joomla statistics

Social Media and Network Mining - Models, Systems and Applications

Tutorial Date and Time: 3:30pm to 5:30pm GMT+8, May 14th 2014.
Venue: Shangri-La's Far Eastern Plaza Hotel 89 Section West, University Road, Tainan City 70146, Taiwan


Mining social media and social networks are now important research topics due to the increasing presence of users' online activities. The users' data generated through these activities represent a wealth of information for business platforms to tap on for marketing and decision making purposes. Similar to other big data research, conducting social media/network data analytics research for consumer insights and turning insights into useful business applications are challenging. In this tutorial, we shall describe some ongoing research efforts to address the above challenges. We shall cover three main topics, namely: (a) social media and network mining framework, (b) social media and network mining techniques, and (c) social media analytics systems and applications. We shall also showcase some working social analytics systems for Twitter data. The tutorial does not assume any data mining or machine learning background. It is tailored for both researchers and practitioners who would like to venture into the social media and network analytics area.

Lecturers: Ee-Peng Lim and Freddy Chua

Ee-Peng Lim and Freddy Chua
Slides in PDF format
Slides 1 by Prof. Ee-Peng Lim
Slides 2 by Dr. Freddy Chua

Topics of Discussion

Social Influence

Social influence can take on different meanings depending on the context of discussion. We base our discussion of social influence on users' items adoption data while preserving the key concepts common to other domains. We say that user x socially influences user y only if user y adopts an item because user x has adopted the item. This simple statement has several implied conditions: We will briefly discuss these three aspects of social influence in our tutorial.
Users' Items Adoption Graph

Static Latent Space Modeling

In users' item adoption data, the number of possible items for users to adopt is very large and can be in the order of millions. A naive way of representing users' behavior is to use a vector with dimensions similar to the number of possible items, with entries representing the raw frequencies of adoption. Unfortunately, using high dimensional vectors for users' behavior would lead to expensive computations during the comparison between users. Comparing only the raw frequencies would also be less informative since we ignore the co-occurrence relationships between different items adopted by the users. A widely adopted method of reducing the computational costs is to perform dimension reduction on the users' item adoption data to obtain vectorized representations in lower dimensions such as order of tens or hundreds.
Static Latent Space Modeling

Static Social Correlation

The static social correlation measure is a quantity between every pair of users that optimizes the likelihood of observing the users item adoption data beyond non-social factors can capture. To model an observed count of adoptions for an item by a user, the non-social approach is to use only the item's latent factor and user's latent factor, while the static social correlation approach is to use the item's latent factor, the user's latent factor as well as the latent factors of the user's friends. The extent of how much the user rely on each of her friends would depend on the user's static social correlation with each of them.

Temporal Latent Space Modeling

A direct extension of static dimension reduction methods on temporal data is to apply any of the static dimension reduction methods on each time step independently of other time steps. But due to temporal sparsity problem where some users have low or no activity in some periods, applying dimension reduction on each time step independently would lead to unrelated latent factors across the different time steps. We would discuss how to perform Dynamic Matrix Factorization that is a generalization of Linear Dynamical Systems in this tutorial.
Temporal Latent Space Modeling

Temporal Social Correlation

Given the time series obtained from Dynamic Matrix Factorization, we then extend the concept of static social correlation for the temporal case.
Temporal Social Correlation