In this study, we present a meta-learning model to adapt the predictions of the network's capacity between viewers who participate in a live video streaming event. We propose the MELANIE model, where an event is formulated as a Markov Decision Process, performing meta-learning on reinforcement learning tasks. By considering a new event as a task, we design an actor-critic learning scheme to compute the optimal policy on estimating the viewers' high-bandwidth connections. To ensure fast adaptation to new connections or changes among viewers during an event, we implement a prioritized replay memory buffer based on the Kullback-Leibler divergence of the reward/throughput of the viewers' connections. Moreover, we adopt a model-agnostic meta-learning framework to generate a global model from past events. As viewers scarcely participate in several events, the challenge resides on how to account for the low structural similarity of different events. To combat this issue, we design a graph signature buffer to calculate the structural similarities of several streaming events and adjust the training of the global model accordingly. We evaluate the proposed model on the link weight prediction task on three real-world datasets of live video streaming events. Our experiments demonstrate the effectiveness of our proposed model, with an average relative gain of 25% against state-of-the-art strategies. For reproduction purposes, our evaluation datasets and implementation are publicly available at https://github.com/stefanosantaris/melanie

In this paper we present a deep graph reinforcement learning model to predict and improve the user experience during a live video streaming event, orchestrated by an agent/tracker. We first formulate the user experience prediction problem as a classification task, accounting for the fact that most of the viewers at the beginning of an event have poor quality of experience due to low-bandwidth connections and limited interactions with the tracker. In our model we consider different factors that influence the quality of user experience and train the proposed model on diverse state-action transitions when viewers interact with the tracker. In addition, provided that past events have various user experience characteristics we follow a gradient boosting strategy to compute a global model that learns from different events. Our experiments with three real-world datasets of live video streaming events demonstrate the superiority of the proposed model against several baseline strategies. Moreover, as the majority of the viewers at the beginning of an event has poor experience, we show that our model can significantly increase the number of viewers with high quality experience by at least 75% over the first streaming minutes. Our evaluation datasets and implementation are publicly available at https://publicresearch.z13.web.core.windows.net © 2021 IEEE.

Accounting for the fact that users have different sequential patterns, the main drawback of state-of-the-art recommendation strategies is that a fixed sequence length of user-item interactions is required as input to train the models. This might limit the recommendation accuracy, as in practice users follow different trends on the sequential recommendations. Hence, baseline strategies might ignore important sequential interactions or add noise to the models with redundant interactions, depending on the variety of users’ sequential behaviours. To overcome this problem, in this study we propose the SAR model, which not only learns the sequential patterns but also adjusts the sequence length of user-item interactions in a personalized manner. We first design an actor-critic framework, where the RL agent tries to compute the optimal sequence length as an action, given the user’s state representation at a certain time step. In addition, we optimize a joint loss function to align the accuracy of the sequential recommendations with the expected cumulative rewards of the critic network, while at the same time we adapt the sequence length with the actor network in a personalized manner. Our experimental evaluation on four real-world datasets demonstrates the superiority of our proposed model over several baseline approaches. Finally, we make our implementation publicly available at https://github.com/stefanosantaris/sar.

Dynamic graph representation learning strategies are based on different neural architectures to capture the graph evolution over time. However, the underlying neural architectures require a large amount of parameters to train and suffer from high online inference latency, that is several model parameters have to be updated when new data arrive online. In this study we propose Distill2Vec, a knowledge distillation strategy to train a compact model with a low number of trainable parameters, so as to reduce the latency of online inference and maintain the model accuracy high. We design a distillation loss function based on Kullback-Leibler divergence to transfer the acquired knowledge from a teacher model trained on offline data, to a small-size student model for online data. Our experiments with publicly available datasets show the superiority of our proposed model over several state-of-the-art approaches with relative gains up to 5% in the link prediction task. In addition, we demonstrate the effectiveness of our knowledge distillation strategy, in terms of number of required parameters, where Distill2Vec achieves a compression ratio up to 7:100 when compared with baseline approaches. For reproduction purposes, our implementation is publicly available at https://stefanosantaris.github.io/Distill2Vec.

Live video streaming has become a mainstay as a standard communication solution for several enterprises worldwide. To efficiently stream high-quality live video content to a large amount of offices, companies employ distributed video streaming solutions which rely on prior knowledge of the underlying evolving enterprise network. However, such networks are highly complex and dynamic. Hence, to optimally coordinate the live video distribution, the available network capacity between viewers has to be accurately predicted. In this paper we propose a graph representation learning technique on weighted and dynamic graphs to predict the network capacity, that is the weights of connections/links between viewers/nodes. We propose VStreamDRLS, a graph neural network architecture with a self-attention mechanism to capture the evolution of the graph structure of live video streaming events. VStreamDRLS employs the graph convolutional network (GCN) model over the duration of a live video streaming event and introduces a self-attention mechanism to evolve the GCN parameters. In doing so, our model focuses on the GCN weights that are relevant to the evolution of the graph and generate the node representation, accordingly. We evaluate our proposed approach on the link prediction task on two real-world datasets, generated by enterprise live video streaming events. The duration of each event lasted an hour. The experimental results demonstrate the effectiveness of VStreamDRLS when compared with state-of-the-art strategies. Our evaluation datasets and implementation are publicly available at https://github.com/stefanosantaris/vstreamdrls.

In this study, we present a dynamic graph representation learning model on weighted graphs to accurately predict the network capacity of connections between viewers in a live video streaming event. We propose EGAD, a neural network architecture to capture the graph evolution by introducing a self-attention mechanism on the weights between consecutive graph convolutional networks. In addition, we account for the fact that neural architectures require a huge amount of parameters to train, thus increasing the online inference latency and negatively influencing the user experience in a live video streaming event. To address the problem of the high online inference of a vast number of parameters, we propose a knowledge distillation strategy. In particular, we design a distillation loss function, aiming to first pretrain a teacher model on offline data, and then transfer the knowledge from the teacher to a smaller student model with less parameters. We evaluate our proposed model on the link prediction task on three real-world datasets, generated by live video streaming events. The events lasted 80 minutes and each viewer exploited the distribution solution provided by the company Hive Streaming AB. The experiments demonstrate the effectiveness of the proposed model in terms of link prediction accuracy and number of required parameters, when evaluated against state-of-the-art approaches. In addition, we study the distillation performance of the proposed model in terms of compression ratio for different distillation strategies, where we show that the proposed model can achieve a compression ratio up to 15:100, preserving high link prediction accuracy. For reproduction purposes, our evaluation datasets and implementation are publicly available at https://stefanosantaris.github.io/EGAD.

The popularity of location-based social networks (LBSNs) has led to a tremendous amount of user check-in data. Recommending points of interest (POIs) plays a key role in satisfying users needs in LBSNs. While recent work has explored the idea of adopting collaborative ranking (CR) for recommendation, there have been few attempts to incorporate temporal information for POI recommendation using CR. In this article, we propose a two-phase CR algorithm that incorporates the geographical influence of POIs and is regularized based on the variance of POIs popularity and users activities over time. The time-sensitive regularizer penalizes user and POIs that have been more time-sensitive in the past, helping the model to account for their long-term behavioral patterns while learning from user-POI interactions. Moreover, in the first phase, it attempts to rank visited POIs higher than the unvisited ones, and at the same time, apply the geographical influence. In the second phase, our algorithm tries to rank users favorite POIs higher on the recommendation list. Both phases employ a collaborative learning strategy that enables the model to capture complex latent associations from two different perspectives. Experiments on real-world datasets show that our proposed time-sensitive collaborative ranking model beats state-of-the-art POI recommendation methods.

Recent studies have shown that incorporating users' reviews into the collaborative filtering strategy can significantly boost the recommendation accuracy. A pressing challenge resides on learning how reviews influence users' rating behaviors. In this paper, we propose an Adversarial Training approach for Review-based recommendations, namely ATR. We design a neural architecture of sequence-to-sequence learning to calculate the deep representations of users' reviews on items following an adversarial training strategy. At the same time we jointly learn to factorize the rating matrix, by regularizing the deep representations of reviews with the user and item latent features. In doing so, our model captures the non-linear associations among reviews and ratings while producing a review for each user-item pair. Our experiments on publicly available datasets demonstrate the effectiveness of the proposed model, outperforming other state-of-the-art methods.

Exploiting the selections of social friends and foes can efficiently face the data scarcity of user preferences and the cold-start problem. In this paper, we present a Social Deep Pairwise Learning model, namely SDPL. According to the Bayesian Pairwise Ranking criterion, we design a loss function with multiple ranking criteria based on the selections of users, and those in their friends and foes to improve the accuracy in the top-k recommendation task. We capture the nonlinearity in user preferences and the social information of trust and distrust relationships by designing a deep learning architecture. In each backpropagation step, we perform social negative sampling to meet the multiple ranking criteria of our loss function. Our experiments on a benchmark dataset from Epinions, among the largest publicly available that has been reported in the relevant literature, demonstrate the effectiveness of the proposed approach, outperforming other state-of-the art methods. In addition, we show that our deep learning strategy plays an important role in capturing the nonlinear associations between user preferences and the social information of trust and distrust relationships, and demonstrate that our social negative sampling strategy is a key factor in SDPL.

Generating friend recommendations in location-based social networks is a challenging task, as we have to learn how different contextual factors influence users' behavior to form social relationships. For example, the contextual information of users' check-in behavior at common locations and users' activities at close regions may impact users' relationships. In this paper we propose a deep pairwise learning model, namely FDPL. Our model first learns the low dimensional latent embeddings of users' social relationships by jointly factorizing them with the available contextual information based on a multi-view learning strategy. In addition, to account for the fact that the contextual information is non-linearly correlated with users' social relationships we design a deep pairwise learning architecture based on a Bayesian personalized ranking strategy. We learn the non-linear deep representations of the computed low dimensional latent embeddings by formulating the top- k friend recommendation task at location-based social networks as a ranking task in our deep pairwise learning strategy. Our experiments on three real world location-based social networks from Brightkite, Gowalla and Foursquare show that the proposed FDPL model significantly outperforms other state-of-the-art methods. Finally, we evaluate the impact of contextual information on our model and we experimentally show that it is a key factor to boost the friend recommendation accuracy at location-based social networks.

With the advent of social media, generating high quality cross-domain recommendations has become more and more important for users of heterogeneous domains. In this study, we propose a collaborative ranking model to generate cross-domain recommendations. Given a target domain, we design an objective function aimed at performing push of relevant items at the top of a recommendation list. Also, as users may have different behaviours in multiple domains in our collaborative ranking model we propose a weighting strategy to control the influence of user preferences from auxiliary domains when producing the recommendation lists. Our experiments on ten cross-domain recommendation tasks show that the proposed approach achieves higher recommendation accuracy than other state-of-the-art methods.

The sparsity of users' preferences can significantly degrade the quality of recommendations in the collaborative filtering strategy. To account for the fact that the selections of social friends and foes may improve the recommendation accuracy, we propose a learning to rank model that exploits users' trust and distrust relationships. Our learning to rank model focusses on the performance at the top of the list, with the recommended items that end-users will actually see. In our model, we try to push the relevant items of users and their friends at the top of the list, while ranking low those of their foes. Furthermore, we propose a weighting strategy to capture the correlations of users' preferences with friends' trust and foes' distrust degrees in two intermediate trust- and distrust-preference user latent spaces, respectively. Our experiments on the Epinions dataset show that the proposed learning to rank model significantly outperforms other state-of-the-art methods in the presence of sparsity in users' preferences and when a part of trust and distrust relationships is not available. Furthermore, we demonstrate the crucial role of our weighting strategy in our model, to balance well the influences of friends and foes on users' preferences.

In this study, we investigate the problem of network completion by considering the similarities between the node attributes. Given a sample of observed nodes with their incident edges, how can we efficiently reconstruct the network by completing the missing edges of unobserved nodes? Apart from the missing edges, in real settings the node attributes may be partially missing, as well as they may introduce noise when completing the network. We propose a network completion method based on joint clustering and similarity learning. The proposed approach differs from competitive strategies, which consider attribute-based similarities at the node-level. First we generate clusters based on the node attributes, thus reducing the noise and the sparsity in the case that the attributes may be missing. We design a joint objective function to jointly factorize the adjacency matrix of the observed edges with the cluster-based similarities of the node attributes, while at the same time the clusters are adapted, accordingly. In addition, we propose an optimization algorithm to solve the network completion problem via alternating minimization. Our experiments on two real world social networks from Facebook and Google+ show that the proposed approach achieves high completion accuracy, compared to other state-of-the-art methods.

With the advent of learning to rank methods, relevant studies showed that Collaborative Ranking (CR) models can produce accurate ranked lists in the top-N recommendation problem. However, in practice several real-world problems decrease their ranking performance, such as the sparsity and cold-start problems, which often occur in recommendation systems for inactive or new users. In this study, to account for the fact that the selections of social friends can improve the recommendation accuracy, we propose a joint CR model based on the users' social relationships. We propose two different CR strategies based on the notions of Social Reverse Height and Social Height, which consider how well the relevant and irrelevant items of users and their social friends have been ranked at the top of the list, respectively. We focus on the top of the list mainly because users see the top-N recommendations in real-world applications, and not the whole ranked list. Furthermore, we formulate a joint objective function to consider both CR strategies, and propose an alternating minimization algorithm to learn our joint CR model. Our experiments on benchmark datasets show that our proposed joint CR model outperforms other state-of-the-art models that either consider social relationships or focus on the ranking performance at the top of the list.

Cross-modal retrieval has been an emerging topic over the last years, as modern applications have to efficiently search for multimedia documents with different modalities. In this study, we propose a cross-modal hashing method by following a cluster-based joint matrix factorization strategy. Our method first builds clusters for each modality separately and then generates a cross-modal cluster representation for each document. We formulate a joint matrix factorization process with the constraint that pushes the documents' representations of the different modalities and the cross-modal cluster representations into a common consensus matrix. In doing so, we capture the inter-modality, intra-modality and cluster-based similarities in a unified latent space. Finally, we present an efficient way to generate the hash codes using the maximum entropy principle and compute the binary codes for external queries. In our experiments with two publicly available data sets, we show that the proposed method outperforms state-of-the-art hashing methods for different cross-modal retrieval tasks.

Recommendation systems have gained a lot of attention because of their importance for handling the unprecedentedly large amount of available content on the Web, such as movies, music, books, etc. Although Collaborative Ranking (CR) models can produce accurate recommendation lists, in practice several real-world problems decrease their ranking performance, such as the sparsity and cold start problems. Here, to account for the fact that the selections of social friends can leverage the recommendation accuracy, we propose SCR, a Social CR model. Our model learns personalized ranking functions collaboratively, using the notion of Social Reverse Height, that is, considering how well the relevant items of users and their social friends have been ranked at the top of the list. The reason that we focus on the top of the list is that users mainly see the top-N recommendations, and not the whole ranked list. In our experiments with a benchmark data set from Epinions, we show that our SCR model performs better than state-of-the-art CR models that either consider social relationships or focus on the ranking performance at the top of the list.

We propose a media storm indexing algorithm using Map-Reduce in our recently proposed CDVC framework. In this study, CDVC is built on Flink, an open-source platform for stream data processing. The question we answer is how to store massive image collections; for instance, with over one million images per second, as well as with varying incoming rate. In our experiments with two benchmark datasets of 80M and 1B image descriptors, we evaluate the proposed algorithm on different indexing workloads, that is, images that come with high volume and different velocity at the scale of 105-106 images per second. Using a limited set of computational nodes, we show that we achieve a significant speed up factor of nine, on average, compared to conventional indexing techniques, in all settings. Finally, we make our source code publicly available.

We present an image tag completion method, namely PMF-SVN, where the key idea is to exploit images' Semantically and Visually similar Neighborhoods (SVNs) in the learning process of a Probabilistic Matrix Factorization (PMF) framework. We propose a two-step SVN formation algorithm that can generate an image set with the images being both visually and semantically similar. Furthermore, we introduce an efficient way to incorporate the formed SVNs into the learning process of PMF, under the constraint that the latent features of each image are averaged by the features of the images that belong to its SVN. In our experiments with benchmark datasets, we show that the proposed PMF-SVN method outperforms competitive baselines, in terms of completion accuracy, by efficiently capturing the semantical and visual associations between images and tags in SVNs.

In several applications, user preferences can be fairly dynamic, since users tend to exploit a wide range of items and modify their tastes accordingly over time. In this paper, we model continuous user-item interactions over time using a tensor that has time as a dimension (mode). To account for the fact that user preferences are dynamic and change individually, we propose a new measure of user-preference dynamics (UPD) that captures the rate with which the current preferences of each user have been shifted. We generate recommendations based on factorizing the tensor, by weighting the importance of past user preferences according to their UPD values. We additionally exploit users' side data, such as demographics, which can help improving the accuracy of recommendations based on a coupled, tensor-matrix factorization scheme. Our empirical evaluation uses a real data set from last.fm, which allows us to demonstrate that user preferences can become very dynamic. Our experimental results show that the proposed method, by taking into account these dynamics, outperforms several baselines.