Recommender systems have become one of the main components of web technologies that help people to cope with information overload. Based on the analysis of past user behavior, these systems filter items according to users’ likes and interests. Two of the most important metrics used to analyze the performance of these systems are the accuracy and diversity of the recommendation lists. Whereas all the efforts exerted in the prediction of the user interests aim at maximizing the former, the latter emerges in various forms, such as diversity in the lists across all user recommendation lists, referred to as aggregate diversity, and diversity in the lists of individuals, known as individual diversity. In this paper, we tackle the combination of these three objectives and justify this approach by showing through experiments that handling these objectives in pairs does not yield satisfactory results in the third one. To that end, we develop a mathematical model that is formulated using multiobjective optimization approaches. To cope with the intractability of this nonlinear integer programming model, its special structure is exploited by a decomposition technique. For the solution of the resulting formulation, we propose an iterative framework that is composed of a clique-generating genetic algorithm, a constructive heuristic, and an improvement heuristic. The former is designed to incorporate all objective functions into the generated cliques and specifically impose a certain level of individual diversity, whereas the latter chooses one clique for each user such that the desired aggregate diversity level is fulfilled. We conduct experiments on three data sets and show that the proposed modeling approach successfully handles all objectives according to the needs of the system and that the proposed methodology is capable of yielding good upper bounds.

Matrix factorization models often reveal the low-dimensional latent structure in high-dimensional spaces while bringing space efficiency to large-scale collaborative filtering problems. Improving training and prediction time efficiencies of these models are also important since an accurate model may raise practical concerns if it is slow to capture the changing dynamics of the system. For the training task, powerful improvements have been proposed especially using SGD, ALS, and their parallel versions. In this paper, we focus on the prediction task and combine matrix factorization with approximate nearest neighbor search methods to improve the efficiency of top-N prediction queries. Our efforts result in a meta-algorithm, MMFNN, which can employ various common matrix factorization models, drastically improve their prediction efficiency, and still perform comparably to standard prediction approaches or sometimes even better in terms of predictive power. Using various batch, online, and incremental matrix factorization models, we present detailed empirical analysis results on many large implicit feedback datasets from different application domains.

Summary Pairwise learning to rank is known to be suitable for a wide range of collaborative filtering applications. In this work, we show that its efficiency can be greatly improved with parallel stochastic gradient descent schemes. Accordingly, we first propose to extrapolate two such state‐of‐the‐art schemes to the pairwise learning to rank problem setting. We then show the versatility of these proposals by showing the applicability of several important extensions commonly desired in practice. Theoretical as well as extensive empirical analyses of our proposals show remarkable efficiency results for pairwise learning to rank in offline and stream learning settings.

Learning to rank with pairwise loss functions has been found useful in collaborative filtering recommender systems. At web scale, the optimization is often based on matrix factorization with stochastic gradient descent (SGD) which has a sequential nature. We investigate two different shared memory lock-free parallel SGD schemes based on block partitioning and no partitioning for use with pairwise loss functions. To speed up convergence to a solution, we extrapolate simple practical algorithms from their application to pointwise learning to rank. Experimental results show that the proposed algorithms are quite useful regarding their ranking ability and speedup patterns in comparison to their sequential counterpart.

The success of a recommender system is generally evaluated with respect to the accuracy of recommendations. However, recently diversity of recommendations has also become an important aspect in evaluating recommender systems. One dimension of diversity is called aggregate diversity, which refers to the diversity of items in the recommendation lists of all users and can be defined with different metrics. The maximization of both accuracy and the aggregate diversity simultaneously renders a multiobjective optimization problem that can be handled by different approaches. In this paper, after providing a thorough analysis of the multiobjective optimization approaches for this problem, we propose a new model that takes into account both accuracy and aggregate diversity. Different from previous works, our model is specifically designed to incorporate distributional diversity metrics, which measure how evenly the items are distributed in the recommendation lists of users. To solve the large-scale instances, we propose a column generation algorithm and a Lagrangian relaxation approach based on the decomposition of the model. We present the results of the mathematical models and the performance of the proposed methodology that are obtained by computational experiments on real-world data sets. These results reveal that our model successfully captures the trade-off between the objectives and reaches very high levels of distributional diversity.

Large scale IP networks contain thousands of network devices such as routers and switches. Massive amounts of logging data is generated by these devices. Analysing this data is both a challenge and an opportunity for finding network problems. Moreover, large IP networks contain devices from different vendors, so it is important to build a system which can work with network devices of different brands. In this study we describe a distributed architecture which can retrieve, store, and process massive amounts of network logging data in real time. Using this architecture we also build a basic anomaly detection system. The system statistically models cumulative counts of logs for different event types for all the devices in the network. The statistical approach lets the system to detect deviations from the normal behaviour without consulting expert knowledge. Our evaluations show that the system effectively handles massive amounts of data and detects anomalies.

Location and time are two critical aspects of most security-related events, and thus, spatiotemporal data analysis plays a central role in many security-related applications. The human brain has great capabilities of developing invariant representations of objects by taking advantage of both spatial similarity of features of objects/events and their relative timings (temporal information). Trace learning rule is one well-known solution for this problem of combining temporal relations with spatial proximity in clustering tasks such as the one performed by self organizing maps. In this work, we investigate a two stage mechanism: i) finding local clusters using spatial proximity, ii) grouping these clusters as suggested by temporal continuity patterns. We show our experimental results on a movie created from face images.