Page 22 - 2017 Brochure
P. 22
earch Laboratories
Data Management and Information Discovery Lab
Data of various types (e.g., sensor data, transactions, in the form of hyperedges in a hypergraph. We have designed
multimedia, social network, Web browsing log, etc.) are being an approximation algorithm and a new index structure based
generated at an ever increasing rate. As computer hardware on FP-Tree. In another project, we have found that mining online
and networks are abundant and inexpensive, the timing has social behavior provides an opportunity to actively identify
never been better to explore possible means of utilizing such social network mental disorders (SNMDs) at an early stage. We
data to develop new technologies that solve difficult problems propose a machine learning framework, namely, Social Network
or to explore applications that were previously impossible. Mental Disorder Detection (SNMDD) and a new SNMD-based
The Data Management and Information Discovery Group was Tensor Model to accurately identify potential cases of SNMDs.
formed with the main objectives of initiating innovative research Our results show that SNMDD is a promising tool for identifying
and strengthening scientific and technological excellence in: users with potential SNMDs. For group therapy, we have
(1) effective collection, representation, storage, processing, formulated the Member Selection for Online Support Group
and analyzing of massive data; and (2) exploring data mining (MSSG) to maximize the similarity of the symptoms between
and machine learning technologies which may efficiently and all selected members, while ensuring that any two members
effectively uncover valuable knowledge within various types are unacquainted to each other. We proved that MSSG is NP-
of data. Currently, research from this group focuses on the Hard and inapproximate within any ratio. Also, we designed a
following topics: (1) Social Influences and Query Processing 3-approximation algorithm with a guaranteed error bound.
for Social Networks; (2) Modeling and Prediction of Real-Time
Bidding in Online Display Advertising; (3) Deep Learning for 2. Modeling and Prediction for Real-Time Bidding
Urban Air Pollution Prediction; (4) Indexing, Data Mining and on Online Display Advertising
Management for Non-Volatile Main Memory. The followings are
the introduction of the problems we have been working on. Online display advertising is now programmatic. The ad
impression displayed for each website visitor may be different
1. Social Influences and Query Processing for and is dynamically decided by a mechanism called Real-Time
Social Networks Bidding (RTB), which brokers interactions between publishers
and advertisers. In the RTB environment, advertisers rely heavily
Our research on social networks focuses on the on having a good prediction model for ad click-through-rates
development of efficient social group queries and social to effectively and efficiently target potential customers and offer
influence. For location-based social networks, we formulate a reasonable bids. We aim to design an appropriate learning
a Social-Spatial Group Query with a new index structure method from historical bidding data with incomplete labels
that can efficiently find spatially close-by friends with tight so that advertisers can have an effective model to accurately
social relations and corresponding rally points at their predict the ad click-through rates and find a good bidding
available time slots. In addition to the spatial and temporal strategy under budgetary constraints.
dimensions, we explore group search that addresses
user preference. The proposed randomized algorithm 3. Deep Learning for Urban Air Pollution Prediction
has a guaranteed performance bound, which can be
exploited in Online-To-Offline social applications and group Air pollution prediction and source identification have become
coupon services. For professional social networks, we critical issues in environmental protection, human health,
propose a new socio-spatial query to form a strong task economics and even politics. In the past three years, there
group that addresses the group distance, professional have been many studies trying to predict fine grained air
skills of group members, and social rapport in the group. pollution using various techniques, including neural networks,
We prove that the problem is inapproximable within any
ratio, but an approximation algorithm exists when a small
error bound is permitted. Recently, we considered social
presence theory in social psychology to formulate a new
problem for organizing friend-making activities via online
social networks. For this we designed a 3-approximation
algorithm.
In addition to social queries, we explore influence diffusion
on a specific target. As such, we find the optimal group of
intermediate nodes that may change a target´s decision.
We also deploy this distributed computation in mobile social
networks to find optimal solutions. Moreover, we incorporate
frequent pattern mining and product bundles for viral marketing
of multiple correlated products. We present a new Social Item
Graph that captures both social influence and frequent patterns
20 研究群 Research Laboratories
Data Management and Information Discovery Lab
Data of various types (e.g., sensor data, transactions, in the form of hyperedges in a hypergraph. We have designed
multimedia, social network, Web browsing log, etc.) are being an approximation algorithm and a new index structure based
generated at an ever increasing rate. As computer hardware on FP-Tree. In another project, we have found that mining online
and networks are abundant and inexpensive, the timing has social behavior provides an opportunity to actively identify
never been better to explore possible means of utilizing such social network mental disorders (SNMDs) at an early stage. We
data to develop new technologies that solve difficult problems propose a machine learning framework, namely, Social Network
or to explore applications that were previously impossible. Mental Disorder Detection (SNMDD) and a new SNMD-based
The Data Management and Information Discovery Group was Tensor Model to accurately identify potential cases of SNMDs.
formed with the main objectives of initiating innovative research Our results show that SNMDD is a promising tool for identifying
and strengthening scientific and technological excellence in: users with potential SNMDs. For group therapy, we have
(1) effective collection, representation, storage, processing, formulated the Member Selection for Online Support Group
and analyzing of massive data; and (2) exploring data mining (MSSG) to maximize the similarity of the symptoms between
and machine learning technologies which may efficiently and all selected members, while ensuring that any two members
effectively uncover valuable knowledge within various types are unacquainted to each other. We proved that MSSG is NP-
of data. Currently, research from this group focuses on the Hard and inapproximate within any ratio. Also, we designed a
following topics: (1) Social Influences and Query Processing 3-approximation algorithm with a guaranteed error bound.
for Social Networks; (2) Modeling and Prediction of Real-Time
Bidding in Online Display Advertising; (3) Deep Learning for 2. Modeling and Prediction for Real-Time Bidding
Urban Air Pollution Prediction; (4) Indexing, Data Mining and on Online Display Advertising
Management for Non-Volatile Main Memory. The followings are
the introduction of the problems we have been working on. Online display advertising is now programmatic. The ad
impression displayed for each website visitor may be different
1. Social Influences and Query Processing for and is dynamically decided by a mechanism called Real-Time
Social Networks Bidding (RTB), which brokers interactions between publishers
and advertisers. In the RTB environment, advertisers rely heavily
Our research on social networks focuses on the on having a good prediction model for ad click-through-rates
development of efficient social group queries and social to effectively and efficiently target potential customers and offer
influence. For location-based social networks, we formulate a reasonable bids. We aim to design an appropriate learning
a Social-Spatial Group Query with a new index structure method from historical bidding data with incomplete labels
that can efficiently find spatially close-by friends with tight so that advertisers can have an effective model to accurately
social relations and corresponding rally points at their predict the ad click-through rates and find a good bidding
available time slots. In addition to the spatial and temporal strategy under budgetary constraints.
dimensions, we explore group search that addresses
user preference. The proposed randomized algorithm 3. Deep Learning for Urban Air Pollution Prediction
has a guaranteed performance bound, which can be
exploited in Online-To-Offline social applications and group Air pollution prediction and source identification have become
coupon services. For professional social networks, we critical issues in environmental protection, human health,
propose a new socio-spatial query to form a strong task economics and even politics. In the past three years, there
group that addresses the group distance, professional have been many studies trying to predict fine grained air
skills of group members, and social rapport in the group. pollution using various techniques, including neural networks,
We prove that the problem is inapproximable within any
ratio, but an approximation algorithm exists when a small
error bound is permitted. Recently, we considered social
presence theory in social psychology to formulate a new
problem for organizing friend-making activities via online
social networks. For this we designed a 3-approximation
algorithm.
In addition to social queries, we explore influence diffusion
on a specific target. As such, we find the optimal group of
intermediate nodes that may change a target´s decision.
We also deploy this distributed computation in mobile social
networks to find optimal solutions. Moreover, we incorporate
frequent pattern mining and product bundles for viral marketing
of multiple correlated products. We present a new Social Item
Graph that captures both social influence and frequent patterns
20 研究群 Research Laboratories
