The use of social media has been rapidly increasing in the last years. Social media, such as Twitter, has become an important source of information for a variety of people. The public availability of data describing some of these social networks has led to a great deal of research in this area. Link prediction, user classification and community detection are some of the main research areas related to social networks. In this paper, we present a user modeling framework that uses Wikipedia to model user interests inside a social network. Our model of user interests reflects the areas a user is interested in, as well as the level of expertise a user has in a certain field.
We provide TwikiMe! as a framework to generate semantically enhanced profiles for
Twitter users. To accomplish it, we use the well established Wikipedia corpus as a
semantic knowledge base. Wikipedia is arguably the most accessed reference Web site
and each of the over 3.5 million existing articles are manually classified by human
curators to one or more categories. Additionally, categories are organized in a graph
where sub-categories reference top level categories. The English Wikipedia has in total
23 top level categories (Main topic classifications), which we use to represent a user
profile (See Figure 2 for the list of top-level categories). Abel et al. [
The creation of semantically enhanced user models consists of three stages (see
Figure 1). The first stage, Extraction, entities are extracted from the user’s tweets. Given
Twitter users streams, we annotate all tweets to detect any mention of entities that can
be linked to Wikipedia articles. In this step, we chose to use the WikipediaMiner [
The second stage, Categorization, we extract the categories of each entity that has been mentioned in the users Tweets. For each category, we follow the path of all parent categories, up to the root category. In some cases, this procedure results in the assignment of several top level categories to an entity. A weight is calculated for each category by first setting a value of 1 for the detected entity. Following the parent categories (which are closer the root category) we divide the weight of each node by the number of siblings categories, resulting in each entity receiving 23 categories’ scores. In the final stage, Aggregation, we perform a linear aggregation over all of the scores for a tweet in order to generate the final user profile.
Our approach of reducing the user profile to 23 topics di ers significantly from the
work of Michelson and Macskassy [
We applied our user profiling method on the Twitter dataset created by Yang and
Leskovec [
Additionally, we calculated how much the user profile changes over time by splitting the user’s tweets into two bins and calculating the inter-user similarity. To perform such evaluation we divided the users’ stream in half (by number of tweets), comparing the profile of the earlier tweets against the latest ones. Table 1 shows the results of the similarity measures. As metric we used the cosine similarity.
Our result show that while the categories related to the tweets of one user stay the same over the time (leading to a very high similarity of 0.85), we see that the average similarity between the user and her friends and followers is rather low with 0.26 and 0.27. Comparing the profiles of the followers and friends we obtained a similarity of 0.24. Based on the high similarity when comparing only tweets from one user, we assume that the topics one user tweets about stays rather constant, and are not necessarily the same as the topics discussed by friends and followers. 3
To ilustrate our approach, we deployed an online version1 of the TwikiMe! framework. The interface guides the user through the twikifying process. First, one must provide a valid Twitter user name and the amount of past tweets to be annotated (due to the outsourcing of WikipediaMiner service, this step is the most time consuming). Once the process is finished, the user is asked to trigger the Categorization step. Finally, the user is presented with the TwikiMe! chart of the given Twitter user.
To address qualitative comparisons, we also provide a comparative graphical interface as depicted in Figure 2. In the comparative chart, it is possible to visualize the percentage results for two distinct users, side-by-side. Additionally, the interface provides a date picker to restrict the comparisons to a given time interval. In Figure 2, one can see the comparison between the Twitter accounts of the Dalai Lama (@DalaiLama) and president Barack Obama (@BarackObama), during the period of April through December 2011. The resulting user profile depicted in the chart shows a clear dominance of Barack Obama in the topics of “people”, “politics” and “business”, while Dalai Lama leads on “life”, “culture” and “geography” topics.
A closer look into di erent “known” users (politicians, researchers and technology news), shows that the resulting TwikiMe! profiles provide expected topic identities. While technology news providers lead on the topics of “business”, “society” and “technology”, researchers lead on the topics of “education” and “applied sciences”. We invite 1 http://twikime.l3s.uni-hannover.de the readers to try it by themselves our TwikiMe! prototype at http://twikime.l3s. uni-hannover.de. The prototype makes available all previously twikified users for comparisons. 4
In this paper, we introduced the TwikiMe! A prototype for generating comprehensible user profiles that are represented compactly as a 23-length vector. The profiles quantifies the users’ implicit interest in each of these di erent categories. We believe that, by exploiting the TwikiMe! profiles, we are able to improve content and user recommendation on Twitter.
As future work, we plan to empirically demonstrate the proof of concept of the generated user profiles as well as improving the strategy of the Categorization step. Finally, we believe that there may be a significant di erence between what a user “produces” and what a user “consumes”. We also plan to evaluate this di erence by generating TwikiMe! user profiles based on tweets that a user follows, rather than just the tweets that a user writes.