Online Opinion Correlation for Large-scale Demographics

1 Project Description

Nowadays, users on the social web are more than eager to express their opinion in the form of reviews,
ratings, tags, likes, etc. We are also witnessing the increasing availability of users' demographics data. As
a result, the need to provide ne-grained analytics of demographics opinion is growing, demanding ecient
and scalable methods of opinion aggregation and correlation. While several approaches exist for mining
opinions from product reviews or micro-blogs, little attention has been devoted to the online aggregation and
correlation of extracted opinions for dierent demographics groups (e.g., \Students in Italy") over time. In
our recent work [2, 3], we introduced methods for aggregating user ratings and tags and extracting meaningful
demographics patterns such as \Teenage girls consistently like Woody Allen Movies".

In this master's project, we aim to formalize a number of opinion aggregation and correlation functions
and explore their ecient implementation. The problem of opinion aggregation relies on two main points.
First, given a demographics group described by a set of opinion proles, the aim is to extract prototypes
best representing the main underlying opinion dynamics. For that, many strategies developed for static
data as well as clustering approaches for temporal data will be explored [6, 7]. The second focus relies
on the description of the extracted prototype dynamics. The challenges for these descriptions are : a) to
cover the variability, uncertainty and condence of the aggregated opinion within demographics groups, b)
to select the most informative and condent time periods by segmenting the opinion prototypes. For that,
interval-valued, histograms-valued data and temporal segmentation strategies can be investigated [1, 5].

This work aims to study families of metrics for temporal data as those developed in [4] to measure, in
a scalable way, correlations between demographics opinion dynamics; and to subsequently reveal the most
correlated ones. Finding demographics groups requires the exploration of all possible combinations of values
for demographics attributes - a task, that becomes quickly intractable, especially for pairs of groups. Our
algorithms in [2, 3] exploit the lattice structure induced by demographics attributes in order to prune the
search space. We propose to adapt them to the newly developed aggregation and correlation functions.

This work will serve as a basis for the denition of a number of problems and the exploration of their
ecient implementation. Examples of such problems are nding demographics groups whose opinion changes
over time, or nding pairs of demographics groups with correlated opinions over time, that is, groups that
react similarly over time to external events. We plan to co-advise a PhD student on this topic.

2 Related Work

Some recent studies have already made a step along this direction. For instance, \A Demographic Analysis
of Online Sentiment during Hurricane Irene" [8] revealed dynamic (temporal) sentiment dierences between
Southern USA and New England, and at the same time a constant (inherent) dierence in the sentiments
expressed by males and females. A similar problem of detecting correlations between multiple time series
has been approached in the area of data streams using a variety of techniques [11, 9, 10]. Among these, [11]
computes correlations using sliding time intervals of specied sizes, composed from a number of sub-intervals of a xed length. Their work is interesting in particular by the use of Discrete Fourier Transformation
(DFT) to compute correlations in an approximate and updateable manner.

3 Co-advisors

Ahlame Douzal is an expert in temporal and sequential machine learning and more particularly in temporal
segmentation strategies that will serve the formalization of sentiment aggregation and correlation task. Sihem
Amer-Yahia is an expert in large-scale indexing and mining of structured datasets, that will serve the task of
adapting existing algorithms to enable the newly developed opinion aggregation and correlation functions.

4 Prerequisites

The candidate should have basic knowledge of database management systems and data mining. Programming
skills including knowledge of Perl/Python will be a plus. We will be using the MovieLens 10 million ratings
dataset 1 that contains user demographics, movie attributes and ratings.

References

[1] A. Chouakria-Douzal. Compression technique preserving correlations of a multivariate temporal se-
quence. In M. R. Berthold, H.-J. Lenz, E. Bradley, R. Kruse, and C. Borgelt, editors, Advances in
Intelligent Data Analysis, volume V, pages 566{577. Springer, 2003.

[2] M. Das, S. Amer-Yahia, G. Das, and C. Yu. Mri: Meaningful interpretations of collaborative ratings.
In VLDB, 2011.

[3] M. Das, S. Thirumuruganathan, S. Amer-Yahia, G. Das, and C. Yu. Who tags what? an analysis
framework. PVLDB, 5(11):1567{1578, 2012.

[4] A. Douzal-Chouakria and C. Amblard. Classication trees for time series. Pattern Recognition,
45(3):1076{1091, 2012.

[5] A. Douzal-Chouakria, L. Billard, and E. Diday. Principal component analysis for interval-valued obser-
vations. Statistical Analysis and Data Mining. Wiley, 4(2):229{246, 2011.

[6] A. Douzal-Chouakria, A. Diallo, and F. Giroud. Adaptive clustering for time series: application for
identifying cell cycle expressed genes. Computational Statistics and Data Analysis, 53(4):1414-1426,
2009.

[7] A. Douzal-Chouakria, A. Diallo, and F. Giroud. A random-periods model for the comparison of a metrics
eciency to classify cell-cycle expressed genes. Pattern Recognition Letters, 31:1601-1617, 2010.

[8] B. Mandel, A. Culotta, J. Boulahanis, D. Stark, B. Lewis, and J. Rodrigue. A demographic analysis of
online sentiment during hurricane irene. In Proceedings of the Second Workshop on Language in Social
Media, pages 27-36, Montreal, Canada, June 2012. Association for Computational Linguistics.

[9] S. Papadimitriou, J. Sun, and C. Faloutsos. Streaming pattern discovery in multiple time-series. In
VLDB, pages 697-708, 2005.

[10] S. Papadimitriou, J. Sun, and P. S. Yu. Local correlation tracking in time series. In ICDM, pages
456{465, 2006.

[11] Y. Zhu and D. Shasha. Statstream: statistical monitoring of thousands of data streams in real time. In
VLDB, pages 358{369, 2002.