance': 0.8,
# 'pickup_latitude': 40.762444,
# 'pickup_longitude': -73.98244,
# 'distance_from_selected': 0.00080072217404248},
# {'trip_distance': 1.4,
# 'pickup_latitude': 40.762767,
# 'pickup_longitude': -73.982293,
# 'distance_from_selected': 0.0011555682584735844},
# {'trip_distance': 8.3,
# 'pickup_latitude': 40.762868,
# 'pickup_longitude': -73.983233,
# 'distance_from_selected': 0.0012508768924205918},
# {'trip_distance': 1.26,
# 'pickup_latitude': 40.760057,
# 'pickup_longitude': -73.983502,
# 'distance_from_selected': 0.0018118976240381972},
# {'trip_distance': 0.0,
# 'pickup_latitude': 40.760644,
# 'pickup_longitude': -73.984531,
# 'distance_from_selected': 0.002067074502774709},
# {'trip_distance': 1.72,
# 'pickup_latitude': 40.762107,
# 'pickup_longitude': -73.98479,
# 'distance_from_selected': 0.0020684557041472677}]Looking at the distance_from_selected it appears that our our trips are still fairly close to our selected trip. Notice that most of the data is within a distance of .002 away, so going to the top 7 nearest neighbors didn't seem to give us neighbors too far from each other, which is a good sign.
Still, it's hard to know what distance in latitude and longitude really look like, so let's map the data.
midtown_location = location(midtown_trip) # [40.76171, -73.98276]
midtown_map = map_from(midtown_location, 16)
closest_markers = markers_from_trips(seven_closest)
add_markers(closest_markers, midtown_map)Ok. These locations stay fairly close to our estimated location of 51st street and 7th Avenue. So they could be a good estimate of a trip distance.
mean_distance(seven_closest) # 1.26Ok, now let's try a different location
charging_bull_closest = nearest_neighbors({'pickup_latitude': 40.7049, 'pickup_longitude': -74.0137}, cleaned_trips, number = 12)mean_distance(charging_bull_closest) # 3.515Ok, so there appears to be a significant difference between choosing a location near Times Square versus choosing a location at Wall Street.
In this lab, we used the nearest neighbors function to predict the length of a taxi ride. To do so, we selected a location, then found a number of taxi rides closest to that location, and finally took the average trip lengths of the nearest taxi rides to find an estimate of the new ride's trip length. You can see that even with just a little bit of math and programming we can begin to make meaningful predictions with data.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent