Gravity R&D has been providing recommendation engines as SaaS solutions since 2009. The company has a strong research focus and recommendation quality has always been their primary differentiating factor. Widely used or open source recommendation algorithms are of little use to our technology team as a result of the superiority of our in-house developed, proprietary algorithms. Gravity R&D experienced many challenges while scaling up their services. The sheer quantity of data handled on a daily basis increased exponentially. This presentation will cover how overcoming these challenges permanently shaped our algorithms and system architecture used to generate these recommendations. Serving personalized recommendations requires real-time computation and data access for every single request. To generate responses in real-time, current user inputs have to be compared against their history in order to deliver accurate recommendations. We then combine this user information with specific details about available items as the next step in the recommendation process. It becomes more difficult to provide accurate recommendations as the number of transactions and items increase. It also becomes difficult because this type of analysis requires the combination of multiple heterogeneous algorithms that all require different inputs. Initially, the architecture was designed for MF based models and serving huge numbers of requests but with a limited number of items. Now, Gravity is using MF, neighborhood based models and metadata based models to generate recommendations for millions of items within their databases. This required a shift from a monolithic architecture with in-process caching to a more service oriented architecture with multi-layer caching. As a result of an increase in the number of components and number of clients, managing the infrastructure can be quite difficult. Even with these challenges, we don't believe that it is worthwhile to use a fully distributed system. It adds unneeded complexity, resources, and overhead to the system. We prefer an approach of firstly optimizing current algorithms and architecture and only moving to a distributed system when no other options are left.

In this paper we will describe a modification of the matrix factorization (MF) algorithm which allows visualizing the user and item characteristics. When applying MF for collaborative filtering, we get a model that represents the attributes of users and items by feature vectors. Some elements of these vectors may have understandable meaning for humans but due to the lack of internal connections between the feature vectors, these are difficult to visualize. In this paper we give a detailed description of a MF method enabling better visualization of features by arranging them into a 2D map, where via the calculation of the feature values we try to position features with similar “meaning” close to each other. To achieve this first we define a neighborhood relation on features, then we modify the MF so that we introduce a new term in the error function which penalize the difference between the neighbor features. We show that this modification slightly decrease the accuracy of the model but we get well visualized feature maps. On the feature maps meanings can be associated with regions, and so we can provide an interesting explanation for the user why he/she was recommended the movie. Such plausible explanations may result in that users will better understand how the system works, which can also increase customer loyalty towards the service provider.

Various morphological and acoustic modeling techniques are evaluated on a less resourced, spontaneous Hungarian large-vocabulary continuous speech recognition (LVCSR) task. Among morphologically rich languages, Hungarian is known for its agglutinative, inflective nature that increases the data sparseness caused by a relatively small training database. Although Hungarian spelling is considered as simple phonological, a large part of the corpus is covered by words pronounced in multiple, phonemically different ways. Data-driven and language specific knowledge supported vocabulary decomposition methods are investigated in combination with phoneme- and grapheme-based acoustic modeling techniques on the given task. Word baseline and morph-based advanced baseline results are significantly outperformed by using both statistical and grammatical vocabulary decomposition methods. Although the discussed morph-based techniques recognize a significant amount of out of vocabulary words, the improvements are due not to this fact but to the reduction of insertion errors. Applying grapheme-based acoustic models instead of phoneme-based models causes no severe recognition performance deteriorations. Moreover, a fully data-driven acoustic modeling technique along with a statistical morphological modeling approach provides the best performance on the most difficult test set. The overall best speech recognition performance is obtained by using a novel word to morph decomposition technique that combines grammatical and unsupervised statistical segmentation algorithms. The improvement achieved by the proposed technique is stable across acoustic modeling approaches and larger with speaker adaptation.

There are many studies focusing on enhancing physiological data in user interfaces. On one hand biofeedback games are using skin conductance and heart rate data to reflect the emotional state of the user, on the other hand BCI research tries to conclude user intentions from EEG signals.

Collaborative filtering (CF) approaches proved to be effective for recommender systems in predicting user preferences in item selection using known user ratings of items. This subfield of machine learning has gained a lot of popularity with the Netflix Prize competition started in October 2006. Two major approaches for this problem are matrix factorization (MF) and the neighbor based approach (NB). In this work, we propose various variants of MF and NB that can boost the performance of the usual ensemble based scheme. First, we investigate various regularization scenarios for MF. Second, we introduce two NB methods: one is based on correlation coefficients and the other on linear least squares. At the experimentation part, we show that the proposed approaches compare favorably with existing ones in terms of prediction accuracy and/or required training time. We present results of blending the proposed methods.