Teamspector is a complex networks framework for experimenting with collaborative data. Currently it includes:
- An IMDb data extraction and pre-processing tools,
- An EDA notebook for exploring IMDb data,
- Scripts to build a social network from IMDb data, and to extract each movie's social and success metrics.
Teamspector needs MongoDB, and should be installed in its own virtual
environment. For installation instructions, see
INSTALL.md. After installing
Teamspector, download the IMDb data sources, extract the data to a local
MongoDB collection, run the pre-processing rules:
$ download imdbws $ extract imdbws $ preprocess imdbws
There is a smaller experiment that runs faster, just to test things. It has the
0. The main experiment has id
1. To run an experiment:
$ experiment imdbws <experiment id>
The IMDb dataset is provided by Amazon: https://www.imdb.com/interfaces/. This website describes in detail the data that's made available. All movies are considered, except for:
- Movies without year of release, - Not feature-length, cinema productions (`titleType != movie`), - Adult movies (`isAdult = True`).
During pre-processing, movies' number of votes (
numVotes) is converted into a
log_votes, as it has an exponential distribution.
The average ratings of movies are normalized into
nrating using a Bayesian
estimate, as there are wild differences in the sample sizes taken into account
to produce the average. We suppose the Bayesian estimate can be trusted only
for movies that received five thousands of votes or more, hence
only defined for movies with
numVotes ≥ 5000.
To calculate movie's success metrics, we check how well a movie did in terms of
nratings, compared to other movies produced in the same year.
The percentile of a movie's
nratings within movies produced
in the same year is computed as
To produce our final movie score metric,
summed. The movie's percentile of this sum is computed into
ypct. For this
percentile, we only consider movies produced in the same year. In other words,
if a movie has
ypct = 0.99, this means the movie is considered better than
99% of the movies produced in the year the movie came out.
Finally, a binary success metric,
top100, is defined to capture if the movie
was one of the best 100 movies produced in the year as measured by
only after 1985 that we have more than 100 movies each year that have a valid
nratings value, hence the
top100 metric is only defined for movies produced
this year or afterwards.
The goal of this experiment is to test the effect of social network structure in human work. The experiment infers the social network graph from a group of workers. Workers are nodes. Nodes who previously worked together are connected via an edge. Edges are unordered, and their strength is proportional to the number of previous works jointly conducted.
Edges and their nodes are removed from the graph after being inactive for eight years. Only productions in which workers are part of the graph's giant component are considered.
For each piece of work produced, we measure aspects of the social structure from workers responsible for the piece of work, and success metrics related to the piece of work. Afterwards, tests are conducted to measure the effect of social structure on work success.
In the context of movie production, works are actors, producers, writers and
directors. Pieces of work are movies, their success metrics are
build_network.py uses the IMDb dataset to build the graph of
workers. For each movie produced, network metrics from workers and success
metrics from movies are stored in a separated table.
The experiment runs in a loop, where each year is analyzed at a time. IMDb only provides the year of release of movies, forcing us to consider movies produced in a given year are all produced and released simultaneously.
We begin with an empty list of network graphs L.
Before each iteration of the loop, each graph in L is inspected and nodes that haven't produced any work for over 8 years are removed along with their edges. If this causes any of the graphs to become disconnected, the disconnected graphs are added to L, while the original graph is removed from L
Then, we look into the productions released in the year. For each production, we find the graphs in L which contain any of the production's workers. Workers not present in the graph are added to the graph as new nodes. Edges between workers are added, or strengthened in case they already exist. If we have selected more than one graph from L, the now connected graphs are joined into a single graph, which is added to L, and the original graphs which were joined are removed from L.
Experiment metrics are only collected starting from 1985, as success metrics
might not be reliable prior to this year. The experiment runs from 1985 to
2012. The reason 2012 was chosen instead of a more recent year is that we don't
know how long it takes after a movie is released for
stabilize. We assume five years is a safe pick.
For a large portion of movies, IMDb provides a list of producers, directors, writers and actors. For some famous movies, more elements from the production team may be available, such as editors, composers, and cinematographers.
Due to computational constraints, the experiment considers only producers, directors and writers. Adding actors, the graphs become exceedingly big for centrality network metrics to be feasibly computed.
For each production team, we collect social metrics related to each worker in the team individually (Ego metrics), to the pairs of workers in the team (Pair metrics) and to the team as a whole (Team metrics).
- Closeness Centrality,
- Betweenness Centrality,
- Clustering Coefficient,
- Square Clustering,
- Network Constraint,
- Node's degreem
- Past Ratings: mean of
nratingfrom node's prior works,
- Past Votes: mean of
log_votesfrom node's prior works,
- Past Experience: the number of prior works the node was part of,
Each production will have as many ego metrics as the number of people in its team. Hence, the maximum, minimum, median and standard deviation from the metrics are calculated.
Via vertex contraction, all nodes that participated in the production being studied are temporarily transformed in a single node. Then, all ego metrics are recalculated for the contracted node representing the team. Besides that, another metric is taken:
- Team size: the number of nodes in the production team.
- Shared Collaborators: the number of nodes connected to both nodes of the pair,
- Neighbour Overlap: the shared collaborators number, divided by the number of nodes connected to either one of the nodes of the pair, the proportion of friends which are connected to both nodes,
- Past experience: the weight of the edge connecting the two nodes,
Production teams with more than two members will have many pairs. Hence, the same aggregate statistics used for the Ego metrics are also used for the Pair metrics.
If you like this project and want to participate, there's a lot of ways you can
help. Check out
CONTRIBUTING.md for more info.