I recently attended the 2014 International Conference on Social Computing, Behavioral-Cultural Modeling, & Prediction in Washington, D.C., where I learned various ways of modeling and making predictions with social-science and social network related problems. Here are some papers that were presented during the conference that I found relevant to Tagged:
1. Multi-objective Optimization for Multi-level Networks
by Brandon Oselio, etc., University of Michigan
The authors analyzed a multi-layered (multiple heterogeneous types of connections between nodes) network as a single entity, and applied multi-objective optimization methods to those networks. They used Pareto optimality to solve the multi-objective optimization problem, with a proposed algorithm for finding an approximate Pareto front in a multi-layer graph bisection (through spectral clustering method).
How we might use this at Tagged: This could be used to solve social discovery problems through an optimization fashion (e.g., developing a recommendation system that considers user behavior+user preference+match distribution+friend connection+messaging).
2. Using Trust Model for Detecting Malicious Activities in Twitter
by Mohini Agarwal & Bin Zhou, University of Maryland
The authors developed a heterogeneous social graph using Tweets and presented an extended trust model on the graph to propagate trustworthiness (with two ways of defining the scores: “Normal Trustworthiness Score” and “Biased Trustworthiness Score”) in the graph. They used the scores to detect malicious activity on Twitter using backward propagation and achieved high accuracy.
How we might use this at Tagged: This would be an interesting method to measure the trustworthiness of users on Tagged and build either product or back-end features.
3. A New Approach for Item Ranking Based on Review Scores Reflecting Temporal Trust Factor
by Kazumi Saito, etc.
Item ranking methods typically rely on the number of reviews or the average review score. This method reflects trust levels by incorporating a trust discount factor into a temporal time decay function. The work then involves using multinomial distributions with the trust discount factor, and z-scores to renormalize variances.
How we might use this at Tagged: This is another way we could define trustworthiness with scoring and recommendations.
4. Identifying Users with Opposing Opinions in Twitter Debates
by Rajadesingan, etc., Arizona State University
This is a semi-supervised framework using a retweet-based label propagation algorithm coupled with a supervised classifier to identify users with differing opinions. This retweet-based label propagation (ReLP) framework looks like this:
The model drastically reduced the manual effort involved in constructing a reliable training set utilizing users’ retweet behavior, and achieved high accuracy.
How we might use this at Tagged: This framework could help us to differentiate users with opposite opinions on the same topic. Great products/features can be developed based on this concept. The label propagation concept in general is also beneficial to us in solving semi-supervised or unsupervised problems.