From Tula to Berlin: analysis of cities on Instagram

It's no secret that you can find photos of any large city on Instagram. What if we try to restore the whole picture using fragments? The information obtained will help to get an idea of ​​unfamiliar places and will be useful for travelers, complementing traditional travel guides.

The idea of ​​analyzing cities by photo is generally not new [ 1 , 2 , 3 ], but to be honest, the articles found in essence do not tell much.

How to collect data from instagram is a topic that has been highlighted several times and is not covered in this article. Something can be pulled through the API, but if there is no access to it, there are alternatives .

Our basic tools are Python and Plotly. At the end, there are links to GitHub and Jupyter laptops for different cities (the graphics are interactive, therefore they contain more information and are recommended for in-person acquaintance). The repository also includes scripts for data collection.
')
In this article we will walk around Berlin. Dataset, considered in this article, contains about 100k photos for ~ 2k locations.

For each location we need:

1. Title
2. Coordinates
3. Photos (10-100 pcs.)
4. The number of posts (variable edge_location_to_media on the location page, it does not appear in the documentation, but by indirect indications means exactly that)

First walk

We put the data on the map. To select the most lively places, combine locations located on the same street in one marker. We will draw maps using Mapbox.


Map of Berlin. Markers mark the number of locations.

Let's get acquainted with the main toponyms. To do this, we need to convert the coordinates of places to addresses - this is the reverse geocoding task. For its decision Google Geocoding API was used. After collecting geodata, we will sort the streets and areas by the number of locations.


For cities such as Moscow, information about the districts is not too important, everything is approximately the same in the center, but Berlin is more heterogeneous and therefore it is useful to distinguish, for example, Kreuzberg from Prenzlauer Berg .

Look at the list of places sorted by popularity.


It mixes "formal" places (monuments, museums, galleries) with "informal" (clubs, bars, shops). To separate one from another, we need data from Wikipedia; Unlike Instagram, its API is available to everyone in full. On one axis, we postpone the number of posts on instagram, on the other - the number of views of the Wikipedia article about this place. In this picture, more “formal” places will be located higher, more popular - to the right.

To reduce errors, we will group locations along the streets, as on the map. In the process of selecting articles for locations, part of the data is lost, so the figure will contain fewer points than the map.
“Insta-wiki” diagram. For several streets marked the most significant places. For more information, see the laptop.

Where to go to take a selfie? Estimate the proportion of photos containing faces. This will help us OpenCV and the cascade of Haar.
Share photos containing faces. Points located on the right side of the picture are popular places for selfies (or vanity fairs)

Going deeper

Next, apply a neural network to determine the environment in the photographs. Used CNN Places365 , trained on dataset collected at MIT [4] . The most appropriate tags for this task were selected. Find out which of them are more common:
Tag Rating. Names are left original. They should not be taken literally: martial_arts_gym is rather a gym, and with a discotheque tag just a dark room can be noted

Let's see which tags correspond to the streets:

The same on the map:


Map of Berlin with the most characteristic tags. Pay attention to the discotheque tag on the right - it is Friedrichshain, an area with a vibrant nightlife.

Hello, Hallo, Hola

One of the ways to learn something about a new city is to compare it with the one that you know. Take feature vectors for locations of two cities and using t-SNE we get two-dimensional coordinates. For greater clarity, the figure hidden locations, lying in the area of ​​the city of the opponent.

Comparison of locations in Berlin and Moscow. Labels indicate the dominant feature in a given area. Clusters of different colors, located nearby, indicate the points of contact between cities, i.e., similar places

Let's look at the difference of signs:
The difference of signs between Berlin and Moscow. It seems that in our capital more often they are photographed in gyms and fitting rooms.

Articles
1. How to Study the City on Instagram
2. What We Instagram: A First Analysis Of Instagram Photo Content and User Types
3. Zooming into an Instagram City: Reading the local through social media
4. Places: A 10 million Image Database for Scene Recognition

Laptops
Tula , Moscow , St. Petersburg , Berlin , Rome , Hong Kong

Github
github.com/pskryuchkov/voyage