Master thesis presented in December 2015 at DCC/UFMG (Departamento de Ciência da Computação da Universidade Federal de Minas Gerais) by Wladston Ferreira Filho.

The source code used to produce this research is available at https://github.com/wladston/teamspector

Outline

Structure

• Chapter:Introduction
• Chapter:Fundamental Concepts
  • Section:Social Network Analysis
  • Section:The iMDB Dataset
  • Section:Machine Learning
• Chapter:Related Work
  • Section:Loosely Related Work
  • Section:Closely Related Work
• Chapter:Interpreting Movie Data and Predicting Success
  • Section:Interpreting the Dataset (Network model, data filtering, success parameters, characterization, performance groups)
  • Section:Defining Features for Movies (movie features, social features, feature-combination, characterization)
  • Movie Success Prediction Model (evaluation methodology)
• Chapter:Analysing the Success Prediction Model
  • Section:Experiments (feature selection and strength, experimental setup)
  • Section:Results
• Chapter:Conclusion and Future Work
  • Section:Contributions
  • Section:Discussion
  • Section:Difficulties Tackled
  • Section:Future work
  • Section:Publications

Overview

[INTRO] Regarding collaboration, forming teams that perform well is an important problem because…. Social elements is key to the problem because…. It's easy to form the network, and research shows links between social elements and team performance. However these are not complete because… A novel approach is… It needs a large dataset, so the iMDB is ideal because… The analisys is non-trivial because of_. We found our model can improve the ways we currently understand and predict team success in this context, and our contributions are …, the rest of the work is organized as…

[FUNDAMENTALS] The concepts from SNA mentioned in this work are (graph, bipartite graph, one-mode-graph, topological metrics, clique, node contraction). Using SNA a bipartite network can be projected into a one-mode network. Special handling is needed when using SNA metrics in projections from a bipartite graph. We can access social characteristics from cliques in a social graph by aggregating metrics from clique's nodes or by getting metrics from the clique's contracted node. IMDb describes movie production generously, but has many amateur movies that just add noise. Its movie co-authorship data is directly mappable to a bipartite graph between movies and agents. iMDB also provides these extra information from movies with are known to be correlated to movie success. The ML concepts mentioned in this work are….

[RELATED WORK] These works find how to better form teams, these works find how to predict team success, both do it with a different approach. These works use a similar approach, but they have these limitations we don't. Our results alone aren't better than previous work but can be combined with other predictive models to previous works to improve them.

[MODEL] This is our model for a social network of movie producers. This is how we filter the dataset. This is how we define movie success. This is how we group movies based on success, and here is the characterization of the groups. Tis is how we define topological features from teams (based on previous studies we use these other features for control), and here is the characterization from such features. Here is how features interact. Here is the final ML model we created for predicting movie success.

[ANALYSIS] We proof that our prediction model works because of…. We know that our features are adequate for prediction because…. These were the experiments we made and its setup characteristics, and these are the results. These are the meanings for all these results.

[CONCLUSION] Here is a summary of what we did, and here is a summary of the new things we found. Here is how the field is changed in result of our work. Here is the reasons we think are behind what we found. Here are the greatest difficulties faced, and how we managed them. Here is how we think the work should continue. Here are the publications generated from this work.

Citation Reference

Not So Similar

• SNA
  • newman2001structure: #sna Main article on deriving social networks from bibliographic data and characterizing it using social analysis metrics.
  • lutter2013there: #movie #sociology #sns Studies career advancement in IMDB to show that some network metrics are related to better career advancements, and that woman are more lilely to drop out of their careers in comparison to man.
  • wasserman2015cross: #sns #significance Studied citations between movies in iMDB to estimate a significance of a movie production. Used the inclusion in the us library of congress national film registry and correct indicator of high significance, found that the metric is performs much better than since expert, and better than groups of experts.
  • elberse2007power: #movies #performance #prediction Assess the impact of famous actors in movie success.
  • fields2011analysis: #sna #recc Studied social net. of musicians, found that they organize according to their similarity, so social parameters can be used for recommendation.
  • stokols2008: #team Studies scientific production and finds team-based work in science is extremely important for scientific breakthroughs.
  • grund2012network: #performance #inter-team-sna Performed social network analysis in football teams and found that network characteristics in graph modeling the way players pass the ball is related to team winning odds.
• Team Formation
  • GunnA15: #team-formation #deterministic #optmization Studies how to better form on-the-fly teams on robots working on human rescue in disaster zones.
  • tseng2004novel: #team-formation #nodes-qualities #optimization Studies optimization techniques for forming optimum teams by combining elements with different habilites.
  • wi2009team: #team-formation #optimization #nodes-qualities Studies optimization of team formation is business, by picking manager and team members based on node specialization and keywords from team's objective.
  • anagnostopoulos2012online: #team-formation #optimization #nodes-qualities Technique for efficiently forming teams with a minimum skill set. They use the iMDB database to emulate a problem in with teams needs to be formed from directors and actions, where genres of movies they have previously acted serve as the skill.

Similar

• Performance
  • chen2010impact: #sna #small-world #performance Impact of small world on patent production, n=16 countries.
  • schilling2007interfirm: #performance #sna Studied n=1106 firms and found that those with dense connections to other films and low closeness are more likely to be innovative.
  • nemoto2011social: #sna #social-capital #performance Finds that when editors with more social capital (as defined by social network analysis metrics) work on an article, it reaches higher levels of quality faster.
  • LiLY2013: #sna #social-capital #performance Social capital can be estimated with metrics from social networks analysis, and researchers with more social capital publish research that have more impact
  • PapagelisMZ11: #sns #influence #degree Studied Flickr, found that the more connections a node has, the more influential it is, regardless of node credibility.
  • singh2011network: #performance #sna Specific kinds of network ties among open source developers are correlated with the development of more popular open source projects.
  • Rebehy2013: #performance #structural-hole #sna Studied 17 real state agents and found that their structural network hole constraint has a strong negative correlation to worker performance.
  • Burt04: #sns #performance Main article about structural holes.
  • burt2005brokerage: #sns #performance #structural-holes #social-captal Studies how sns and structural holes and be used to explain competitive advantage and social capital.
  • burt2009structural: #sna #performance #structutal-holes Shows that competitive advantage is created from differences in access to structural holes.
  • uzzi2005collaboration: #small-world #sna #performance #team Studied network of Broadway musical producers and found that small world coefficient in the network is related to the production of better work.
• Prediction
  • OghinaBTR2012: #movie #performance #social-activity-mining Studied 70 movies, extracting textual and social network activity for predicting the rating the movie would have in iMDB.
  • Ghiassi2015: #movie #prediction #performance Pre-production movie revenue forecast using artificial neural networks.
  • KimKHC13: #movie #forecast #social-activity-mining Shows that mining a large number of comments from social network services can capture the public opinion related to a movie, and help forecast whether the movie would be a hit.
  • COSN2013: #sns #prediction Uses financial data to predict success of kickstart fund-raising campaign, and an extra set of features extracted from social characteristics of backers and social network activity on twitter. They find that doing a combined predictor using both features yields the best results, with the social features providing a relative improvement of 4%.

Reference

On producing a good thesis https://www.ece.nus.edu.sg/stfpage/eleamk/phd/phdth2.html