7. CONCLUSIONS AND FUTURE WORK
In this work we approach the novel problems of characterizing sentiment evolution in a demographic group and identifying correlated groups, which address the large-scale sentiment aggregation. We design efficient algorithms for sentiment aggregation based on a careful indexing of time and demographics into hierarchies and demonstrate that our problems can be solved effectively on a large scale using clever pruning, top-k and compression methods.
Our approach allows observing sentiment behavior at a much finer level of detail than currently possible, helping to identify cases that are counter-intuitive and can only be observed by processing large amounts of data. Moreover, it enables an unprecedented scale-up of traditional social studies and raises new data analysis opportunities, useful for sociology and marketing researchers.
We outline some interesting problems and extensions of this paper, which we plan to work on. We consider only a disjoint type of relation, although it is possible to expand the notion of relations between groups to any arbitrary path in a demographics lattice, and use it as a filtering argument to our problems. Also we are investigating the case where disjoint groups appear to be the same sets of users due to a strict dependency among attributes. Filtering high correlations between such groups is possible when their sets of users are known and can be done as a preprocessing step. Alternatively, we can compare the volume of sentiments between these groups, which becomes possible since our DTree storage preserves this information.