Some extensions to representation and encoding of structure in models of distributional semantics

This thesis makes several contributions towards improved methods for encoding structure in computational models of word meaning. New methods are proposed and evaluated which address the requirement of being able to easily encode linguistic structural features within a computational representation while retaining the ability to scale to large volumes of textual data. Various methods are implemented and evaluated on a range of evaluation tasks to demonstrate the effectiveness of the proposed methods.

Recognising facial expressions with noisy data

Techniques to improve the automated analysis of natural and spontaneous facial expressions have been developed. The outcome of the research has applications in several fields including national security (eg: expression invariant face recognition); education (eg: affect aware interfaces); mental and physical health (eg: depression and pain recognition).

A novel method for finding similarities between unordered trees using matrix data model

Trees are capable of portraying the semi-structured data which is common in web domain. Finding similarities between trees is mandatory for several applications that deal with semi-structured data. Existing similarity methods examine a pair of trees by comparing through nodes and paths of two trees, and find the similarity between them. However, these methods provide unfavorable results for unordered tree data and result in yielding NP-hard or MAX-SNP hard complexity. In this paper, we present a novel method that encodes a tree with an optimal traversing approach first, and then, utilizes it to model the tree with its equivalent matrix representation for finding similarity between unordered trees efficiently. Empirical analysis shows that the proposed method is able to achieve high accuracy even on the large data sets.

Improving PLDA speaker verification with limited development data

This paper analyses the probabilistic linear discriminant analysis (PLDA) speaker verification approach with limited development data. This paper investigates the use of the median as the central tendency of a speaker’s i-vector representation, and the effectiveness of weighted discriminative techniques on the performance of state-of-the-art length-normalised Gaussian PLDA (GPLDA) speaker verification systems. The analysis within shows that the median (using a median fisher discriminator (MFD)) provides a better representation of a speaker when the number of representative i-vectors available during development is reduced, and that further, usage of the pair-wise weighting approach in weighted LDA and weighted MFD provides further improvement in limited development conditions. Best performance is obtained using a weighted MFD approach, which shows over 10% improvement in EER over the baseline GPLDA system on mismatched and interview-interview conditions.

Taxonomic resolution and quantification of freshwater macroinvertebrate samples from an Australian dryland river : The benefits and costs of using species abundance data

In studies using macroinvertebrates as indicators for monitoring rivers and streams, species level identifications in comparison with lower resolution identifications can have greater information content and result in more reliable site classifications and better capacity to discriminate between sites, yet many such programmes identify specimens to the resolution of family rather than species. This is often because it is cheaper to obtain family level data than species level data. Choice of appropriate taxonomic resolution is a compromise between the cost of obtaining data at high taxonomic resolutions and the loss of information at lower resolutions. Optimum taxonomic resolution should be determined by the information required to address programme objectives. Costs saved in identifying macroinvertebrates to family level may not be justified if family level data can not give the answers required and expending the extra cost to obtain species level data may not be warranted if cheaper family level data retains sufficient information to meet objectives. We investigated the influence of taxonomic resolution and sample quantification (abundance vs. presence/absence) on the representation of aquatic macroinvertebrate species assemblage patterns and species richness estimates. The study was conducted in a physically harsh dryland river system (Condamine-Balonne River system, located in south-western Queensland, Australia), characterised by low macroinvertebrate diversity. Our 29 study sites covered a wide geographic range and a diversity of lotic conditions and this was reflected by differences between sites in macroinvertebrate assemblage composition and richness. The usefulness of expending the extra cost necessary to identify macroinvertebrates to species was quantified via the benefits this higher resolution data offered in its capacity to discriminate between sites and give accurate estimates of site species richness. We found that very little information (<6%) was lost by identifying taxa to family (or genus), as opposed to species, and that quantifying the abundance of taxa provided greater resolution for pattern interpretation than simply noting their presence/absence. Species richness was very well represented by genus, family and order richness, so that each of these could be used as surrogates of species richness if, for example, surveying to identify diversity hot-spots. It is suggested that sharing of common ecological responses among species within higher taxonomic units is the most plausible mechanism for the results. Based on a cost/benefit analysis, family level abundance data is recommended as the best resolution for resolving patterns in macroinvertebrate assemblages in this system. The relevance of these findings are discussed in the context of other low diversity, harsh, dryland river systems.

A comparison of spatially explicit landscape representation methods and their relationship to stream condition

1. Biodiversity, water quality and ecosystem processes in streams are known to be influenced by the terrestrial landscape over a range of spatial and temporal scales. Lumped attributes (i.e. per cent land use) are often used to characterise the condition of the catchment; however, they are not spatially explicit and do not account for the disproportionate influence of land located near the stream or connected by overland flow. 2. We compared seven landscape representation metrics to determine whether accounting for the spatial proximity and hydrological effects of land use can be used to account for additional variability in indicators of stream ecosystem health. The landscape metrics included the following: a lumped metric, four inverse-distance-weighted (IDW) metrics based on distance to the stream or survey site and two modified IDW metrics that also accounted for the level of hydrologic activity (HA-IDW). Ecosystem health data were obtained from the Ecological Health Monitoring Programme in Southeast Queensland, Australia and included measures of fish, invertebrates, physicochemistry and nutrients collected during two seasons over 4 years. Linear models were fitted to the stream indicators and landscape metrics, by season, and compared using an information-theoretic approach. 3. Although no single metric was most suitable for modelling all stream indicators, lumped metrics rarely performed as well as other metric types. Metrics based on proximity to the stream (IDW and HA-IDW) were more suitable for modelling fish indicators, while the HA-IDW metric based on proximity to the survey site generally outperformed others for invertebrates, irrespective of season. There was consistent support for metrics based on proximity to the survey site (IDW or HA-IDW) for all physicochemical indicators during the dry season, while a HA-IDW metric based on proximity to the stream was suitable for five of the six physicochemical indicators in the post-wet season. Only one nutrient indicator was tested and results showed that catchment area had a significant effect on the relationship between land use metrics and algal stable isotope ratios in both seasons. 4. Spatially explicit methods of landscape representation can clearly improve the predictive ability of many empirical models currently used to study the relationship between landscape, habitat and stream condition. A comparison of different metrics may provide clues about causal pathways and mechanistic processes behind correlative relationships and could be used to target restoration efforts strategically.

A comparison of techniques for the reconstruction of leaf surfaces from scanned data

The foliage of a plant performs vital functions. As such, leaf models are required to be developed for modelling the plant architecture from a set of scattered data captured using a scanning device. The leaf model can be used for purely visual purposes or as part of a further model, such as a fluid movement model or biological process. For these reasons, an accurate mathematical representation of the surface and boundary is required. This paper compares three approaches for fitting a continuously differentiable surface through a set of scanned data points from a leaf surface, with a technique already used for reconstructing leaf surfaces. The techniques which will be considered are discrete smoothing D2-splines [R. Arcangeli, M. C. Lopez de Silanes, and J. J. Torrens, Multidimensional Minimising Splines, Springer, 2004.], the thin plate spline finite element smoother [S. Roberts, M. Hegland, and I. Altas, Approximation of a Thin Plate Spline Smoother using Continuous Piecewise Polynomial Functions, SIAM, 1 (2003), pp. 208--234] and the radial basis function Clough-Tocher method [M. Oqielat, I. Turner, and J. Belward, A hybrid Clough-Tocher method for surface fitting with application to leaf data., Appl. Math. Modelling, 33 (2009), pp. 2582-2595]. Numerical results show that discrete smoothing D2-splines produce reconstructed leaf surfaces which better represent the original physical leaf.

Graph theory based representation of building information models for access control applications

Building information models are increasingly being utilised for facility management of large facilities such as critical infrastructures. In such environments, it is valuable to utilise the vast amount of data contained within the building information models to improve access control administration. The use of building information models in access control scenarios can provide 3D visualisation of buildings as well as many other advantages such as automation of essential tasks including path finding, consistency detection, and accessibility verification. However, there is no mathematical model for building information models that can be used to describe and compute these functions. In this paper, we show how graph theory can be utilised as a representation language of building information models and the proposed security related functions. This graph-theoretic representation allows for mathematically representing building information models and performing computations using these functions.

Large-scale analysis of soccer matches using spatiotemporal tracking data

Although the collection of player and ball tracking data is fast becoming the norm in professional sports, large-scale mining of such spatiotemporal data has yet to surface. In this paper, given an entire season's worth of player and ball tracking data from a professional soccer league (approx 400,000,000 data points), we present a method which can conduct both individual player and team analysis. Due to the dynamic, continuous and multi-player nature of team sports like soccer, a major issue is aligning player positions over time. We present a "role-based" representation that dynamically updates each player's relative role at each frame and demonstrate how this captures the short-term context to enable both individual player and team analysis. We discover role directly from data by utilizing a minimum entropy data partitioning method and show how this can be used to accurately detect and visualize formations, as well as analyze individual player behavior.

Identifying product families using data mining techniques in manufacturing paradigm

Identifying product families has been considered as an effective way to accommodate the increasing product varieties across the diverse market niches. In this paper, we propose a novel framework to identifying product families by using a similarity measure for a common product design data BOM (Bill of Materials) based on data mining techniques such as frequent mining and clus-tering. For calculating the similarity between BOMs, a novel Extended Augmented Adjacency Matrix (EAAM) representation is introduced that consists of information not only of the content and topology but also of the fre-quent structural dependency among the various parts of a product design. These EAAM representations of BOMs are compared to calculate the similarity between products and used as a clustering input to group the product fami-lies. When applied on a real-life manufacturing data, the proposed framework outperforms a current baseline that uses orthogonal Procrustes for grouping product families.

Text mining with semantic annotation : using enriched text representation for entity-oriented retrieval, semantic relation identification and text clustering

This project is a step forward in the study of text mining where enhanced text representation with semantic information plays a significant role. It develops effective methods of entity-oriented retrieval, semantic relation identification and text clustering utilizing semantically annotated data. These methods are based on enriched text representation generated by introducing semantic information extracted from Wikipedia into the input text data. The proposed methods are evaluated against several start-of-art benchmarking methods on real-life data-sets. In particular, this thesis improves the performance of entity-oriented retrieval, identifies different lexical forms for an entity relation and handles clustering documents with multiple feature spaces.

Representing team behaviours from noisy data using player role

Due to their unobtrusive nature, vision-based approaches to tracking sports players have been preferred over wearable sensors as they do not require the players to be instrumented for each match. Unfortunately however, due to the heavy occlusion between players, variation in resolution and pose, in addition to fluctuating illumination conditions, tracking players continuously is still an unsolved vision problem. For tasks like clustering and retrieval, having noisy data (i.e. missing and false player detections) is problematic as it generates discontinuities in the input data stream. One method of circumventing this issue is to use an occupancy map, where the field is discretised into a series of zones and a count of player detections in each zone is obtained. A series of frames can then be concatenated to represent a set-play or example of team behaviour. A problem with this approach though is that the compressibility is low (i.e. the variability in the feature space is incredibly high). In this paper, we propose the use of a bilinear spatiotemporal basis model using a role representation to clean-up the noisy detections which operates in a low-dimensional space. To evaluate our approach, we used a fully instrumented field-hockey pitch with 8 fixed high-definition (HD) cameras and evaluated our approach on approximately 200,000 frames of data from a state-of-the-art real-time player detector and compare it to manually labeled data.

Space and time efficiency of the forest-of-quadtrees representation

A forest of quadtrees is a refinement of a quadtree data structure that is used to represent planar regions. A forest of quadtrees provides space savings over regular quadtrees by concentrating vital information. The paper presents some of the properties of a forest of quadtrees and studies the storage requirements for the case in which a single 2m × 2m region is equally likely to occur in any position within a 2n × 2n image. Space and time efficiency are investigated for the forest-of-quadtrees representation as compared with the quadtree representation for various cases.

On probability-based inference under data missing by design

Whether a statistician wants to complement a probability model for observed data with a prior distribution and carry out fully probabilistic inference, or base the inference only on the likelihood function, may be a fundamental question in theory, but in practice it may well be of less importance if the likelihood contains much more information than the prior. Maximum likelihood inference can be justified as a Gaussian approximation at the posterior mode, using flat priors. However, in situations where parametric assumptions in standard statistical models would be too rigid, more flexible model formulation, combined with fully probabilistic inference, can be achieved using hierarchical Bayesian parametrization. This work includes five articles, all of which apply probability modeling under various problems involving incomplete observation. Three of the papers apply maximum likelihood estimation and two of them hierarchical Bayesian modeling. Because maximum likelihood may be presented as a special case of Bayesian inference, but not the other way round, in the introductory part of this work we present a framework for probability-based inference using only Bayesian concepts. We also re-derive some results presented in the original articles using the toolbox equipped herein, to show that they are also justifiable under this more general framework. Here the assumption of exchangeability and de Finetti's representation theorem are applied repeatedly for justifying the use of standard parametric probability models with conditionally independent likelihood contributions. It is argued that this same reasoning can be applied also under sampling from a finite population. The main emphasis here is in probability-based inference under incomplete observation due to study design. This is illustrated using a generic two-phase cohort sampling design as an example. The alternative approaches presented for analysis of such a design are full likelihood, which utilizes all observed information, and conditional likelihood, which is restricted to a completely observed set, conditioning on the rule that generated that set. Conditional likelihood inference is also applied for a joint analysis of prevalence and incidence data, a situation subject to both left censoring and left truncation. Other topics covered are model uncertainty and causal inference using posterior predictive distributions. We formulate a non-parametric monotonic regression model for one or more covariates and a Bayesian estimation procedure, and apply the model in the context of optimal sequential treatment regimes, demonstrating that inference based on posterior predictive distributions is feasible also in this case.