Upul Senanayake

The Pagerank-Index: Going beyond Citation Counts in Quantifying Scientific Impact of Researchers

Quantifying and comparing the scientific output of researchers has become critical for governments, funding agencies and universities. Comparison by reputation and direct assessment of contributions to the field is no longer possible, as the number of scientists increases and traditional definitions about scientific fields become blurred. The h-index is often used for comparing scientists, but has several well-documented shortcomings. In this paper, we introduce a new index for measuring and comparing the publication records of scientists: the pagerank-index (symbolised as π). The index uses a version of pagerank algorithm and the citation networks of papers in its computation, and is fundamentally different from the existing variants of h-index because it considers not only the number of citations but also the actual impact of each citation. We adapt two approaches to demonstrate the utility of the new index. Firstly, we use a simulation model of a community of authors, whereby we create various ‘groups’ of authors which are different from each other in inherent publication habits, to show that the pagerank-index is fairer than the existing indices in three distinct scenarios: (i) when authors try to ‘massage’ their index by publishing papers in low-quality outlets primarily to self-cite other papers (ii) when authors collaborate in large groups in order to obtain more authorships (iii) when authors spend most of their time in producing genuine but low quality publications that would massage their index. Secondly, we undertake two real world case studies: (i) the evolving author community of quantum game theory, as defined by Google Scholar (ii) a snapshot of the high energy physics (HEP) theory research community in arXiv. In both case studies, we find that the list of top authors vary very significantly when h-index and pagerank-index are used for comparison. We show that in both cases, authors who have collaborated in large groups and/or published less impactful papers tend to be comparatively favoured by the h-index, whereas the pagerank-index highlights authors who have made a relatively small number of definitive contributions, or written papers which served to highlight the link between diverse disciplines, or typically worked in smaller groups. Thus, we argue that the pagerank-index is an inherently fairer and more nuanced metric to quantify the publication records of scientists compared to existing measures.

Node Assortativity in Complex Networks: An Alternative Approach

Abstract Assortativity quantifies the tendency of nodes being connected to similar nodes in a complex network. Degree Assortativity can be quantified as a Pearson correlation. However, it is insufficient to explain assortative or disassortative tendencies of individual nodes or links, which may be contrary to the overall tendency of the network. A number of ‘local’ assortativity measures have been proposed to address this. In this paper we define and analyse an alternative formulation for node assortativity, primarily for undirected networks. The alternative approach is justified by some inherent shortcomings of existing local measures of assortativity. Using this approach, we show that most real world scale-free networks have disassortative hubs, though we can synthesise model networks which have assortative hubs. Highlighting the relationship between assortativity of the hubs and network robustness, we show that real world networks do display assortative hubs in some instances, particularly when high robustness to targeted attacks is a necessity.

Optimisation of strategy placements for public good in complex networks

Game theory has long been used to model cognitive decision making in societies. While traditional game theoretic modelling has focussed on well-mixed populations, recent research has suggested that the topological structure of social networks play an important part in the dynamic behaviour of social systems. Any agent or person playing a game employs a strategy (pure or mixed) to optimise pay-off. Previous studies have analysed how selfish agents can optimise their payoffs by choosing particular strategies within a social network model. In this paper we ask the question that, if agents were to work towards the common goal of increasing the public good (that is, the total network utility), what strategies they should adapt within the context of a heterogeneous network. We consider a number of classical and recently demonstrated game theoretic strategies, including cooperation, defection, general cooperation, Pavlov, and zero-determinant strategies, and compare them pairwise. We use the Iterative Prisoners Dilemma game simulated on scale-free networks, and use a genetic-algorithmic approach to investigate what optimal placement patterns evolve in terms of strategy. In particular, we ask the question that, given a pair of strategies are present in a network, which strategy should be adopted by the hubs (highly connected people), for the overall betterment of society (high network utility). We find that cooperation as opposed to defection, Pavlov as opposed to general cooperator, general cooperator as opposed to zero-determinant, and pavlov as opposed to zero-determinant, strategies will be adopted by the hubs, for the overall increased utility of the network. The results are interesting, since given a scenario where certain individuals are only capable of implementing certain strategies, the results give a blueprint on where they should be placed in a complex network for the overall benefit of the society.

Influence of vaccination strategies and topology on the herd immunity of complex networks

It is well known that non-vaccinated individuals may be protected from contacting a disease by vaccinated individuals in a social network through community protection (herd immunity). Such protection greatly depends on the underlying topology of the social network, the strategy used in selecting individuals for vaccination, and the interplay between these. In this paper, we analyse how the interplay between topology and immunization strategies influences the herd immunity of social networks. First, we introduce an area under curve measure which can quantify the levels of herd immunity in a social network. Then, using this measure, we analyse the above mentioned interplay in three ways: (1) by comparing vaccination strategies across topologies, (2) by analysing the influence of selected topological metrics, and (3) by considering the influence of network growth on herd immunity. For qualitative comparison, we consider three classical topologies (scale-free, random, and small-world) and three vaccination strategies (natural, random, and betweenness-based immunization). We show that betweenness-based vaccination is the best strategy of immunization in static networks, regardless of topology, but its prominence over other strategies diminishes in dynamically growing topologies. We find that the network features that lead to ‘small-worldness’ in networks (low diameter and high clustering) discourage herd immunity, regardless of the vaccination strategy, while preferential mixing (high assortativity) encourages it. In terms of growth, we demonstrate that herd immunity of random networks actually increases with growth, if the proportion of survivors to a secondary infection is considered, while the community protection in scale-free and small-world networks decreases with growth. Our work highlights the complex balance between social network structure and vaccination strategies in influencing community protection, and contributes a numerical measure to quantify this.

The p-index: Ranking Scientists Using Network Dynamics

Abstract The indices currently used by scholarly databases, such as Google scholar, to rank scientists, do not attach weights to the citations. Neither is the underlying network structure of citations considered in computing these metrics. This results in scientists cited by well-recognized journals not being rewarded, and may lead to potential misuse if documents are created purely to cite others. In this paper we introduce a new ranking metric, the p-index (pagerank-index), which is computed from the underlying citation network of papers, and uses the pagerank algorithm in its computation. The index is a percentile score, and can potentially be implemented in public databases such as Google scholar, and can be applied at many levels of abstraction. We demonstrate that the metric aids in fairer ranking of scientists compared to h-index and its variants. We do this by simulating a realistic model of the evolution of citation and collaboration networks in a particular field, and comparing h-index and p-index of scientists under a number of scenarios. Our results show that the p-index is immune to author behaviors that can result in artificially bloated h-index values.

The performance of page rank algorithm under degree preserving perturbations

Page rank is a ranking algorithm based on a random surfer model which is used in Google search engine and many other domains. Because of its initial success in Google search engine, page rank has become the de-facto choice when it comes to ranking nodes in a network structure. Despite the ubiquitous utility of the algorithm, little is known about the effect of topology on the performance of the page rank algorithm. Hence this paper discusses the performance of page rank algorithm under different topological conditions. We use scale-free networks and random networks along with a custom search engine we implemented in order to experimentally prove that the performance of page rank algorithm is deteriorated when the random network is perturbed. In contrast, scale-free topology is proven to be resilient against degree preserving perturbations which aids the page rank algorithm to deliver consistent results across multiple networks that are perturbed to varying proportions. Not only does the top ranking results emerge as stable nodes, but the overall performance of the algorithm is proven to be remarkably resilient which deepens our understanding about the risks in applying page rank algorithm without an initial analysis on the underlying network structure. The results conclusively suggests that while page rank algorithm can be applied to scale-free networks with relatively low risk, applying page rank algorithm to other topologies can be risky as well as misleading. Therefore, the success of the page rank algorithm in real world in search engines such as Google is at least partly due to the fact that the world wide web is a scale-free network. Since the world wide web is constantly evolving, we postulate that if the topological structure of the world wide web changes significantly so that it loses its scale-free nature to some extent, the page rank algorithm will not be as effective.

Deep Learning Approach for Classification of Mild Cognitive Impairment Subtypes.

Timely intervention in individuals at risk of dementia is often emphasized, and Mild Cognitive Impairment (MCI) is considered to be an effective precursor to Alzheimers disease (AD), which can be used as an intervention criterion. This paper attempts to use deep learning techniques to recognise MCI in the elderly. Deep learning has recently come to attention with its superior expressive power and performance over conventional machine learning algorithms. The current study uses variations of auto-encoders trained on neuropsychological test scores to discriminate between cognitively normal individuals and those with MCI in a cohort of community dwelling individuals aged 70-90 years. The performance of the auto-encoder classifier is further optimized by creating an ensemble of such classifiers, thereby improving the generalizability as well. In addition to comparable results to those of conventional machine learning algorithms, the auto-encoder based classifiers also eliminate the need for separate feature extraction and selection while also allowing seamless integration of features from multiple modalities.

A memory-efficient heuristic for maximum matching in scale-free networks

The maximum matching problem has been extensively studied, and several algorithms have been proposed which can maximize the percentage of matching. Nevertheless, these algorithms are designed without consideration of the topology of the networks on which they are intended to be applied. However, recent research has shown that many distributed systems from social, technical and biological domains which can be represented as networks display the scale-free structure, which has well-known topological characteristics such as the power-law degree distribution. In this paper, we describe a simple iterative heuristic for maximum matchings in scale-free networks which takes advantage of such characteristics. We show that our heuristic is no worse than the best known version of the Blossom algorithm in terms of time-complexity, and much better than any version of Blossom in terms of memory usage (space-complexity) when applied to typical scale-free networks. The heuristic due to its simplicity and memory efficiency is a viable alternative to Blossom in most real world applications.