Introduction

With the emergence of the social Web in the mid-2000s, the Web has evolved from a static Web, where users were only able to consume information, to a Web where users are also able to interact and produce information. This evolution which is commonly known as Social Web or Web 2.0, has introduced new freedoms for the user in his relation with the Web by facilitating his interactions with other users who have similar tastes or share similar resources. Examples of Web 2.0 include social platforms and networks (such as MySpace, Facebook, and LinkedIn), collaborative tagging sites (like delicious, CiteULike, and Flickr), microblogging sites (like Twitter and Yammer), and even crowdsourced databases (e.g., Bioinformatics sequence databases such as GenBank or UniProt).

These platforms are commonly used as means to communicate with other users, exchange messages, share resources (photos, videos, or bioinformatics sequences), comment, tag, maintain profiles, interact and play online games, etc. In addition to dedicated social platforms, traditional content providers like newspapers, tend to be more social since they provide to users means for sharing, commenting, and linking documents together, e.g., sharing. These collaborative tasks are among the most important factors for the increasingly quantity of available social data. Hence, the need to exploit this unprecedented quantity of data is critical for many companies and their business. This has led to the development of Data-driven Decision-making techniques to extract valuable knowledge from this data, based on fast, effective and scalable data analysis and machine learning algorithms, which are at the heart of my research interests focusing on descriptive, predictive, and prescriptive analytics of large-scale data.

Specifically, my research addresses data and information quality in crowdsourced databases, as well as the crucial problem of enabling users to consume relevant information with respect to their interests and needs. This task is commonly realized through content recommendation (such as movie Netflix, books on Kobo or products on Amazon) or Information Retrieval ( on platforms like Google or Bing). However, conventional models of recommendation and Information Retrieval don’t consider this side social information, and proposing novel extensions are also at the heart of my work. As two exemplars of my research:

• In Social Information Retrieval [1], I have extended conventional Information Retrieval models in order to incorporate side social information. Specifically, I have introduced a novel ranking function [2], a new Information Retrieval modeling schema [3, 4], and a query expansion algorithm [5, 6] all based on innovative social features. These contributions have been integrated into a social web search engine called LAICOS [7], which represented some of my major contributions to the area of Social Information Retrieval.
• In data quality for collaborative bioinformatics databases (e.g., [8, 9, 10, 11]), I have addressed the problem of identifying low-quality and suspicious records based on literature analysis. These databases contain hundreds of millions of records derived from submissions by individual laboratories, as well as from bulk submissions from large-scale sequencing centers. Their diversity and scale means that they suffer from a range of data quality issues including errors, discrepancies, redundancies, ambiguities, and incompleteness. In that context, I have combined a set of information retrieval and machine learning techniques while using the published literature as a background knowledge to detect literature-inconsistent records.

With the increasing level of business competitiveness, investigating more intelligent and faster ways of analyzing, exploring, using, and mining the data is a key aspect for many businesses. Hence, my objective is to go deep in the development and the usage of data analytics in different areas from both academic and industry perspectives. To elaborate in more detail, following is a summary of my ongoing research that enables this large-scale data processing for information retrieval and recommendation.

Mining Massive Media

Social media sites such as Twitter present a double-edged sword for users. On one hand these sources contain a vast amount of novel and topical content that challenge traditional news media sources in terms of their timeliness and diversity. Yet on the other hand they also contain a vast amount of chatter and otherwise low-value content for most users’ information needs where filtering out irrelevant content is extremely time-consuming. My previous work on twitter topic classification [12] has noted this need for topic-based filtering and has adopted a range of variations on supervised classification techniques to build effective topic filters. However, there remain many fundamental research questions that need to be answered on a longitudinal study of the performance of such supervised topic classifiers. For example, can we identify temporally unstable features? How can we design a learning algorithm that automatically downweights the influence of temporally unstable features?

Visualizing large-scale Data for Information Retrieval

In many applications that involve searching geo-temporal data, results are often displayed on a map or other interactive visual interfaces. However, in many settings of interactive visual search, it is time-consuming to individually examine all matching results. This suggests the need to aggregate results, for example, via an unsupervised clustering method like K-means. These ideas build on the “cluster hypothesis” of information retrieval, which posits that documents in the same cluster are likely to address similar relevance needs for the search task. Unfortunately, the use of unsupervised methods does not necessarily guarantee that the clusters themselves are relevant. Therefore, it is critical to develop relevance-driven clustering methods for Visual Information Displays (VIDs) that are capable of displaying highly relevant clusters for queries.

Since “retrieved” information elements are not individually selected, but rather clustered using the relevance signal of each element, the problem is clearly an optimization problem of how to restrict cluster settings to show the user the most relevant information to his query. While this notably diverges from the standard information retrieval setting where ranked documents are chosen individually, these additional constraints do not change the overall objective to select relevant information given the user’s information need as represented by the given relevance measure. In this work, I intend to provide a unified framework that abstracts presentation-specific details of different information clusters for VIDs and facilitates an optimization perspective of clustering w.r.t. a provided relevance measure. I also intend to define a new specific objective function for optimization that is suitable for this visualization task. Finally, I aim to propose new, fast, effective and scalable algorithms that optimize the objective function that I define for this visualization task.

Machine Learning and Social Media Analysis for Financial Analytics

Social media platforms such as Twitter or StockTwits are widely used for sharing advice between investors, traders, and entrepreneurs. The popularity of these platforms, the vast amount of information generated by active members, and their usage by influential people imply that this social media content is often predictive of financial markets.

In this project, I plan to develop novel algorithms to model the influence of social media on stock market performance. These novel methods are based on machine learning principles that will provide financial analysts with powerful tools for summarizing social media sentiment and opinion regarding individual stocks as well as providing quantitative predictions for use in investment and trading decisions.

Specifically, the research questions I intend to answer in this project are: how can we model and capture the influence of users, brand and industry mentions over time? How can we deal with varying reliability of users for different stocks, or users with non-stationary (i.e., fluctuating) influence over time? How can we automatically identify brands that may be relevant to a stock ticker to reduce the number of features in our prediction models and improve the quality of fit (i.e., it is easier to learn accurate parameters in smaller models since it is harder to overfit in these cases)?

Future Research Directions

While my work has made contributions to large-scale data processing and retrieval, there still remains a vast amount of research to improve data processing, retrieval and recommendation technologies in an online framework. To this end, following are a few key areas where my future research will help close the loop between data and decisions:

1. Combining diverse side information for recommendation and IR: users are interacting on diverse online services such as social network platforms, collaborative tagging sites, microblogging sites, and online game platforms. These services provide valuable information to enhance recommendation and IR. A critical question is how to extend existing solutions into a unified framework that considers multiple and diverse data sources? I have attempted to answer this question for recommendation using a co-matrix factorization approach of multiple data sources [13].
2. Real-time analysis of large-scale data: considering stream data analysis is important for most existing data mining techniques. But how to do it? Although it’s a significant challenge in data mining, different options are possible. One direction consists of exploring the possibility of updating models without their complete re-computation (incremental learning).
3. Large-scale data visualization: many applications such as real-time monitoring of social network activities or urban traffic congestion reports involve high information and cognitive load tasks. In this context, it is common to have large-scale information visualization interfaces to concurrently display system status, alerts, and events. However, displaying all information elements simultaneously would result in a saturated and unreadable display. Thus, the development of novel adaptive user interfaces (AUIs) are required to help focus the user’s attention by filtering information relevant to their current task. How to optimally set the parameters of filters is a problem that I intend to formalize as a Mixed-integer Linear Programming for which we have optimal solvers.

Overall, my vision for the future is simple: social media is connecting everything, Web services are becoming more interactive and personalized, and every embedded system that impacts our lives — from our daily commute to our consumer needs — will become increasingly adaptive. This means that we will have to transform immense quantities of data into valuable knowledge that should be used for decision-making. This is what I aim to do in my research.