Abstract: The objective of the thesis is to explore how complex data can be treated using unsupervised machine learning techniques, in which additional information is injected to guide the exploratory process. Starting from specific problems, our contributions take into account the different dimensions of the complex data: their nature (image, text), the additional information attached to the data (labels, structure, concept ontologies) and the temporal dimension. A special attention is given to data representation and how additional information can be leveraged to improve this representation.

---------

DOI: 10.1007/s10844-013-0235-x
Abstract: Feature-based format is the main data representation format used by machine learning algorithms. When the features do not properly describe the initial data, performance starts to degrade. Some algorithms address this problem by internally changing the representation space, but the newly-constructed features are rarely comprehensible. We seek to construct, in an unsupervised way, new features that are more appropriate for describing a given dataset and, at the same time, comprehensible for a human user. We propose two algorithms that construct the new features as conjunctions of the initial primitive features or their negations. The generated feature sets have reduced correlations between features and succeed in catching some of the hidden relations between individuals in a dataset. For example, a feature like sky ∧ ¬building ∧ panorama would be true for non-urban images and is more informative than simple features expressing the presence or the absence of an object. The notion of Pareto optimality is used to evaluate feature sets and to obtain a balance between total correlation and the complexity of the resulted feature set. Statistical hypothesis testing is used in order to automatically determine the values of the parameters used for constructing a data-dependent feature set. We experimentally show that our approaches achieve the construction of informative feature sets for multiple datasets.

---------

Abstract: Abstract—In this paper, we propose a new time-aware dissimilarity measure that takes into account the temporal dimension. Observations that are close in the description space, but distant in time are considered as dissimilar. We also propose a method to enforce the segmentation contiguity, by introducing, in the objective function, a penalty term inspired from the Normal Distribution Function. We combine the two propositions into a novel time-driven constrained clustering algorithm, called TDCK-Means, which creates a partition of coherent clusters, both in the multidimensional space and in the temporal space. This algorithm uses soft semi-supervised constraints, to encourage adjacent observations belonging to the same entity to be assigned to the same cluster. We apply our algorithm to a Political Studies dataset in order to detect typical evolution phases. We adapt the Shannon entropy in order to measure the entity contiguity, and we show that our proposition consistently improves temporal cohesion of clusters, without any significant loss in the multidimensional variance.

---------

Ref HAL: hal-00616247_v2
Abstract: We propose a system which employs conceptual knowledge to improve topic models by removing unrelated words from the simplified topic description. We use WordNet to detect which topical words are not conceptually similar to the others and then test our assumptions against human judgment. Results obtained on two different corpora in different test conditions show that the words detected as unrelated had a much greater probability than the others to be chosen by human evaluators as not being part of the topic at all. We prove that there is a strong correlation between the said probability and an automatically calculated topical fitness and we discuss the variation of the correlation depending on the method and data used.

---------

Ref HAL: hal-00616245_v1
Abstract: The growing number of statistical topic models led to the need to better evaluate their output. Traditional evaluation means estimate the model's fitness to unseen data. It has recently been proven than the output of human judgment can greatly differ from these measures. Thus the need for methods that better emulate human judgment is crucial. In this paper we present a system that computes the utility of individual topics from a given model on the basis of information drawn from a given ontology, in this case WordNet. The notion of utility is regarded as the ability to attribute concepts to each topic and separate words related to the topic from the unrelated ones based on that concept. In multiple experiments we prove the correlation between the automatic evaluation method and the answers received from human evaluators, for two different corpora and difficulty levels. By changing the evaluation focus from a statistical one to a conceptual one we were able to detect which topics are conceptually meaningful and rank them accordingly.

---------