Venkateswara Rao Kagita

Virtual user approach for group recommender systems using precedence relations

We describe virtual user strategy and then examine its properties.The experimental results show that virtual strategy achieve better Precision and Recall.We propose incremental algorithms to update virtual user profile.A new measure called monotonicity is introduced to judge the efficiency of a recommender system.Virtual-user profile + Monotonicity yield recommendations having higher accuracy on benchmark datasets. In this paper, we propose a novel virtual user strategy using precedence relations and develop a new scheme for group recommender systems. User profiles are provided in terms of the precedence relations of items as used by group members. A virtual user for a group is constructed by taking transitive precedence of items of all members into consideration. The profile of the virtual user represents the combined profile of the group. There has not been any earlier attempt to define virtual user profile using precedence relations. We show that the proposed framework exhibits many interesting properties. Earlier approaches construct virtual user profile by considering the set of common items used by all members of the group. In the present work, we propose a method of computing weightage for each item, not necessarily common to all members, using transitive precedence. We also introduce a new measure called monotonicity to measure the performance of any recommender system. In a top-k recommendation, monotonicity tries to measure the number of items continued to be recommended when a technique is utilized incrementally. We experimented extensively for different combinations of parameter settings and for different group sizes on MovieLens data. We show that our framework has better performance in terms of precision and recall when compared with other methods. We show that our recommendation framework exhibits robust monotonicity.

Efficient computation for probabilistic skyline over uncertain preferences

We propose an efficient algorithm that can compute skyline probability exactly for reasonably large database.We introduce the concept of zero-contributing set which has zero effect in the signed aggregate of joint probabilities.We propose an incremental algorithm to compute skyline probability in dynamic scenarios wherein objects are added incrementally.The theoretical concepts developed helps to devise an efficient technique to compute skyline probability of all objects in the database.Detailed experimental analysis for real and synthetic datasets are reported to corroborate our findings. Efficient computation of skyline probability over uncertain preferences has not received much attention in the database community as compared to skyline probability computation over uncertain data. All known algorithms for probabilistic skyline computation over uncertain preferences attempt to find inexact value of skyline probability by resorting to sampling or to approximation scheme. Exact computation of skyline probability for database with uncertain preferences of moderate size is not possible with any of the existing algorithms. In this paper, we propose an efficient algorithm that can compute skyline probability exactly for reasonably large database. The inclusion-exclusion principle is used to express skyline probability in terms of joint probabilities of all possible combination. In this regard we introduce the concept of zero-contributing set which has zero effect in the signed aggregate of joint probabilities. Our algorithm employs a prefix-based k-level absorption to identify zero-contributing sets. It is shown empirically that only a very small portion of exponential search space remains after level wise application of prefix-based absorption. Thus it becomes possible to compute skyline probability with respect to large datasets. Detailed experimental analysis for real and synthetic datasets are reported to corroborate this claim. We also propose an incremental algorithm to compute skyline probability in dynamic scenarios wherein objects are added incrementally. Moreover, the theoretical concepts developed in this paper help to devise an efficient technique to compute skyline probability of all objects in the database. We show that the exponential search space is pruned once and then for each individual object skyline probability can be derived by inspecting a portion of the pruned lattice. We also use a concept of revival of absorbed pairs. We believe that this process is more efficient than computing the skyline probability individually.

Collaborative filtering using multiple binary maximum margin matrix factorizations

In MMMF, ratings matrix with multiple discrete values is treated by specially extending hinge loss function to suit multiple levels.We view this process as analogous to extending two-class classifier to a unified multi-class classifier.Alternatively, multi-class classifier can be built by arranging multiple two- class classifiers in a hierarchical manner.In this paper, we investigate this aspect for collaborative filtering and propose an efficient and novel framework of multiple bi-level MMMFs.We compare our method with nine well-known algorithms on two benchmark datasets and show that our method outperforms these methods on NMAE measure. Maximum Margin Matrix Factorization (MMMF) has been a successful learning method in collaborative filtering research. For a partially observed ordinal rating matrix, the focus is on determining low-norm latent factor matrices U (of users) and V (of items) so as to simultaneously approximate the observed entries under some loss measure and predict the unobserved entries. When the rating matrix contains only two levels (1), rows of V can be viewed as points in k-dimensional space and rows of U as decision hyperplanes in this space separating +1 entries from 1 entries. When hinge/smooth hinge loss is the loss function, the hyperplanes act as maximum-margin separator. In MMMF, rating matrix with multiple discrete values is treated by specially extending hinge loss function to suit multiple levels. We view this process as analogous to extending two-class classifier to a unified multi-class classifier. Alternatively, multi-class classifier can be built by arranging multiple two-class classifiers in a hierarchical manner. In this paper, we investigate this aspect for collaborative filtering and propose an efficient and novel framework of multiple bi-level MMMFs. There is substantial saving in computational overhead. We compare our method with nine well-known algorithms on two benchmark datasets and show that our method outperforms these methods on NMAE measure. We also show that our method yields latent factors of lower ranks and the trade-off between empirical and generalization error is low.

Collaborative filtering by PSO-based MMMF

Matrix factorization (MF) techniques are one of the most succesful realisations of recommender systems based on collaborative filtering/prediction (CF). For instance, in a movie recommender system based on CF, the inputs to the system are user ratings on movies (items) the users have already seen. To predict user preferences on movies they have not yet watched one needs to understand the patterns in the partially observed rating matrix. It is possible to visualize this setting as a matrix completion problem, i.e., completing entries in a partially observed data matrix. Then the objective is to compute user latent factor and item latent factor such that the rating matrix is completed. The factorization is usually accomplished by minimizing an objective function using gradient descent or its variants such as conjugate gradient or stochastic gradient descent. In this paper we make use of a particular MF technique called Maximum Margin Matrix Factorization (MMMF) and show that it is suitable for multi-level discrete rating matrix. The factorization is accomplished by minimizing the hinge loss objective function. We propose to improve the gradient search by combining a component of particle Swarm Optimisation (PSO) search. Though earlier attempts of improving PSO search by adding gradient information exist, the main objective of the present work is to improvise gradient/stochastic-gradient search. Our proposed algorithm finds better minimizing points early (fewer number of iterations) not only for the loss function but also for other performance metrics of collaborative filtering such as RMSE and MAE. There has not been any earlier attempt to combine particle swarm optimisation with maximum margin matrix factorisation for collaborative filtering.

Group Recommender Systems: A Virtual User Approach Based on Precedence Mining

The recommendation framework based on precedence mining as outlined in [3] is limited to personal recommendation and cannot be trivially extended for group recommendation scenario. In this paper, we extend the precedence mining model for group recommendation by proposing a novel way of defining a virtual user by taking transitive precedence relation into account. We obtained experimental results for different combinations of parameter settings and for different group-sizes on MovieLens data-set based on our virtual-user model. We show that our framework has better performance in terms of precision and recall when compared with other methods.

Precedence Mining in Group Recommender Systems

We extend the Precedence mining model for personal recommendation as outlined in Parameswaran et.al., [6] in three different ways. Firstly, we show how precedence mining model can be used for recommending items of interest to a group of users (group recommendation) and compare and contrast our model with traditional group recommendation models like collaborative and Hybrid. Secondly, we extend the precedence mining model to incorporate ratings for items and experimental results show that the goodness of recommendation is improved. The third extension is related to the issue of new items being ignored which is a fundamental problem plaguing collaborative and precedence mining algorithms. When recommendations are based on other users interests (like in Collaborative recommender systems) the possibility of not recommending a new item which has not been consumed by many of the users is high though the new item may be of interest to the target user. We outline two models, Vector precedence-mining and Hybrid precedence-mining that addresses this issue.

Proximal maximum margin matrix factorization for collaborative filtering

We propose an alternative and new MMMF scheme for discrete-valued rating matrix.Our work draws motivation of recent advent of proximal support vector machines.The propose method overcomes the problem of overtting.We validate our hypothesis by conducting experiments on real and synthetic datasets. Maximum Margin Matrix Factorization (MMMF) has been a successful learning method in collaborative filtering research. For a partially observed ordinal rating matrix, the focus is on determining low-norm latent factor matrices U (of users) and V (of items) so as to simultaneously approximate the observed entries under some loss measure and predict the unobserved entries. When the rating matrix contains only two levels (1), rows of V can be viewed as points in k-dimensional space and rows of U as decision hyperplanes in this space separating +1 entries from 1 entries. The concept of optimizing a loss function to determine the separating hyperplane is prevalent in support vector machines (SVM) research and when hinge/smooth hinge loss is used, the hyperplanes act as a maximum-margin separator. In MMMF, a rating matrix with multiple discrete values is treated by specially extending hinge loss function to suit multiple levels. MMMF is an efficient technique for collaborative filtering but it has several shortcomings. A prominent shortcoming is an overfitting problem wherein if learning iteration is prolonged to decrease the training error the generalization error grows. In this paper, we propose an alternative and new maximum margin factorization scheme for discrete-valued rating matrix to overcome the problem of overfitting. Our work draws motivation from a recent work on proximal support vector machines (PSVMs) wherein two parallel hyperplanes are used for binary classification and points are classified by assigning them to the class corresponding to the closest of two parallel hyperplanes. In other words, proximity to decision hyperplane is used as the classifying criterion. We show that a similar concept can be used to factorize the rating matrix if the loss function is suitably defined. The present scheme of matrix factorization has advantages over MMMF (similar to the advantages of PSVM over standard SVM). We validate our hypothesis by carrying out experiments on real and synthetic datasets.

Bi-directional Search for Skyline Probability

Computing the skyline probability of an object for a database, wherein the probability of preferences between pairs of data objects are uncertain, requires computing the probability of union of events from the probabilities of all possible joint probabilities. From the literature it can be seen that for a database of size n it requires computation of 2 n joint probabilities of all possible combinations. All known algorithms for probabilistic skyline computation over uncertain preferences attempt to find inexact value of skyline probability by resorting to sampling or to approximation schemes. In this paper we use a concept called zero-contributing set of a power set lattice to denote portion of the lattice (a sub-lattice) such that the signed aggregate of joint probabilities corresponding to this set is zero. When such sets can be implicitly identified, the corresponding terms can be removed, saving substantial computational efforts. We propose an efficient heuristic that employs a bi-directional search traversing level wise the power set lattice from top and from bottom and prunes the exponential search space based on zero-contribution.

Threshold-based direct computation of skyline objects for database with uncertain preferences

Skyline queries aim at finding a set of skyline objects from the given database. For categorical data, the notion of preferences is used to determine skyline objects. There are many real world applications where the preference can be uncertain. In such contexts, it is relevant to determine the probability that an object is a skyline object in a database with uncertain pairwise preferences. Skyline query is to determine a set of objects having skyline probability greater than a threshold. In this paper, we address this problem. To the best of our knowledge, there has not been any technique which handles this problem directly. There have been proposals to compute skyline probability of individual objects but applying these for skyline query is computationally expensive. In this paper, we propose a holistic algorithm that determines the set of skyline objects for a given threshold and a database of uncertain preferences. We establish the relationship between skyline probability and the probability of the union of events. We guide our search to prune objects which are unlikely to be skyline objects. We report extensive experimental analysis to justify the efficiency of our algorithm.

Multi-label classification using hierarchical embedding

Multi-label learning deals with the classification of data with multiple labels.Output space with many labels is tackle by modeling inter-label correlations.Use of parametrization and embedding have been the prime focus.A piecewise-linear embedding using maximum margin matrix factorization is proposed.Our experimental analysis manifests the superiority of our proposed method. Multi-label learning is concerned with the classification of data with multiple class labels. This is in contrast to the traditional classification problem where every data instance has a single label. Multi-label classification (MLC) is a major research area in the machine learning community and finds application in several domains such as computer vision, data mining and text classification. Due to the exponential size of the output space, exploiting intrinsic information in feature and label spaces has been the major thrust of research in recent years and use of parametrization and embedding have been the prime focus in MLC. Most of the existing methods learn a single linear parametrization using the entire training set and hence, fail to capture nonlinear intrinsic information in feature and label spaces. To overcome this, we propose a piecewise-linear embedding which uses maximum margin matrix factorization to model linear parametrization. We hypothesize that feature vectors which conform to similar embedding are similar in some sense. Combining the above concepts, we propose a novel hierarchical matrix factorization method for multi-label classification. Practical multi-label classification problems such as image annotation, text categorization and sentiment analysis can be directly solved by the proposed method. We compare our method with six well-known algorithms on twelve benchmark datasets. Our experimental analysis manifests the superiority of our proposed method over state-of-art algorithm for multi-label learning.