International Journal of Semantic Computing

At the onset of any campaign in an online media, an advertiser usually provides a specific demographic “target” to a content publisher. But limited individual profile information leads to low accuracy in targeting. A third party (like Nielsen, Comscore) validates the percentage success of “in-target” impressions. Low accuracy in targeting is expensive as the publisher gets paid only for the impressions which were on target. However, publishers could get access to the demographic mix information for each show from the third party. In this work, we propose a new approach to incorporate user level latent features developed from the show-wise demographic mix provided by the third party, along with session level features of viewers to improve the demographic predictions. In congruence with current industry practice, we train our model on a small labeled data and establish the effectiveness of our approach over existing approaches through experiments.

Especially software running on mobile devices does increasingly rely on contextual information such as time and location. And whenever a software product is affected by context, this context has to be replicated for testing and debugging. This paper introduces an external context manipulation interface for a previously developed learning item scheduler. The scheduler determines when to present a learning item in a learning game based on previous interaction in order to maximize learning efficiency and is based on psychological models. As inter-presentation-intervals can be in the range of days to months, system testing cannot be conducted in a conventional manner. Hence, virtual time hops can be used to fast forward to any specific point in virtual time which would make the software act like it was system time. The approach has shown to be a valuable debugging and testing aid and can be extended for other contextual information sources.

Most multimedia documents available today are agnostic to data semantics. Moreover, their specification language offers little to ease authoring of meaningful content. In this paper, we present the main entities of a new version (3.1) of the Nested Context Model (NCM), which concentrate efforts at integrating support for enriched concept description to the model. These extensions enable the specification of relationships between concept descriptions and multimedia content in the hypermedia way, composing what we call hyperknowledge in this paper. NCM previous version (3.0) is a hypermedia conceptual model. NCL (Nested Context Language), which is part of international standards and ITU recommendations, was engineered according to NCM 3.0 definitions. The extensions discussed in this paper contribute not only for advances in the NCL, but mainly as a conceptual model for hyperknowledge document engineering.

Instant sharing of user-generated video recordings has become a widely used service on platforms such as YouNow, Facebook.Live or uStream. Yet, providing such services with a high QoE for viewers is still challenging, given that mobile upload speed and capacities are limited, and the recording quality on mobile devices greatly depends on the users’ capabilities. One proposed solution to address these issues is video composition. It allows to switch between multiple recorded video streams, selecting the best source at any given time, for composing a live video with a better overall quality for the viewers. Previous approaches have required an in-depth visual analysis of the video streams, which usually limited the scalability of these systems. In contrast, our work allows the stream selection to be realized solely on context information, based on video- and service-quality aspects from sensor and network measurements.

The implemented monitoring service for a context-aware upload of video streams is evaluated in different network conditions, with diverse user behavior, including camera shaking and user mobility. We have evaluated the system’s performance based on two studies. First, in a user study, we show that a higher efficiency for the video upload as well as a better QoE for viewers can be achieved when using our proposed system. Second, by examining the overall delay for the switching between streams based on sensor readings, we show that a composition view change can efficiently be achieved in approximately four seconds.

With the explosion of multimedia data, semantic event detection from videos has become a demanding and challenging topic. In addition, when the data has a skewed data distribution, interesting event detection also needs to address the data imbalance problem. The recent proliferation of deep learning has made it an essential part of many Artificial Intelligence (AI) systems. Till now, various deep learning architectures have been proposed for numerous applications such as Natural Language Processing (NLP) and image processing. Nonetheless, it is still impracticable for a single model to work well for different applications. Hence, in this paper, a new ensemble deep learning framework is proposed which can be utilized in various scenarios and datasets. The proposed framework is able to handle the over-fitting issue as well as the information losses caused by single models. Moreover, it alleviates the imbalanced data problem in real-world multimedia data. The whole framework includes a suite of deep learning feature extractors integrated with an enhanced ensemble algorithm based on the performance metrics for the imbalanced data. The Support Vector Machine (SVM) classifier is utilized as the last layer of each deep learning component and also as the weak learners in the ensemble module. The framework is evaluated on two large-scale and imbalanced video datasets (namely, disaster and TRECVID). The extensive experimental results illustrate the advantage and effectiveness of the proposed framework. It also demonstrates that the proposed framework outperforms several well-known deep learning methods, as well as the conventional features integrated with different classifiers.

Estimating house prices is essential for homeowners and investors alike with both needing to understand the value of their asset, and to understand real estate assets as part of an overall portfolios. Commonly-used indices like the National Association of Realtors (NAR) median home price index, or the celebrated Case-Shiller Home Price Index are reported exclusively over a large geographic areas, i.e., a metropolitan, whereby home price dynamics are lost. In this paper, we propose a improved method to capture price dynamics over time at the most granular level possible — a single home. Using over 16 years of home sale data, from the year 2000 to 2016, we estimate home price index for each house. Once home price dynamics is captured, its possible to aggregate price dynamics to construct a price index over geographies of any kind, e.g., ZIP code. This particular index relies on a so-called ‘gradient boosted’ model, a methodology framework relying on multiple calibration parameters and heavily dependent on sampling techniques. We demonstrate that this approach offers several strengths compared to the commonly reported indices, the ‘median sale’ and ‘repeat sales’ indices.