Personal City-Level Location Prediction Based on Behavior via Machine Learning
, Munich
Before I started my Erasmus I thought it could be fun to collect some data about myself while traveling so that I can visualize and analyze them later.
Here I will try to do exactly that without diving into any theoretical explanations of why I did what I did, but I'll add some interpretations of the results.
Goals
This post has two goals:
- Explore and visualize my personal data from my Erasmus time (and a bit after)
- Predict the city I am in based on my behavior (number of steps on a day, money spent, and number of photos/videos taken)
For goal 1, I will use data I collected myself from my phone and smartwatch, as well as data that Google collected about me.
For goal 2, I will use only the data that I collected myself.
I'll consider goal 2 a success if I can feed in the following day that wasn't part of the training data and get a prediction that is correct.
On December 19, 2023 I was in Munich, I took 2575 steps, spent no money, and took no photos/videos.
example = {
'steps': 2575,
'money_spent_offline': 0,
'money_spent_online': 0,
'money_spent_total': 0,
'photos_videos': 0
}
I don't aim to get the best results possible, but rather to get a rough idea of how well this can work and I don't aim to explain any theory.
Exploration and Visualization
Google Data
Let's start by exploring some of the data that Google saved about me. In total, Google saved 2164 place visits.
All map data used for images © OpenStreetMap contributors
These are the top 5 places I visited repeatedly and the number of times I have been there:
El Viajero Sedentario 163
University of Seville 66
Torneo (San Lorenzo) 31
Quiero Dulce 21
REWE 20
... ...
Giselastraße 1
German post office 1
Gazzo 1
Garching-Forschungszentrum 1
La Sagrada Familia 1
Google only saves the names of visited places, but doesn't categorize them. I tried categorizing by place name and what the name contains. This is very inaccurate, as lots of places won't have their category in their name. E.g., cafés often won't have "café" in their name.
What should be accurate is at least my home addresses, university, and airport visits. The airport visits don't all correspond to flights, but also include picking up or dropping off friends.
Here's the result:
Other 1258
Seville Home 483
Munich Home 74
University 73
Market 51
Cafe 42
Airport 40
Berlin Home 38
Train Station 33
Bar 32
Restaurant 23
Beach 9
Museum 8
Google can also save activities it recognizes like flying, walking, bus rides, etc.
This is what the first five entries of the activity data set look like (some coordinates censored):
activityType
startTimestamp
endTimestamp
startLatitude
startLongitude
endLatitude
endLongitude
473
IN_PASSENGER_VEHICLE
2022-09-01 02:20:28.243000+00:00
2022-09-01 03:04:01+00:00
?????
?????
52.365605
13.508949
474
FLYING
2022-09-01 04:23:15.995000+00:00
2022-09-01 06:10:42.876000+00:00
52.365978
13.505437
50.044165
8.562142
475
FLYING
2022-09-01 07:18:53.480000+00:00
2022-09-01 10:23:56.245000+00:00
50.044132
8.562804
37.423092
-5.899401
476
IN_PASSENGER_VEHICLE
2022-09-01 10:57:11.553000+00:00
2022-09-01 11:16:23.291000+00:00
37.423862
-5.900755
37.381626
-6.000961
477
WALKING
2022-09-01 11:37:37.062000+00:00
2022-09-01 12:08:41.935000+00:00
37.382230
-6.001706
?????
?????
So now it is possible to plot the activities on a map.
Flights
Here are all my flight connections:
Train Rides
All train rides:
Flights and Train Rides
Flights and train rides combined:
Bus and Train Rides
A few of the bus rides were mixed up with train rides, so I'll show train rides and bus rides combined:
Walking
For walking, here are heatmaps of where I walked most in Sevilla, first zoomed out and then closer:
For the place visits Google saves, it looks rather similar to the walking heatmap, as I mostly walked everywhere anyway:
My own data
Originally, the data I collected about myself contained rows for about every half hour. It recorded my coordinates, the location name on a city level, the country name, the number of Wi-Fi networks in reach, and a timestamp.
I aggregated this to one row per day for simplicity. Then I needed to assign a single city to each day, which leads to some places where I spent little time to not show up in the data.
Then I added more columns: I scraped my bank account for my payments history, processed my smartwatch data for heart rate, counted my photos and videos, and used a weather service for temperature and condition data.
As my phone broke and I didn't have one for a few days, I also filled in some missing data manually or, as in the case of steps for example, filled in the mean of the variable. This might introduce bias, but as there were only a few days missing, I think it's acceptable.
In the end I also added some augmented data, e.g., the day of the week, season, or a simplified weather condition.
This is, in short, how I arrived at the dataset I will use.
The final dataset contains 416 rows and 30 columns. I think the most interesting columns here are
date, city, max_temp, apparent_max_temp, weather_condition, steps, money_spent_offline, money_spent_online, wifi_count_mean, heart_rate_combined, photos_videos
All temperature values are in °C, money spent in €, and heart rate in bpm.
These are the first 10 rows (days) of the dataset:
weekday
weekend
month
season
city
country
max_temp
apparent_max_temp
temp_difference
humidity
weather_condition
simplified_weather
steps
activity
money_spent_offline
money_spent_online
money_spent_total
wifi_count_mean
watch_1_steps
watch_2_steps
watch_steps_combined
watch_1_mean_heart_rate
watch_2_mean_heart_rate
heart_rate_combined
watch_type
photos_videos
media_activity
generic_lat
generic_lon
date
2022-08-25
Thursday
Weekday
August
Summer
Berlin
Germany
27.3
28.7
-1.4
Moderate
Light drizzle
Rain
10399
Moderate
7.20
0.00
7.20
9.25
NaN
NaN
NaN
NaN
NaN
NaN
NaN
0
NaN
52.52437
13.41053
2022-08-26
Friday
Weekday
August
Summer
Berlin
Germany
26.9
29.2
-2.3
Humid
Heavy rain
Rain
3991
Low
0.00
0.00
0.00
11.57
NaN
NaN
NaN
NaN
NaN
NaN
NaN
1
Moderate
52.52437
13.41053
2022-08-27
Saturday
Weekend
August
Summer
Berlin
Germany
23.0
25.1
-2.1
Humid
Moderate rain
Rain
2423
Low
0.00
0.00
0.00
2.79
NaN
NaN
NaN
NaN
NaN
NaN
NaN
1
Moderate
52.52437
13.41053
2022-08-28
Sunday
Weekend
August
Summer
Berlin
Germany
21.9
20.8
1.1
Moderate
Light drizzle
Rain
2143
Low
0.00
0.00
0.00
9.46
NaN
NaN
NaN
NaN
NaN
NaN
NaN
5
Moderate
52.52437
13.41053
2022-08-29
Monday
Weekday
August
Summer
Berlin
Germany
21.7
20.9
0.8
Moderate
Light drizzle
Rain
8250
Moderate
9.95
0.00
9.95
12.47
NaN
NaN
NaN
NaN
NaN
NaN
NaN
9
Moderate
52.52437
13.41053
2022-08-30
Tuesday
Weekday
August
Summer
Berlin
Germany
20.9
20.4
0.5
Moderate
Light drizzle
Rain
6856
Moderate
15.95
0.00
15.95
7.83
NaN
NaN
NaN
NaN
NaN
NaN
NaN
0
NaN
52.52437
13.41053
2022-08-31
Wednesday
Weekday
August
Summer
Berlin
Germany
21.3
19.7
1.6
Moderate
Light drizzle
Rain
10668
Moderate
0.00
0.00
0.00
21.10
NaN
NaN
NaN
NaN
NaN
NaN
NaN
0
NaN
52.52437
13.41053
2022-09-01
Thursday
Weekday
September
Autumn
Seville
Spain
34.7
33.4
1.3
Moderate
Mostly cloudy
Cloudy
28784
High
76.96
0.00
76.96
21.83
NaN
NaN
NaN
NaN
NaN
NaN
NaN
30
High
37.38283
-5.97317
2022-09-02
Friday
Weekday
September
Autumn
Seville
Spain
33.6
31.9
1.7
Moderate
Partly cloudy
Cloudy
13744
High
34.28
21.00
55.28
15.05
NaN
NaN
NaN
NaN
NaN
NaN
NaN
14
Moderate
37.38283
-5.97317
2022-09-03
Saturday
Weekend
September
Autumn
Seville
Spain
32.1
31.0
1.1
Moderate
Mainly clear
Clear
35051
High
42.00
71.93
113.93
12.60
NaN
NaN
NaN
NaN
NaN
NaN
NaN
58
High
37.38283
-5.97317
I visited at least the following 28 cities. The count is the number of days I spent there.
There are cities missing, like Ronda for example, because I didn't spend a lot of time there. I didn't bother to fix this.
Seville 223
Munich 65
Berlin 48
Cádiz 21
Istanbul 7
Milano 6
Lisbon 4
Madrid 4
Valencia 4
Barcelona 4
Las Palmas de Gran Canaria 3
Vinaròs 3
Merzouga 3
Benidorm 3
Cartagena 2
Córdoba 2
Alhama de Aragón 2
Fès 2
Dénia 1
Zaragoza 1
Castelló de la Plana 1
Lagoa 1
Granada 1
Málaga 1
Zuheros 1
Barbate 1
Jerez de la Frontera 1
Santiponce 1
And these are the 6 countries I visited (on 3 continents, two of which I visited for the first time):
Spain 280
Germany 113
Turkey 7
Italy 6
Portugal 5
Morocco 5
The following are some summary statistics like top value per variable, mean, or min and max. I censored the money_spent variables as I will also do later on.
weekday
weekend
month
season
city
country
max_temp
apparent_max_temp
temp_difference
humidity
weather_condition
simplified_weather
steps
activity
money_spent_offline
money_spent_online
money_spent_total
wifi_count_mean
watch_1_steps
watch_2_steps
watch_steps_combined
watch_1_mean_heart_rate
watch_2_mean_heart_rate
heart_rate_combined
watch_type
photos_videos
media_activity
generic_lat
generic_lon
count
416
416
416
416
416
416
416.000000
416.000000
416.000000
416
416
416
416.000000
416
416.000000
416.000000
416.000000
416.000000
199.000000
168.000000
359.000000
199.000000
166.000000
358.000000
358.000000
416.000000
342
416.000000
416.00000
unique
7
2
12
4
28
6
NaN
NaN
NaN
3
12
4
NaN
3
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
2
NaN
NaN
top
Thursday
Weekday
September
Autumn
Seville
Spain
NaN
NaN
NaN
Moderate
Mainly clear
Clear
NaN
Moderate
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
Moderate
NaN
NaN
freq
60
297
60
135
223
280
NaN
NaN
NaN
273
87
149
NaN
188
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
NaN
269
NaN
NaN
mean
NaN
NaN
NaN
NaN
NaN
NaN
23.865144
22.922356
0.942788
NaN
NaN
NaN
11577.788462
NaN
?????
?????
?????
12.007359
11392.187500
10999.627329
11247.524547
71.333333
71.798742
71.540037
0.444134
17.302885
NaN
41.001306
0.05572
std
NaN
NaN
NaN
NaN
NaN
NaN
7.779006
9.021076
1.961471
NaN
NaN
NaN
7793.126558
NaN
?????
?????
?????
4.198189
7440.075266
7734.197996
7629.722817
4.250173
4.732582
4.519293
0.497565
37.645968
NaN
5.841556
9.06130
min
NaN
NaN
NaN
NaN
NaN
NaN
-1.400000
-5.700000
-6.000000
NaN
NaN
NaN
107.000000
NaN
0.000000
0.000000
0.000000
0.000000
384.000000
74.000000
74.000000
62.000000
63.000000
62.000000
0.000000
0.000000
NaN
28.099730
-15.41343
25%
NaN
NaN
NaN
NaN
NaN
NaN
19.100000
17.650000
-0.100000
NaN
NaN
NaN
5637.750000
NaN
?????
?????
?????
8.937500
6163.000000
5181.500000
5597.500000
68.000000
68.000000
68.000000
0.000000
1.000000
NaN
37.382830
-5.97317
50%
NaN
NaN
NaN
NaN
NaN
NaN
24.400000
24.200000
0.900000
NaN
NaN
NaN
9872.000000
NaN
?????
?????
?????
11.265000
9943.000000
9039.000000
9440.000000
71.000000
71.798742
71.000000
0.000000
4.000000
NaN
37.382830
-5.97317
75%
NaN
NaN
NaN
NaN
NaN
NaN
29.600000
29.400000
2.100000
NaN
NaN
NaN
16053.500000
NaN
?????
?????
?????
14.622500
15540.500000
14017.000000
15096.500000
73.500000
74.750000
74.000000
1.000000
14.000000
NaN
48.137430
11.57549
max
NaN
NaN
NaN
NaN
NaN
NaN
43.300000
43.300000
6.300000
NaN
NaN
NaN
39452.000000
NaN
?????
?????
?????
34.750000
37912.000000
35623.000000
37912.000000
83.000000
88.000000
88.000000
1.000000
300.000000
NaN
52.524370
28.94966
Visualizations
Steps
Starting with the steps variable, here is a histogram of it:
and the daily step count over time:
or the same information as a heatmap:
The day I took the most steps was the day of the Carnival of Cádiz. The night before we also were in a club so the night already contributed a lot:
weekday
weekend
month
season
city
country
max_temp
apparent_max_temp
temp_difference
humidity
weather_condition
simplified_weather
steps
activity
money_spent_offline
money_spent_online
money_spent_total
wifi_count_mean
watch_1_steps
watch_2_steps
watch_steps_combined
watch_1_mean_heart_rate
watch_2_mean_heart_rate
heart_rate_combined
watch_type
photos_videos
media_activity
generic_lat
generic_lon
date
2023-02-18
Saturday
Weekend
February
Winter
Cádiz
Spain
19.1
15.1
4.0
Dry
Light drizzle
Rain
39452
High
16.14
0.0
16.14
13.15
37912.0
NaN
37912.0
78.0
NaN
78.0
0.0
24
Moderate
36.52672
-6.2891
Next we can also observe if the step count differs between different days of the week. To me, it looks like there is a significant difference between Saturday and the rest of the week.
There would be a statistical test necessary to find out how likely this difference would be to appear due to randomness, but I will leave this out because I spent way too much time on this whole thing already.
This is the comparison between the weekend and not weekend:
And the following is the comparison between different weather conditions. The following two graphs show the same information, but the second one contains simplified weather conditions which summarizes different strengths of the same weather condition (e.g. light rain and heavy rain are both rain).
As I would expect, the step count is higher when the weather is better.
We can also compare the step count between different cities. The following graph shows the mean step count per city.
The most mean steps are in cities where I only stayed for a short time because I walked around a lot to see as much as possible in the short time I had.
Temperature
Changing to another variable, here is the daily maximum and apparent maximum temperature per day:
The hottest day I had was in Seville when the temperature rose to 43.3°C. I'm sure there were hotter days even, but I also escaped to Cádiz sometimes.
weekday
weekend
month
season
city
country
max_temp
apparent_max_temp
temp_difference
humidity
weather_condition
simplified_weather
steps
activity
money_spent_offline
money_spent_online
money_spent_total
wifi_count_mean
watch_1_steps
watch_2_steps
watch_steps_combined
watch_1_mean_heart_rate
watch_2_mean_heart_rate
heart_rate_combined
watch_type
photos_videos
media_activity
generic_lat
generic_lon
date
2023-06-26
Monday
Weekday
June
Summer
Seville
Spain
43.3
43.3
0.0
Moderate
Clear
Clear
7300
Moderate
187.41
164.62
352.03
15.72
NaN
8419.0
8419.0
NaN
67.0
67.0
1.0
4
Moderate
37.38283
-5.97317
I never experienced temperatures this high before. But, I was also surprised how it was less of a problem than expected. I think the reason for this is that the humidity was not that high.
This means the apparent temperature was similar to the actual temperature. We can plot the difference between the apparent and actual temperature. In this next graph, if the difference is positive, the apparent temperature is lower than the actual temperature.
From this I would conclude that the humidity is smaller. If the difference is negative I assume that the humidity is higher.
I'm not sure if this is completely correct under all circumstances (probably not, e.g., if there's a lot of wind), but I use it as an approximation for humidity.
With this logic, the driest day was on January 13, 2023 in Munich with a difference of 6.3°C:
weekday
weekend
month
season
city
country
max_temp
apparent_max_temp
temp_difference
humidity
weather_condition
simplified_weather
steps
activity
money_spent_offline
money_spent_online
money_spent_total
wifi_count_mean
watch_1_steps
watch_2_steps
watch_steps_combined
watch_1_mean_heart_rate
watch_2_mean_heart_rate
heart_rate_combined
watch_type
photos_videos
media_activity
generic_lat
generic_lon
date
2023-01-13
Friday
Weekday
January
Winter
Munich
Germany
10.8
4.5
6.3
Dry
Moderate drizzle
Rain
13961
High
44.61
98.07
142.68
15.38
14159.0
NaN
14159.0
70.0
NaN
70.0
0.0
0
NaN
48.13743
11.57549
And the most humid day was on July 12, 2023 in Cartagena with a difference of -6.0°C. It was so humid there that you could feel the humidity in the air and on your skin. Sweating didn't have a cooling effect anymore. It also felt hotter than, for example, in Seville, although the actual temperature in Cartagena was way lower.
weekday
weekend
month
season
city
country
max_temp
apparent_max_temp
temp_difference
humidity
weather_condition
simplified_weather
steps
activity
money_spent_offline
money_spent_online
money_spent_total
wifi_count_mean
watch_1_steps
watch_2_steps
watch_steps_combined
watch_1_mean_heart_rate
watch_2_mean_heart_rate
heart_rate_combined
watch_type
photos_videos
media_activity
generic_lat
generic_lon
date
2023-07-12
Wednesday
Weekday
July
Summer
Cartagena
Spain
34.0
40.0
-6.0
Humid
Partly cloudy
Cloudy
16871
High
68.48
8.0
76.48
10.63
NaN
17496.0
17496.0
NaN
70.0
70.0
1.0
16
Moderate
37.60512
-0.98623
Weather
Next we can look at the (simplified) weather conditions I observed. The most common weather condition was clear, closely followed by rain. The few snowy days were probably mostly from the time I spent in Munich:
Looking only at the weather while I was in Seville, it was mostly clear, but there were also lots of rainy days. The longest period of consecutive days with rain was 7 days, starting on 2022-12-11.
In Munich, it was mostly rainy/snowy, because I was there mostly in winter:
From the days I spent in Cádiz there were also quite a few rainy days:
Photos/Videos
Next we can look at the number of photos/videos taken per day. The following graph shows the number of photos/videos taken per day.
The day I took the most photos/videos was November 20, 2022 in Córdoba. I took 300 photos/videos that day.
weekday
weekend
month
season
city
country
max_temp
apparent_max_temp
temp_difference
humidity
weather_condition
simplified_weather
steps
activity
money_spent_offline
money_spent_online
money_spent_total
wifi_count_mean
watch_1_steps
watch_2_steps
watch_steps_combined
watch_1_mean_heart_rate
watch_2_mean_heart_rate
heart_rate_combined
watch_type
photos_videos
media_activity
generic_lat
generic_lon
date
2022-11-20
Sunday
Weekend
November
Autumn
Córdoba
Spain
16.4
15.5
0.9
Moderate
Mostly cloudy
Cloudy
23850
High
131.1
4.99
136.09
11.8
26638.0
NaN
26638.0
80.0
NaN
80.0
0.0
300
High
37.89155
-4.77275
I took at least one photo/video on 342 out of 416 days which is around 82.21% of all days.
In total, I took 7198 photos/videos which is an average of 17.3 photos/videos per day, however the median is 4.
There is also an observable difference between different weather conditions again. The following graph shows the mean number of photos/videos taken per day for different (simplified) weather conditions.
I again won't do a statistical test.
Wi-Fi
Next up is the mean Wi-Fi count per day which is plotted in the following graph. The mean Wi-Fi count is the mean count of all the Wi-Fi networks my phone detected in the same day.
I imagined that this could be an indicator of how civilized my location is.
The day with the most Wi-Fi networks detected was April 7, 2023 in Barcelona with a mean of 34.75 Wi-Fi networks detected.
weekday
weekend
month
season
city
country
max_temp
apparent_max_temp
temp_difference
humidity
weather_condition
simplified_weather
steps
activity
money_spent_offline
money_spent_online
money_spent_total
wifi_count_mean
watch_1_steps
watch_2_steps
watch_steps_combined
watch_1_mean_heart_rate
watch_2_mean_heart_rate
heart_rate_combined
watch_type
photos_videos
media_activity
generic_lat
generic_lon
date
2023-04-07
Friday
Weekday
April
Spring
Barcelona
Spain
18.9
17.0
1.9
Moderate
Mainly clear
Clear
26486
High
34.06
0.0
34.06
34.75
26786.0
NaN
26786.0
72.0
NaN
72.0
0.0
74
High
41.38879
2.15899
Heart Rate
In October 2022 I also started measuring my heart rate. In May 2023 I switched to a different device.
From the plot there is already a day that especially sticks out: June 3, 2023. On this day I had a mean heart rate of 88 bpm. It was the day I traveled from Seville to Munich to surprise my girlfriend on her birthday.
One could say that my heart rate could have been higher because I traveled on that day and because I was very active, but I also had lots of other days with these circumstances where my heart rate didn't stick out like this.
I think it's super cute to see that my heart rate was noticeably higher because I was so excited.
weekday
weekend
month
season
city
country
max_temp
apparent_max_temp
temp_difference
humidity
weather_condition
simplified_weather
steps
activity
money_spent_offline
money_spent_online
money_spent_total
wifi_count_mean
watch_1_steps
watch_2_steps
watch_steps_combined
watch_1_mean_heart_rate
watch_2_mean_heart_rate
heart_rate_combined
watch_type
photos_videos
media_activity
generic_lat
generic_lon
date
2023-06-03
Saturday
Weekend
June
Summer
Munich
Germany
21.0
19.7
1.3
Moderate
Mainly clear
Clear
16038
High
93.50
0.0
93.50
19.25
NaN
13295.0
13295.0
NaN
88.0
88.0
1.0
4
Moderate
48.13743
11.57549
The most relaxing day was on February 25, 2023 in Seville with a mean heart rate of 62 bpm.
I also happen to know what I did that day: I went to my favorite café, El Viajero Sedentario, for two and a half hours in the morning and afterward I relaxed on the rooftop enjoying the weather that was slowly getting warmer. Although there were a few rain drops on that day, there was also a long time of sun I could enjoy.
weekday
weekend
month
season
city
country
max_temp
apparent_max_temp
temp_difference
humidity
weather_condition
simplified_weather
steps
activity
money_spent_offline
money_spent_online
money_spent_total
wifi_count_mean
watch_1_steps
watch_2_steps
watch_steps_combined
watch_1_mean_heart_rate
watch_2_mean_heart_rate
heart_rate_combined
watch_type
photos_videos
media_activity
generic_lat
generic_lon
date
2023-02-25
Saturday
Weekend
February
Winter
Seville
Spain
15.2
11.6
3.6
Dry
Light drizzle
Rain
1975
Low
9.9
0.0
9.9
7.77
1887.0
NaN
1887.0
62.0
NaN
62.0
0.0
1
Moderate
37.38283
-5.97317
Money Spent
I tracked spending money "offline" and "online". Offline here means that I paid in person with my card and online means that I paid online with my card.
I tried only paying with card so most of my payments should be covered.
Comparing total money spent between different weather conditions we can see that the mean total money spent per day is the highest on clear days and on Sundays.
Y-scales are censored on purpose.
Y-scales are censored on purpose.
Y-scales are censored on purpose.
Offline Money Spent
For the offline spending the largest payment (excluding ATM withdrawals) was a flamenco show group evening at a bar that I prepaid for my friends.
weekday
weekend
month
season
city
country
max_temp
apparent_max_temp
temp_difference
humidity
weather_condition
simplified_weather
steps
activity
money_spent_offline
money_spent_online
money_spent_total
wifi_count_mean
watch_1_steps
watch_2_steps
watch_steps_combined
watch_1_mean_heart_rate
watch_2_mean_heart_rate
heart_rate_combined
watch_type
photos_videos
media_activity
generic_lat
generic_lon
date
2023-05-26
Friday
Weekday
May
Spring
Seville
Spain
28.1
27.4
0.7
Moderate
Moderate rain
Rain
10283
Moderate
?????
13.9
?????
12.76
NaN
10209.0
10209.0
NaN
73.0
73.0
1.0
23
Moderate
37.38283
-5.97317
Online Money Spent
The largest online payment was prepaying for our vacation rental in Cádiz.
weekday
weekend
month
season
city
country
max_temp
apparent_max_temp
temp_difference
humidity
weather_condition
simplified_weather
steps
activity
money_spent_offline
money_spent_online
money_spent_total
wifi_count_mean
watch_1_steps
watch_2_steps
watch_steps_combined
watch_1_mean_heart_rate
watch_2_mean_heart_rate
heart_rate_combined
watch_type
photos_videos
media_activity
generic_lat
generic_lon
date
2023-05-07
Sunday
Weekend
May
Spring
Seville
Spain
32.1
31.0
1.1
Moderate
Clear
Clear
24232
High
28.4
?????
?????
11.63
NaN
21945.0
21945.0
NaN
85.0
85.0
1.0
150
High
37.38283
-5.97317
At some point I also had to buy a new phone because I destroyed my old one while jumping around to music with my housemates. This was the second most expensive day.
weekday
weekend
month
season
city
country
max_temp
apparent_max_temp
temp_difference
humidity
weather_condition
simplified_weather
steps
activity
money_spent_offline
money_spent_online
money_spent_total
wifi_count_mean
watch_1_steps
watch_2_steps
watch_steps_combined
watch_1_mean_heart_rate
watch_2_mean_heart_rate
heart_rate_combined
watch_type
photos_videos
media_activity
generic_lat
generic_lon
date
2022-10-15
Saturday
Weekend
October
Autumn
Seville
Spain
28.9
28.2
0.7
Moderate
Clear
Clear
6076
Moderate
8.28
?????
?????
11.42
NaN
NaN
NaN
NaN
NaN
NaN
NaN
0
NaN
37.38283
-5.97317
And the day of the Barcelona vacation rental and flight payments comes after that.
weekday
weekend
month
season
city
country
max_temp
apparent_max_temp
temp_difference
humidity
weather_condition
simplified_weather
steps
activity
money_spent_offline
money_spent_online
money_spent_total
wifi_count_mean
watch_1_steps
watch_2_steps
watch_steps_combined
watch_1_mean_heart_rate
watch_2_mean_heart_rate
heart_rate_combined
watch_type
photos_videos
media_activity
generic_lat
generic_lon
date
2023-03-05
Sunday
Weekend
March
Spring
Seville
Spain
16.2
14.6
1.6
Moderate
Light rain
Rain
1972
Low
17.28
?????
?????
10.22
2076.0
NaN
2076.0
65.0
NaN
65.0
0.0
2
Moderate
37.38283
-5.97317
Correlations
Let's look at correlations between the different variables. The following correlations are the most interesting ones for me:
steps, heart_rate_combined: 0.62***
steps, photos_videos: 0.46***
photos_videos, heart_rate_combined: 0.37***
temp_difference, apparent_max_temp: -0.70***
Interpretation: Usually, on days I took more steps I also had a higher heart rate. This is very expected because I was more active on these days.
Also, on days I took more steps I also took more photos and videos. I like taking photos so this is also not surprising.
On days I took more photos I also usually had a higher heart rate, but this correlation is rather weak.
Finally, on days with a higher difference between the max and apparent max temperature there was usually a lower apparent max temperature. I think a reason for this can be that I was mostly in dry-ish climates.
Unlike in the previous data post, there are no significant correlations with money_spent anymore.
That may be a less interesting result now, but it's good because it increases the potential for more accurately predicting my location later.
Here are two example plots showing the weak correlation with money spent (money scales are censored on purpose):
One could also expect a correlation between the max temperature and the number of steps. However, this correlation is very weak (0.25):
Linear Regression
My goal with linear regression here is not to make accurate predictions, but to find interesting relationships between the variables.
Of course, these relationships are not causal.
Also, I didn't test the assumptions of linear regression here, so the results are not very reliable.
Steps as independent variable
First I run a linear regression with the number of steps as the dependent variable. Through backwards elimination of independent variables that are insignificant at the 1% level I end up with the following model.
apparent_max_temp, wifi_count_mean, and photos_videos are the only significant independent variables.
An interpretation of the result would be that if the apparent max temperature increased by 1°C, I took on average 209.36 more steps, holding the other variables constant.
If the mean Wi-Fi count increased by 1, I took on average 362.81 more steps, holding the other variables constant.
If the number of photos and videos I took increased by 1, I took on average 89.26 more steps, holding the other variables constant.
OLS Regression Results
==============================================================================
Dep. Variable: steps R-squared: 0.299
Model: OLS Adj. R-squared: 0.294
Method: Least Squares F-statistic: 58.49
Date: Sun, 24 Dec 2023 Prob (F-statistic): 1.61e-31
Time: 13:37:10 Log-Likelihood: -4243.7
No. Observations: 416 AIC: 8495.
Df Residuals: 412 BIC: 8512.
Df Model: 3
Covariance Type: nonrobust
=====================================================================================
coef std err t P>|t| [0.025 0.975]
-------------------------------------------------------------------------------------
const 877.9449 1305.529 0.672 0.502 -1688.385 3444.274
apparent_max_temp 209.3622 35.732 5.859 0.000 139.123 279.601
wifi_count_mean 362.8066 76.920 4.717 0.000 211.603 514.011
photos_videos 89.2580 8.573 10.412 0.000 72.406 106.110
==============================================================================
Omnibus: 40.938 Durbin-Watson: 1.468
Prob(Omnibus): 0.000 Jarque-Bera (JB): 54.677
Skew: 0.725 Prob(JB): 1.34e-12
Kurtosis: 4.027 Cond. No. 177.
==============================================================================
Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
Photos/Videos ~ steps
Next, I run a linear regression with the number of photos and videos as the dependent variable and just the steps as the independent variable.
As noted in note [2], there seem to be some problems with the model. Like I said before, I didn't test the assumptions of linear regression here, so the results are not very reliable.
Nevertheless, the model estimates that if my number of steps increased by 1, the number of photos and videos I took increased by 0.0022 on average.
The negative intercept is significant, but has no practical meaning in this case.
OLS Regression Results
==============================================================================
Dep. Variable: photos_videos R-squared: 0.208
Model: OLS Adj. R-squared: 0.206
Method: Least Squares F-statistic: 108.4
Date: Sun, 24 Dec 2023 Prob (F-statistic): 1.05e-22
Time: 13:37:11 Log-Likelihood: -2050.7
No. Observations: 416 AIC: 4105.
Df Residuals: 414 BIC: 4114.
Df Model: 1
Covariance Type: nonrobust
==============================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
const -8.1747 2.949 -2.772 0.006 -13.971 -2.379
steps 0.0022 0.000 10.412 0.000 0.002 0.003
==============================================================================
Omnibus: 428.657 Durbin-Watson: 1.399
Prob(Omnibus): 0.000 Jarque-Bera (JB): 15285.273
Skew: 4.624 Prob(JB): 0.00
Kurtosis: 31.219 Cond. No. 2.50e+04
==============================================================================
Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
[2] The condition number is large, 2.5e+04. This might indicate that there are
strong multicollinearity or other numerical problems.
Heart Rate ~ steps
Let's look at the relationship between the heart rate and just the number of steps.
The model estimates that if my number of steps increased by 1, my mean heart rate for that day increased by 0.0004 bpm on average.
OLS Regression Results
===============================================================================
Dep. Variable: heart_rate_combined R-squared: 0.387
Model: OLS Adj. R-squared: 0.385
Method: Least Squares F-statistic: 224.3
Date: Sun, 24 Dec 2023 Prob (F-statistic): 1.14e-39
Time: 13:38:03 Log-Likelihood: -960.01
No. Observations: 358 AIC: 1924.
Df Residuals: 356 BIC: 1932.
Df Model: 1
Covariance Type: nonrobust
==============================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
const 67.3933 0.334 201.593 0.000 66.736 68.051
steps 0.0004 2.41e-05 14.977 0.000 0.000 0.000
==============================================================================
Omnibus: 49.215 Durbin-Watson: 1.443
Prob(Omnibus): 0.000 Jarque-Bera (JB): 81.853
Skew: 0.822 Prob(JB): 1.68e-18
Kurtosis: 4.669 Cond. No. 2.48e+04
==============================================================================
Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
[2] The condition number is large, 2.48e+04. This might indicate that there are
strong multicollinearity or other numerical problems.
heart rate ~ watch_type + steps
As I switched to a different watch in May 2023, I also wanted to see if the devices differ in their heart rate measurements.
I wasn't sure if I should include this section, because the included independent variables are very limited and I only tried linear relationships. Resulting R-squared values are rather low. Although these limitations might lead to very unreliable results, I'll still include this section to show my thinking.
Another way to ask if the devices differ in their measurement is if the watch type influences the measured mean heart rate.
For this I run a linear regression with the mean heart rate as the dependent variable and the watch type and the number of steps as the independent variables because I want to account for the influence of the number of steps.
The estimated model's watch_type coefficient is not significant, so the watch type doesn't seem to influence the measured mean heart rate (which would be good because if both are accurate there shouldn't be a difference).
OLS Regression Results
===============================================================================
Dep. Variable: heart_rate_combined R-squared: 0.388
Model: OLS Adj. R-squared: 0.384
Method: Least Squares F-statistic: 112.3
Date: Sun, 24 Dec 2023 Prob (F-statistic): 1.61e-38
Time: 13:38:25 Log-Likelihood: -959.71
No. Observations: 358 AIC: 1925.
Df Residuals: 355 BIC: 1937.
Df Model: 2
Covariance Type: nonrobust
==============================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
const 67.2716 0.370 181.645 0.000 66.543 68.000
watch_type 0.2888 0.377 0.765 0.445 -0.453 1.031
steps 0.0004 2.41e-05 14.937 0.000 0.000 0.000
==============================================================================
Omnibus: 47.203 Durbin-Watson: 1.441
Prob(Omnibus): 0.000 Jarque-Bera (JB): 77.079
Skew: 0.800 Prob(JB): 1.83e-17
Kurtosis: 4.614 Cond. No. 3.31e+04
==============================================================================
Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
[2] The condition number is large, 3.31e+04. This might indicate that there are
strong multicollinearity or other numerical problems.
heart rate ~ watch_type + photos_videos + apparent_max_temp
I also tried to use the number of photos/videos I took and the apparent max temperature as a proxy for how active I was on a day.
Here the watch type coefficient isn't significant either.
OLS Regression Results
===============================================================================
Dep. Variable: heart_rate_combined R-squared: 0.161
Model: OLS Adj. R-squared: 0.154
Method: Least Squares F-statistic: 22.59
Date: Sun, 24 Dec 2023 Prob (F-statistic): 2.10e-13
Time: 13:38:28 Log-Likelihood: -1016.1
No. Observations: 358 AIC: 2040.
Df Residuals: 354 BIC: 2056.
Df Model: 3
Covariance Type: nonrobust
=====================================================================================
coef std err t P>|t| [0.025 0.975]
-------------------------------------------------------------------------------------
const 69.1102 0.579 119.364 0.000 67.972 70.249
watch_type -0.0576 0.515 -0.112 0.911 -1.071 0.956
photos_videos 0.0408 0.006 7.338 0.000 0.030 0.052
apparent_max_temp 0.0784 0.028 2.803 0.005 0.023 0.133
==============================================================================
Omnibus: 33.655 Durbin-Watson: 1.559
Prob(Omnibus): 0.000 Jarque-Bera (JB): 43.765
Skew: 0.699 Prob(JB): 3.14e-10
Kurtosis: 3.990 Cond. No. 122.
==============================================================================
Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
Goal 2: Prediction
Now let's get to the second part: predicting my location on a city level based on my behavior.
I filtered out all days I stayed in a city that I stayed in for less than 7 days. This leaves 5 possible cities to predict: Seville, Munich, Berlin, Cádiz, and Istanbul.
Then I applied encoding and standardization where necessary/applicable.
My goal here isn't to produce the absolute best model, but rather to try out different approaches and see what might work better than other approaches here.
Neural Network
The first approach I tried was to use a neural network. For this I also used SMOTE to balance the classes.
NN: First Try
Just for the first neural network try I used all variables except variables coming from the watch, or variables directly hinting at my location (coordinates and country).
Without much trying around I ended up with the following architecture:
Three dense layers, two of which with a ReLU activation, and the last one with a softmax activation.
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense (Dense) (None, 64) 3776
dense_1 (Dense) (None, 32) 2080
dense_2 (Dense) (None, 5) 165
=================================================================
Total params: 6,021
I used the Adam optimizer, a categorical cross entropy loss, and accuracy metrics. After training for 10 epochs I got the following results:
Orange: train accuracy
Blue: validation accuracy
Orange: train loss
Blue: validation loss
NN Second Try: Only Behavioral Variables
For the second try with a neural network I only left the behavioral variables (steps, money spent, photos/videos).
This was my goal from the beginning.
I also tried out a different architecture with one ReLU layer, one softmax layer, and L2 regularization.
Model: "sequential_3"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_7 (Dense) (None, 64) 704
dense_8 (Dense) (None, 5) 325
=================================================================
Total params: 1,029
Trainable params: 1,029
Non-trainable params: 0
_________________________________________________________________
Orange: train accuracy
Blue: validation accuracy
Orange: train loss
Blue: validation loss
Here we end up with a pretty bad validation accuracy. It also looks a lot like the model is overfitting.
I didn't try to improve this a lot because of time reasons and because I think the most promising way to fix this would be to have more data which I can't do anyway unless I wait for multiple years.
New example not included in dataset
Nevertheless, the model can already correctly predict the city of my test day:
December 19, 2023
example = {
'steps': 2575,
'money_spent_offline': 0,
'money_spent_online': 0,
'money_spent_total': 0,
'photos_videos': 0
}
Actual: Munich
Predicted: Munich
However, with a validation accuracy this low, this might as well be just a random guess.
KNN
The next method I tried was KNN, where the highest accuracy I got was 0.42 with K=3.
This is already better than the NN, but I suspected the next methods would lead to better results.
Random Forest
For random forest I did a random search for the best parameters.
Here I ended up with an accuracy of 0.65.
XGBoost
And for XGBoost I got an accuracy of 0.69 (nice).
With a random search I even found parameters that lead to an accuracy of 0.7. This appears to be the best model I could find without investing more time into this.
And I think it's already pretty good. It can also correctly predict the city of my test day:
New example not included in dataset
December 19, 2023
example = {
'steps': 2575,
'money_spent_offline': 0,
'money_spent_online': 0,
'money_spent_total': 0,
'photos_videos': 0
}
Actual: Munich
Predicted: Munich
But of course, it's not always correct:
December 05, 2023
example = {
'steps': 3579,
'money_spent_offline': 14.1,
'money_spent_online': 0,
'money_spent_total': 14.1,
'photos_videos': 0
}
Actual: Munich
Predicted: Berlin
For this day it got the city wrong, but at least it was still in the same country.
Counting how often a feature is used to split the data across all trees, we can see that the number of steps is the most important feature, followed by money spent offline, the number of photos/videos taken, and money spent online. 'activity' and 'media_activity' are just augmentations I derived from the number of steps and photos/videos.
Conclusion
The XGBoost model turned out to work best for this work. As I didn't try improving much on the neural network, further exploring might lead to better results. Also, the accuracy metric I used could be complemented by other metrics.
I could have incorporated more behavioral features like which transportation modes I used on the day, but I didn't want to spend too much time on this project.
For now, I'm very content with the result already.
It shows that my behavior changes enough between different cities to be able to predict the city I'm in. It is not applicable to anyone besides myself, but I suspect that lots of people would have similar results if they were to try this.
Although doing this was fun, I'm not sure if there are any practical applications for the prediction of my location, as every device has GPS built in and knows its location.
As GPS usage is rather battery-consuming, there might be use cases for predicting the location without GPS usage. As this is done on a city-level, it might be useful for, e.g., weather services.
At least this post might inspire similar fun personal projects for others which can help to learn about data collection, exploration, and prediction.
I wouldn't continue trying to improve this model, but I can imagine to keep collecting my data for further explorations and visualizations over the years.
I hope you enjoyed this little project.
What is left to write for me is a general summary post of the whole time.