This project aims to understand the connections between painters and art styles. Originally, this included creating a painter dataset, but this evolved into a seperate project called PainterPalette. These notebooks use the PainterPalette dataset, creating links between painters based on similarity of style, geographical and social interaction, and also connections between styles and movements.

There have been multiple networks created by us of painters (based on being at the same places at the same time+nationality, additionally style similarity, or who influenced whom networks). We also created a network of styles and movements.

REDO: From 2024.03.02 the analyses were repeated, because PainterPalette now has more painters, data and cleaner attributes, taking more painters from Art500k. Now organized a bit better and simplified.
More fresh analysis coming soon (summer-autumn 2024).

Some old analysis (2023)/visualizations of the resulting networks:

• Movement network:
  

• InfluencedOn network:
  

• A map of most common styles in the dataset in different countries:
  

• Time-and-place network:
  

• Time-and-place network without reordering nodes based on time (older version):
  

(One can see the clusters of movements despite not connecting painters based on style similarities.)

We created our own dataset called Painter Palette: a dataset with ~3200 painters (Art500k data on 10000 painters), data on their styles, movements, nationality, birthyear, first and last year of painting in the Art500k dataset, birthplace, places of their paintings, influences, friends and coworkers, teachers. It's created by assembling data from paintings from the Art500k dataset, and data from paintings from the WikiArt dataset, with some manual additions. This dataset is available in the PainterPalette repository.

import pandas as pd

url = "https://raw.githubusercontent.com/me9hanics/PainterPalette/main/datasets/artists.csv"
artists = pd.read_csv(url)
artists

The artists.csv file contains all information about painters, each row representing a painter, columns representing an attribute. An example of a few painters from the dataset:

From the data, we created networks of painters, styles, and movements. The types of networks we created are:

-style networks: Nodes are styles, two styles are connected if a painter painted in both styles, at least n times (usually n=150, the painters who have painted in both styles multiple times are counted multiple times)
-movement networks: two movements have a link if there are two paintings with the same style, but from different movements (same exact idea as previously, threshold at 100)
-painter networks: Painter influence network from the InfluencedOn data (“pre-defined”, an edge between two artists is created if in the Art500k dataset there was any info on one artist inspiring another on a painting), and a time-and-place network (which is dynamic in a sense) where two painters share a connection roughly if they were at the same place(s) around the same time(s). The reason why these two attributes were considered, and style not, is because these three are the major factors in what connections (network) an artist has, but if we were to analyze stylistic connections then the network would be biased (higher percentage of links between same style/movement painters).

Updated analysis: coming soon

Old analysis:

TL;DR:
Newer styles have more common artists, but they are less strongly connected. New artists explore more styles. Similarly, new movements are connected to more styles, but more loosely. The painter networks are (small world) and seem to have a (scale-free) property, even when looked at a large time period (600+ years). Naively building style and movement networks upon painters start to lose the scale-free property, the degree distribution is ruined by convoluted Poisson distributions (unless one creates higher level networks cleverly, see Villegas 2023: Laplacian renormalization group), they tend to be small-world only after thresholding to a sparse network. Artists both move to cities representative of their style (Homophily), and are influenced by where they live/work (Social Contagion).
Dynamically, we created a framework to modify linking methods to include a probabilistic time-interval component (each node can only gain new edges till it "dies"), from which we found that whilst triangular closure helps generating a network having a clustering coefficient closer to the real one, it approximately creates a less accurate network than a simple preferential attachment model. (Both methods outperform Erdős-Rényi, but the preferential attachment model seems to give a closer result the preferential attachment model with triangular closure.)

All in all, we can say that homogeneous movement connections, and similarly, "friend of friend" connections were much more prominent in the first few centuries, when there were only few types of movements in the dataset (gothic, renaissance (early, high, northern), baroque), the percentages of these ratios were high and even in the 18th century, the decrease was slow. However, from the 1800s, especially in the time interval 1850-1950 (the last 10 decades of the dataset), the ratios decreased very fast, the closed triangles ratio fell to around 12.5% and homogeneous connections now do not even make up 30% of all connections. This is likely caused by two things. One being the increase of international connections, as observed in the data that painters have more locations (visited more places, more distant places). Probably the deeper cause of this is the development in transport and transportation infrastructure. The other cause is the large increase in number of styles, even during one period, at any place. Probably the two events helped "develop" the other, painters being familiar with more styles as they travel, and more styles meant painters exploring more styles, likely causing an increase in growth of connections with painters in multiple styles, even international connections being made. These events helped the increase of inter-movement connections.
As for comparison to random graphs, we see that there is much more "organization", through all periods there are more triangles (& clusters) and more homogeneous connections than in random graphs. The next step would be to find what really drives this network's structure (how links are being made), and how to model it.

Since the networks seems to be scale-free, we could think it's driven by preferential attachment, but since preferential attachment linking does not create hubs, we may also think that some triangular closure connections are also driving the network. I'd assume this to be true, but from further analysis done, where I created a framework for linking models (Erdős-Rényi, preferential attachment, pref. attachment with 1 triangle closed each time) to only develop connections between nodes in a certain "time interval" in dynamic networks, and just from "dummy" analysis it seems that even if preferential attachment does not create enough triangles, a simple preferential model creates some better results. See notebook with slides and report.

Linking painters/people/entities together:
PageRank / Wiki Connections

Wiki Connections: partial dataset http://www.iesl.cs.umass.edu/data/data-wiki-links
smaller dataset: https://snap.stanford.edu/data/wikispeedia.html

Philosopher's web: Only available after paying 10$ for pro user
Philosophy NLP data: https://philosophydata.com/phil_nlp.zip

Network of the people connected to Francis Bacon, the network contains mostly born in the 16th century and are English so most philosophers in this list are not super relevant, there is no Kant, Nietzsche, etc. But a good example with a great visualization.

http://www.sixdegreesoffrancisbacon.com/?ids=10000473&min_confidence=60&type=network

import igraph as ig #To install: conda install -c conda-forge python-igraph  
people = pd.read_csv('datasets/SDFB_people_.csv')
relationships = pd.read_csv('datasets/SDFB_relationships_.csv')

#I used igraph, because it's faster than networkx, and graph-tool sucks on Windows
network = relationships.rename(columns={'id': 'relationship_id', }).drop(columns=['created_by', 'approved_by', 'citation'])
print(network.head(), '\n')
cols = network.columns.tolist()
cols = cols[1:3] + cols[0:1] + cols[3:]
network = network[cols]
network = network[network['person1_index'] != 10050190] #for some reason, there is no person with this id, I did a loop
# I used the documentation here: https://python.igraph.org/en/stable/generation.html#from-pandas-dataframe-s  this I followed
# this is important too: https://python.igraph.org/en/stable/api/igraph.Graph.html#DataFrame  
g = ig.Graph.DataFrame(network, directed=False, vertices=people[['id', 'display_name','historical_significance','birth_year','death_year']], use_vids=False)
print(g.summary().replace(',', '\n'))

filtered = g.vs.select(_degree = 0) #https://python.igraph.org/en/stable/tutorial.html#selecting-vertices-and-edges
g.delete_vertices(filtered)

import cairo #Needed for plotting #import cairocffi as cairo  # can do matplotlib too
#layout = g.layout(layout='auto')
#ig.plot(g, layout = layout) #ig.plot(g) #looks even worse

layout = g.layout(layout='reingold_tilford_circular') #kamada_kawai requires too much computing, 'fruchterman_reingold' is too dense
visual_style = {}
visual_style["vertex_size"] = 5
visual_style["vertex_color"] = "blue"
visual_style['bbox'] = (900, 900)
visual_style["layout"] = layout
#ig.plot(g, **visual_style) #Commented out because it takes big memory
# Needs improvement, but it's a start

https://global.health/ they got nice data on diseases, probably time-variant too, such as monkeypox, ebola.

Modeling of Biological + Socio-tech systems (MOBS) Lab: https://www.mobs-lab.org/

Website: me9hanics.github.io