Browse Prior Art Database

Interactive Visual Exploration for Multi-Faceted User Behaviors

IP.com Disclosure Number: IPCOM000187459D
Original Publication Date: 2009-Sep-07
Included in the Prior Art Database: 2009-Sep-07

Publishing Venue

IBM

Abstract

With the prosperity of online social networks on the internet, e.g. Facebook, LinkedIn and Live space, there are overwhelming requirements on the techniques of exploring latent information embedded in a social network. Such information is expressed by multi-faceted user behaviors, including pairwise communications as well as personal behaviors, such as writing an article in blog, commenting others’ articles/photos or sending a message/gift to others. Exploring the latent information behind these behaviors can facilitate community building (e.g., identify the role of a user) and foster opportunistic collaboration (e.g., sending customized advertisement to different communities).

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 35% of the total text.

Page 1 of 14

Interactive Visual Exploration for Multi-Faceted User Behaviors

As mentioned above, our method is comprised of three steps. We present the detail description of each step in following subsections:

3.1 Data Processing

Data processing aimed at acquiring the data representation of multi-faceted user behaviors in a social network. Such multi-faceted user behaviors may include both pairwise communications and personal updates.

We represent pairwise communication as adjacent matrix of communication graph, which is asymmetry considering the direction of user behaviors. In the matrix, each entry records the count of the communication behaviors. In the light of the diversity of user behavior, we may acquire a few matrices, where each one represents a certain type of pariwise communication. For the personal updates, they are represented in a diagonal matrix, where entries record the count of the personal update.

In this way, each kind of user behaviors is represented as a matrix. Alternatively, several similar types of user behaviors could be combined to an integrated one by summing related matrices up as needed.

3.2 Extract Overlapped Communities

After data processing, multi-faceted user behaviorsare represented as some matrices. In this step, latent overlapped communities and the relationships between these communities and users are extracted by inputting these matrices.

A lot of data mining methods could be adapted to extract the overlapped communities such as fuzzy clustering (e.g. [3] [4]). In our method, we developed a probabilistic model named U-U model to extract these latent information, which embeds the users and latent communities in geometric space. The relationships between communities and users are expressed in the probabilistic meaning bythe distance in the geometric space. Compared withother data mining method, U-U model facilitates the visual expression of the relationships between user and communities as we will mention in the visualization step.

1

Page 2 of 14

3.2.1 U-U Model

U-U Model is interpreted as a probabilistic model from user to user. We build the model by the intuitive consideration that user behaviors are accompanied with some specific intention/topic. We regard such intentions/topics as the latent communities, and each user behavior is interpreted as a process which is first choosing a latent community and then specifically choosing a receiver related to the chose communities.

Suppose we have N users and K latent communities, U-U model is a probabilistic model to find the correlation of users and communities with coordinates (

X

for users and

for communities).

The portion of relationships between users and communities are given as follows:

(1)

where )

F

|f is the user's portion of cij th community.

In summary, U-U model follows the following generating procedure for a set of users and commu...