Research

How you act is who you are? Predicting personality with social behavior

Fri, 20 Jul 2012 08:38:15 -0400

One’s personality impacts many aspects of her interactions with the outside world. The impact of personality on job performance is well studied. It also has an impact on personal relationships. There is a growing body of work in trying to predict one’s personalty based on their activity in social media sites likes Twitter and Facebook [1,2,3]. This work predominantly analyses the textual content and in some cases the ties between individuals.

We ask a different question: can we predict personality based on one’s social behavior? Can we use how people act towards as indicators of personality? While personality impacts our behavior, personality aspects do not show up in all situations. An aggressive person will not act aggressively in all situations. In fact, there are many situations where act like others. Studies have looked at when distinctive characteristics of our personality are more visible. Research indicates that in environments that satisfy one’s basic psychological needs like relatedness to others, show of competence, and autonomy, “people tend to act like themselves”. We then ask the question whether Twitter satisfies these needs?

Can we predict personality based on how people behave on Twitter? The answer is yes, we can predict personality as well as you would with analysis of textual content of Tweets and relationships of an individual. In our recent paper (asonam_cameraReady.pdf) that will appear in the Asonam 2012 Conference proceedings, we consider a set of indicators that look at how much one’s behavior varies towards different friends and what type of friends she has by looking at friends’ behavior. For behavior, we measure a large range of actions like the bandwidth of total and reciprocal communication, the amount of message forwarding, the amount of priority given to friends and delays incurred, etc. We also compute a set of detailed textual features used in the previous work. We then predict personality using our features vs. the textual features against the results of a big five personality survey. We find that we can predict personality as good as any textual analysis tool.

Here are some behaviors that are best predictors of different personality traits:

• Neuroticism: anxious, insecure, sensitive. Neurotics are moody, tense, and easily tipped into experiencing negative emotions. Behavior that best predicts neuroticism is a lot of variation of behavior such as high variance of message text and response time, sending messages that are not propagated and having friends who do not propagate messages through retweets.

• Extroversion: outgoing, amicable, assertive. Friendly and energetic, extroverts draw inspiration from social situations. Extraverted individuals write long messages, but they typically tend not to be propagated by others.

• Agreeableness: cooperative, helpful, nurturing. People who score high in agreeableness are peace-keepers who are generally optimistic and trusting of others. Agreeable individuals tend to follow the norms of Twitter like mentioning topics, they tend to respond uniformly to friends and have friends who are responsive.

• Openness to Experience: curious, intelligent, imaginative. High scorers tend to be artistic and sophisticated in taste and appreciate diverse views, ideas, and experiences. This trait is best predicted by low variance of timing between tweets and almost uniform propagation behavior from friends.

• Conscientiousness: responsible, organized, persevering. Conscientious individuals are extremely reliable and tend to be high achievers, hard workers, and planners. Conscientiousness is highly correlated with a lot of behaviors that indicate that people with this trait tend to act like others.

Overall, the timing between messages, text length and propagations appear to be the most informative features for understanding personality.

[1] J. Golbeck, C. Robles, and K. Turner, “Predicting personality with social media,” in Proceedings of the 2011 annual conference extended abstracts on Human factors in computing systems, CHI EA ’11, (New York, NY, USA), pp. 253–262, ACM, 2011.
[2] J. Golbeck, C. Robles, M. Edmondson, and K. Turner, “Predicting personality from twitter,” in Proceedings of the 3rd IEEE Interna- tional Conference on Social Computing, (Boston, Massachusetts, USA), pp. 149–156, 2011.
[3] D. Quercia, M. Kosinski, D. Stillwell, and J. Crowcroft, “Our twitter profiles, our selves: Predicting personality with twitter,” in Proceedings of IEEE SocialCom, 2011.

ABC: Attentive Betweenness Centrality

Tue, 10 Jul 2012 22:10:47 -0400

Who are the most important people in a network? There are many different measures that describe the importance of a person based on her relationships and where these relationships place her in the network. For example, one well-known measure of betweenness considers how crucial a person is by looking at whether they are on a critical path for others in the network. If a person is on a large number of shortest paths between people in the network, then they are important. For example, in the above example, node A and B are on the shortest path between nodes X and Y. In fact, they are on the shortest path between all nodes in the left and right subgraphs. In essence this means that if X wants to send a message to Y, then A is the best conduit. This gives nodes A and B a crucial advantage. The betweenness measure is based on this idea: it finds the fraction of shortest paths between any pair of nodes that pass through a given node.

The problem with betweenness is that it disregards path that are almost the shortest path. For example, paths that pass through E,C,D are not as short as those that pass through A and B. But, these paths are still valuable. Especially if node A becomes unavailable, they provide a crucial alternate path. However, in betweenness centrality, nodes E,C,D will get almost no credit. To address this problem, we introduce a new measure called ABC centrality. Instead of looking at whether a node lies on the shortest path, it looks at the amount of flow that can pushed through a node. But, the longer the path, lesser the amount of flow. In fact, the flow gets split across all the outgoing edges and reduced by a factor of alpha at each step. To see an example, let’s look at the following figure:

Suppose node A wants to send flow to node B along the rightmost path. The flow will be divided by a half because node A has two outgoing edges. Now, the flow is not divided further. But, at each step, it attenuates by a factor of alpha (a). As a result, total flow reaching B along this path is alpha square divided by 2. Now, we compute the importance of the nodes along this path by looking at how much each node contributed to this flow. This process gives credit to all the nodes for the different paths they are on. ABC betweenness simply adds the credit for a node along all the different paths.

Our algorithm appears in the SocialComm 2012 Conference. We are not the first to recognize the shortcoming of the betweenness measure. In our paper (SocialComm2012_ABC_Centrality.pdf), we show that our algorithm better approximates the original betweenness measure than any of the other proposed algorithms, and at the same time overcoming the new problems introduced by these algorithms. The code for our algorithm can be found here.

For example, for the well known Karate Club dataset, our algorithm captures the nodes at the center of the two main factions very well as shown below.
Plus, the parameter alpha allows us to adjust how much attention we pay to the length of the paths. So, what types of properties do these nodes with high betweenness have? We find that nodes with high ABC scores are more diverse. For example, for the Internet Movie Database (IMDB), the actors with high ABC values also star in movies in many different genres. They tend to act as a bridge or connector between actors who specialize in a specific genre. Here are some of the actors with the highest ABC scores. If you do not know who they are, we guarantee you that you will recognize each actor as soon as you see a picture of him:

Actions Speak as Loud as Words

Wed, 1 Feb 2012 23:28:47 -0500

Did you ever think about what types of words you use when speaking to a close friend? How about a colleague who is not so close? When speaking in a professional context, we use different types of words than we are engaged in personal communication. Words tell a lot about our relationships. We can look at words to understand what types of relationships we have with our friends. In fact, words are even used to understand the personality of individuals. For example, the word cloud above shows the types of words more common among friends.

How about actions? What can we tell from the actions? While different users may have hundreds of friends and followers on Twitter, they may not behave/act the same towards them. Some friends they may immediately reply to and some friends they never do. Some friends may forward their messages and some do not. The timing, the reciprocity and the sheer number are all indicators of our relationships with our friends. In our recent WWW paper (Adali_WWW2012.pdf), we develop a large number of indicators of friendship based on actions and a new methodology for comparing actions to words. We show that actions are as good as words in understanding relationships. We can find the same two type of relationships from either actions or text: intimate and formal relationships.

We look at actions and text together and ask the following question: Given just the actions, what are the dominant textual categories that best differentiate between different actions? It turns out, there are exactly two such categories. The graph we found using this method is shown below. The categories to the left are the more formal relationships, while the categories to the right are the more intimate categories. We can also find the dominant behavior for each text category.

We also ask the same question in reverse, looking at just the text, what types of action are most significant in our study. It turns out action to text mapping returns the same two groups either way: the same action maps to same groups of text categories. In fact text and action are equally powerful in understanding the relationships. There are two main categories: intimate and formal relationships. Below is the word cloud for our more formal relationship category.

Time Based Models of Trust

Sat, 1 Oct 2011 23:28:46 -0400

Trust has been studied in many different contexts, but is coming into focus again in social networks. A great deal of information is exchanged today on social networks. Sometimes this exchange is between people who know each other and sometimes it is not. In any case, people are forced to make quick decisions about trust in many social contexts based on limited information. We examine the impact time plays in trust models.

In our first piece of work, we consider a model that incorporates factors impacting trust from many different time scales. At the lowest level, we consider the impact cognitive heuristics and biases have on trust decisions. These are the quickest factors that are included in the trust decision. The second level considers the more deliberate, system 2 type, trust decisions based on prior experiences with the trustee and other social factors. This is further followed by additional data that can be gathered from the network to support or refute the trust decision, if the trustor chooses to engage in them. This work is presented at the Trust in Agent Societies Workshop held in conjunction with AAMAS 2011.

In our second piece of work, we consider the case where the trustor has to incorporate two factors into the decision. The first one is called competence, it measures how capable is the trustee to accomplish a task. The second one is called willingness, it measures how much energy or bandwidth the trustee is willing to allocate to the trustee. Both factors play a significant role in deciding whether to trust someone for a time-critical mission. For example, if the objective is to spread information to the network quickly, the right combination of both factors is needed. We develop models based on two factors and show how they can be incorporated into decision making. This work is being presented at the 2012 IEEE Conference on Cognitive Methods in Situation Awareness and Decision Support, COGSIMA 2012.

Prominence in Social Networks

Tue, 1 Mar 2011 23:17:54 -0500

in this work, we consider prominence computation in social networks. Instead of only considering relationships between people, we ask a different questions: What about the links between the objects people create? How are they connected? People collaborate on objects and these objects form natural groups: like research collaborations, research areas or venues they appear in. We first use data mining algorithms to find the natural groupings between objects. These groupings show us that prominent people tend to belong to prominent groups with prominent objects.

Using this intuition, we compute prominence using an iterative algorithm. We show that our algorithm beats in performance using many different measures of performance many well known algorithms like hits, pagerank and various centrality measures for many different data sets like the Internet movie database (IMDB), Enron mail data and Academic research collaborations (DBLP).

The preliminary version of this work appears in ICWSM 2011.