MLDAS banner

2018 Symposium Invited Speakers' Abstracts

Temporal Regularized Tensor Factorization for Monitoring and Forecasting of Traffic Congestion
Abdelkader Baggag - Qatar Computing Research Institute (QCRI) - HBKU

Abstract: In this paper, we investigate the problem of missing data in the context of real-time monitoring and forecasting of traffic congestion for road networks. We assume that the city has deployed sensors for speed reading in a subset of edges. Our objective is to infer speed readings for the remaining edges in the network as well as missing values. We propose a tensor representation for the series of road network snapshots, and develop a regularized factorization method to estimate the missing values, while learning the latent factors of the network. The regularizer, which incorporates spatial properties of the road network, improves the quality of the results. The learned factors are used in an autoregressive algorithm to predict the future state of the road network with a long horizon. Extensive numerical experiments with real traffic data from the cities of Doha and Aarhus demonstrate that the proposed approach is appropriate for imputing missing data and predicting traffic states.

Next utterance ranking based on context response similarity
Basma El Amel Boussaha - University of Nantes

Abstract: Building dialogue systems that converse with humans in order to help them in their daily tasks is being a priority. Some systems converse by generating dialogues word by word whereas others retrieve the best utterance among a set of candidate responses. These retrieval systems rank the candidate responses by their relevance to the history of the conversation (context), the best response is then chosen. Approaches based on deep neural networks performed well on this task. In this work, we improve a state of the art approach based on an LSTM dual encoder and propose a new response retrieval dialogue system. Based on syntactic and semantic similarities between the context and the response extracted from word embeddings, our approach learns to match the context with the best response. Experimental results on the Ubuntu Dialogue Corpus show an important improvement of about 7%, 6% and 2% on Recall@(1, 2 and 5) compared to the best state of the art system.

Research Roadmap for Automatic Persona Generation: Principles and Open Questions
Joni Salminen - Qatar Computing Research Institute (QCRI) - HBKU

Abstract: As the quantity of online analytics data has dramatically increased, computational techniques are deployed to make sense of this data. In this perspective manuscript, we propose employing personas as a form of making large amounts of customer analytics information useful to decision makers in software development, business, and other domains where understanding customer behavior is important. Toward this end, we develop a system capable of handling hundreds of millions of customer interactions from tens of thousands of pieces of online content. Our approach identifies customer segments by their online behavior, associates the segments with demographic data, and creates rich persona profiles by dynamically adding characteristics, such as name, photo, and descriptive quotes. This manuscript characterizes the open research questions in automatic persona generation, outlining a research agenda that aims at making data analytics more useful for human decision makers.

Capsule-Net for Urdu Digits Recognition
Talha Iqbal - COMSATS Institute of Information Technology

Abstract: A capsule is formed when a group of additional neurons is added to existing convolutional layer in a typical convolutional neural network (CNN). Capsules have activity vector that represents instantiation parameters of an object or part of an object. Capsule network has recently been introduced by Hinton to overcome the shortcomings of typical CNN model trained with back-propagation. In this work, we investigate the use of capsule networks for recognition of handwritten digits of Urdu. Our results show that a multi-layer capsule network achieves better results (98.5% accuracy) than deep auto-encoder (97.3% accuracy), especially when we have digits that are highly overlapped.