Bharat Bhasker

A web recommendation system considering sequential information

Abstract With the rapid growth of information technology, the current era is witnessing an exponential increase in the generation and collection of web data. Projecting the right information to the right person is becoming more difficult day by day, which in turn adds complexity to the decision making process. Recommendation systems are intelligent systems that address this issue. They are widely used in e-commerce websites to recommend products to users. Most of the popular recommendation systems consider only the content information of users and ignore sequential information. Sequential information also provides useful insights about the behavior of users. We have developed a novel system that considers sequential information present in web navigation patterns, along with content information. We also consider soft clusters during clustering, which helps in capturing the multiple interests of users. The proposed system has utilized similarity upper approximation and singular value decomposition (SVD) for the generation of recommendations for users. We tested our approach on three datasets, the MSNBC benchmark dataset, simulated dataset and CTI dataset. We compared our approach with the first order Markov model as well as random prediction model. The results validate the viability of our approach.

Personalised recommendations in e-commerce

Most of the current personalised recommender systems use either collaborative filtering or data mining for offering recommendations. However, such methods are beset with problems of sparsity and scalability. In this paper, we present a System for Personalised REcommendations in E-commerce (SPREE) that combines the strengths of both collaborative filtering and data mining for providing better recommendations. We experimentally evaluate our system and show the benefits using a set of real and synthetic datasets. We also propose a novel similarity metric for efficiently computing collaborative users. Experimental results show that the proposed similarity metric is up to 12 orders of magnitude faster and has better predictive capabilities compared to other similarity metrics.

HeteClass: A Meta-path based framework for transductive classification of objects in heterogeneous information networks

Abstract Transductive classification using labeled and unlabeled objects in a heterogeneous information network for knowledge extraction is an interesting and challenging problem. Most of the real-world networks are heterogeneous in their natural setting and traditional methods of classification for homogeneous networks are not suitable for heterogeneous networks. In a heterogeneous network, various meta-paths connecting objects of the target type, on which classification is to be performed, make the classification task more challenging. The semantic of each meta-path would lead to the different accuracy of classification. Therefore, weight learning of meta-paths is required to leverage their semantics simultaneously by a weighted combination. In this work, we propose a novel meta-path based framework, HeteClass, for transductive classification of target type objects. HeteClass explores the network schema of the given network and can also incorporate the knowledge of the domain expert to generate a set of meta-paths. The regularization based weight learning method proposed in HeteClass is effective to compute the weights of symmetric as well as asymmetric meta-paths in the network, and the weights generated are consistent with the real-world understanding. Using the learned weights, a homogeneous information network is formed on target type objects by the weighted combination, and transductive classification is performed. The proposed framework HeteClass is flexible to utilize any suitable classification algorithm for transductive classification and can be applied on heterogeneous information networks with arbitrary network schema. Experimental results show the effectiveness of the HeteClass for classification of unlabeled objects in heterogeneous information networks using real-world data sets.

Quickly locating efficient, equitable deals in automated negotiations under two-sided information uncertainty

This paper develops an automated negotiation procedure inclusive of mechanism design and agent design for bilateral multi-issue negotiations under two-sided information uncertainty. The proposed negotiation mechanism comprises a protocol called MUP (Monotonic Utility-granting Protocol) and a matching strategy called WYDIWYG (What You Display Influences What You Get). The proposed preference elicitation procedure makes the agents faithful surrogates of the user they represent while the proposed Frontier Tracking Proposal Construction Algorithm (FTPCA) makes them learn the opponents flexibility in negotiation and respond appropriately. The mechanism design and the agent design together help in locating efficient and equitable deals quickly. The efficiency, stability, simplicity, distribution symmetry and incentive compatibility of the proposed procedure are demonstrated through negotiation simulation experiments.

An alternative approach for clustering web user sessions considering sequential information

Clustering is a prominent technique in data mining applications. It generates groups of data points that are similar to each other in a given aspect. Each group has some inherent latent similarity which is computed using the similarity measures. Clustering web users based on navigational pattern has always been an interesting as well as a challenging task. A web user, based on its navigational pattern, may belong to multiple categories. Intrinsically, web user navigation pattern exhibits sequential property. When dealing with sequence data, a similarity measure should be chosen, which captures both the order as well as content information during computation of similarity among sequences. In this paper, we have utilized the Sequence and Set Similarity Measure S^{3}M with rough set based similarity upper approximation clustering algorithm to group web users based on their navigational patterns. The quality of cluster formed using rough set based clustering algorithm with S^{3}M measure has been compared with the well known clustering algorithm, Density based spatial clustering of applications with noise DBSCAN. The experimental results show the viability of our approach.

An upper approximation based community detection algorithm for complex networks

The emergence of multifarious complex networks has attracted researchers and practitioners from various disciplines. Discovering cohesive subgroups or communities in complex networks is essential to understand the dynamics of real-world systems. Researchers have made persistent efforts to investigate and infer community patterns in complex networks. However, real-world networks exhibit various characteristics wherein existing communities are not only disjoint but are also overlapping and nested. The existing literature on community detection consists of limited methods to discover co-occurring disjoint, overlapping and nested communities.In this work, we propose a novel rough set based algorithm capable of uncovering true community structure in networks, be it disjoint, overlapping or nested. Initial sets of granules are constructed using neighborhood connectivity around the nodes and represented as rough sets. Subsequently, we iteratively obtain the constrained connectedness upper approximation of these sets. To constrain the sets and merge them during each iteration, we utilize the concept of relative connectedness among the nodes. We illustrate the proposed algorithm on a toy network and evaluate it on fourteen real-world benchmark networks. Experimental results show that the proposed algorithm reveals more accurate communities and significantly outperforms state-of-the-art techniques. A rough set based community detection algorithm for complex networks has been proposed.Experiments have been performed on fourteen benchmark networks from diverse domains.Comparative analysis of the proposed algorithm has been performed with the relevant state-of-the-art methods.The performance of proposed algorithm is superior to state-of-the-art methods.

Product recommendations for cross-selling in electronic business

A system applicable in electronic commerce environments that combines the strengths of both collaborative filtering and data mining for providing better recommendations is presented. It captures the item-to-item relationship through association rule mining and then uses purchase behaviour of collaborative users for generating the recommendations. It was implemented and evaluated on a set of real datasets. Our methodology results in improved quality of recommendations measured in terms of recall and coverage metrics.

Collaborative Filtering Based on the Entropy Measure

This paper introduces a new memory based approach to ratings based collaborative filtering. Unlike existing memory based collaborative filtering approaches, this approach exploits the predictable portions of even some complex relationships between users while selecting the mentors for an active user through the use of the novel notion of selective predictability, which can be measured using the Entropy measure. The proposed approach has been tested using the MovieLens dataset, and it is expected that this approach should work equally well for any given dataset. This flexibility would make it possible to make use of this approach in a wide variety of application domains including e-commerce where recommendations need to be provided to users based on the ratings provided implicitly or explicitly by different users to different items in the past. However the items should represent a relatively homogeneous group like movies, music albums, compact, disks, books, software, research articles etc.

MaxDomino: efficiently mining maximal sets

We present MaxDomino, an algorithm for mining maximal frequent sets using a novel concept of dominancy factor of a transaction. We also propose a hashing scheme to collapse the database to a form that contains only unique transactions. Unlike traditional bottom up approach with look-aheads, MaxDomino employs a top down strategy with selective bottom up search for mining maximal sets. Using the connect dataset [Benchmark dataset created by University California, Irvine], our experimental results reveal that MaxDomino outperforms GenMax at higher support levels. Furthermore, our scalability tests show that MaxDomino yields an order of magnitude improvement in speed over GenMax. MaxDomino is especially efficient when the maximal frequent sets are longer.

Privacy preserving graph publishing using fuzzy set

Privacy preserving graph publishing is gaining importance in recent times mainly because of inherent privacy issues existing in publishing graph and social network data. Therefore, graph structure needs to be anonymized before publishing. This study proposes anonymization of a graph using fuzzy sets to preserve the graphs privacy while maintaining the utility that can be derived from the graph. We have conducted the experiments on four different datasets, and the results suggest that the proposed approach would not only help in protecting the privacy of data but also in maintaining the quality of data for analysis. To check the robustness of the proposed approach, we have validated the effectiveness of the approach on five key community detection algorithms on three performance measures.