Refine
Has Fulltext
- yes (2)
Document Type
- Doctoral Thesis (2) (remove)
Language
- English (2)
Is part of the Bibliography
- yes (2)
Keywords
- mixture models (2) (remove)
The prediction of the ground shaking that can occur at a site of interest due to an earthquake is crucial in any seismic hazard analysis. Usually, empirically derived ground-motion prediction equations (GMPEs) are employed within a logic-tree framework to account for this step. This is, however, challenging if the area under consideration has only low seismicity and lacks enough recordings to develop a region-specific GMPE. It is then usual practice to adapt GMPEs from data-rich regions (host area) to the area with insufficient ground-motion recordings (target area). Host GMPEs must be adjusted in such a way that they will capture the specific ground-motion characteristics of the target area. In order to do so, seismological parameters of the target region have to be provided as, for example, the site-specific attenuation factor kappa0. This is again an intricate task if data amount is too sparse to derive these parameters.
In this thesis, I explore methods that can facilitate the selection of non-endemic GMPEs in a logic-tree analysis or their adjustment to a data-poor region. I follow two different strategies towards this goal.
The first approach addresses the setup of a ground-motion logic tree if no indigenous GMPE is available. In particular, I propose a method to derive an optimized backbone model that captures the median ground-motion characteristics in the region of interest. This is done by aggregating several foreign GMPEs as weighted components of a mixture model in which the weights are inferred from observed data. The approach is applied to Northern Chile, a region for which no indigenous GMPE existed at the time of the study. Mixture models are derived for interface and intraslab type events using eight subduction zone GMPEs originating from different parts of the world. The derived mixtures provide satisfying results in terms of average residuals and average sample log-likelihoods. They outperform all individual non-endemic GMPEs and are comparable to a regression model that was specifically derived for that area.
The second approach is concerned with the derivation of the site-specific attenuation factor kappa0. kappa0 is one of the key parameters in host-to-target adjustments of GMPEs but is hard to derive if data amount is sparse. I explore methods to estimate kappa0 from ambient seismic noise. Seismic noise is, in contrast to earthquake recordings, continuously available. The rapidly emerging field of seismic interferometry gives the possibility to infer velocity and attenuation information from the cross-correlation or deconvolution of long noise recordings. The extraction of attenuation parameters from diffuse wavefields is, however, not straightforward especially not for frequencies above 1 Hz and at shallow depth. In this thesis, I show the results of two studies. In the first one, data of a small-scale array experiment in Greece are used to derive Love wave quality factors in
the frequency range 1-4 Hz. In a second study, frequency dependent quality factors of S-waves (5-15 Hz) are estimated by deconvolving noise recorded in a borehole and at a co-located surface station in West Bohemia/Vogtland. These two studies can be seen as preliminary steps towards the estimation of kappa0 from seismic noise.
Learning a model for the relationship between the attributes and the annotated labels of data examples serves two purposes. Firstly, it enables the prediction of the label for examples without annotation. Secondly, the parameters of the model can provide useful insights into the structure of the data. If the data has an inherent partitioned structure, it is natural to mirror this structure in the model. Such mixture models predict by combining the individual predictions generated by the mixture components which correspond to the partitions in the data. Often the partitioned structure is latent, and has to be inferred when learning the mixture model. Directly evaluating the accuracy of the inferred partition structure is, in many cases, impossible because the ground truth cannot be obtained for comparison. However it can be assessed indirectly by measuring the prediction accuracy of the mixture model that arises from it. This thesis addresses the interplay between the improvement of predictive accuracy by uncovering latent cluster structure in data, and further addresses the validation of the estimated structure by measuring the accuracy of the resulting predictive model. In the application of filtering unsolicited emails, the emails in the training set are latently clustered into advertisement campaigns. Uncovering this latent structure allows filtering of future emails with very low false positive rates. In order to model the cluster structure, a Bayesian clustering model for dependent binary features is developed in this thesis. Knowing the clustering of emails into campaigns can also aid in uncovering which emails have been sent on behalf of the same network of captured hosts, so-called botnets. This association of emails to networks is another layer of latent clustering. Uncovering this latent structure allows service providers to further increase the accuracy of email filtering and to effectively defend against distributed denial-of-service attacks. To this end, a discriminative clustering model is derived in this thesis that is based on the graph of observed emails. The partitionings inferred using this model are evaluated through their capacity to predict the campaigns of new emails. Furthermore, when classifying the content of emails, statistical information about the sending server can be valuable. Learning a model that is able to make use of it requires training data that includes server statistics. In order to also use training data where the server statistics are missing, a model that is a mixture over potentially all substitutions thereof is developed. Another application is to predict the navigation behavior of the users of a website. Here, there is no a priori partitioning of the users into clusters, but to understand different usage scenarios and design different layouts for them, imposing a partitioning is necessary. The presented approach simultaneously optimizes the discriminative as well as the predictive power of the clusters. Each model is evaluated on real-world data and compared to baseline methods. The results show that explicitly modeling the assumptions about the latent cluster structure leads to improved predictions compared to the baselines. It is beneficial to incorporate a small number of hyperparameters that can be tuned to yield the best predictions in cases where the prediction accuracy can not be optimized directly.