Search

While complex machine learning methods can achieve great performance, human-interpretable details of their internal reasoning is to a large extent unavailable. Interpretable machine learning can remedy the lack of access to model reasoning but remains an elusive feat to fully achieve. Here we propose ranked selectors as a method for post-hoc explainability of classification outcomes from arbitrary classification models, with an initial emphasis on tabular data of moderate dimensions. The method is based on constructing a set of selectors, or rules, delimiting a local class consistent domain with maximal cover around the item of interest. The extended adjacent feature space is probed to achieve a ranking of the selectors. The method supports the use of an explicit foil class Q, allowing the formulation of contrastive queries in the form 'Why inference P instead of alternative inference Q?'. The answer is given as a short list of disjoint rules, a format previously shown to be amenable to human interpretation. We demonstrate the proposed method in open datasets, and elaborate on its stability aspects relative to other comparable methods.

We perform a computational study of the propagation of the oxygen concentration within a two-dimensional slice of a heterogeneous tumour region where the position and shape of the blood vessels are known. Exploiting the parameters space, we explore which effect is noticeable what concerns the formation of hypoxic zones. We use this information to anticipate a patient-specific radiation treatment with a potentially controlled response of the cancer growth.

For low dimensional classification problems we propose the novel DIOPT approach which considers the construction of a discretized feature space. Predictions for all cells in this space are obtained by means of a reference classifier and the class labels are stored in a lookup table generated by enumerating the complete space. This then leads to extremely high classification throughput as inference consists only of discretizing the relevant features and reading the class label from the lookup table index corresponding to the concatenation of the discretized feature bin indices. Since the size of the lookup table is limited due to memory constraints, the selection of optimal features and their respective discretization levels is paramount. We propose a particular supervised discretization approach striving to achieve maximal class separation of the discretized features, and further employ a purpose-built memetic algorithm to search towards the optimal selection of features and discretization levels. The inference run time and classification accuracy of DIOPT is compared to benchmark random forest and decision tree classifiers in several publicly available data sets. Orders of magnitude improvements are recorded in classification runtime with insignificant or modest degradation in classification accuracy for many of the evaluated binary classification tasks.

The focus of this thesis was to study if the level of fatigue could be assessed via nocturnal heart rate variability (HRV). HRV is a phenomenon that can be used to evaluate the status of the autonomic nervous system. A commonly used method for assessing the level of fatigue via HRV is an orthostatic test. The test is widely used by athletes in order to maintain the right balance between training and recovery. The test should be done every day as soon as possible after waking up, and therefore it may seem too laborious for nonprofessional athletes. Evaluating fatigue via nocturnal HRV would make monitoring fatigue less laborious by automating the measurement. The data that was used in this study consisted of 234 overnight recordings of RRi (time interval between adjacent heart beats) and 234 orthostatic test from 11 different persons. The results of the orthostatic test were then compared to different HRV variables calculated from the nocturnal RRi measurements. The nocturnal HRV variables were calculated from two different segments from the RRi data. The results reveal that the standard deviation of the RRi (std RRi) from the estimated slow-wave sleep (SWS) segment correlated with the mean RRi in the orthostatic test in standing position. The average correlation was minor (τ = 0.26 ± 0.26) and it was calculated with non-parametric Kendall’s Tau test. Out of 11 participants, the correlation was statistically significant (p < 0.05) for eight participants. It was concluded that the std RRi can not yet replace the orthostatic test in fatigue analysis. Despite the significant correlation, further analysis on the similarity of the std RRi and the orthostatic test should be conducted. Also due to interindividual differences, the suitability of this nocturnal parameter needs to be verified for each individual before one can consider leaving out the orthostatic test., Tässä diplomityössä tutkittiin ihmisen rasituksen tason tunnistamista unenaikaisen sykevälivaihtelun (HRV) avulla. Sykevälivaihtelun avulla pyritään selvittämään autonomisen hermoston tila, joka puolestaan reagoi fyysiseen harjoitteluun ja siitä aiheutuvaan rasitukseen. Rasitusta arvioidaan ortostaattisen sykevälivaihtelutestin avulla, joka on erityisesti urheilijoiden käytössä. Ortostaattinen testi tulisi suorittaa päivittäin, aamulla heti heräämisen jälkeen. Tämän vuoksi se voi tuntua työläältä. Unen aikana tehtävä automaattinen rasittuneisuuden mittaus voisi helpottaa oman rasituksen tason seurantaa. Tutkimuksessa käytetty aineisto koostui 234:stä ortostaattisesta testistä ja yhtä monesta läpi yön kestäneestä RRi (peräkkäisten sydämen lyöntien välinen aika) mittauksesta. Aineisto oli kerätty 11:ltä eri ihmiseltä. Unenaikaset sykevälivaihtelumuuttujat laskettiin kahdella erilaisella aikaikkunalla. Rasituksen arviointia unenaikaisesta sykevälivaihtelusta tutkittiin vertaamalla unenaikaista sykevälivaihtelua ortostaattisen testin tuloksiin. Vertailu tehtiin laskemalla korrelaatio yksittäisen henkilön unenaikaisen sykevälivaihtelun ja ortostaattisen testin tuloksien välille. Voimakkain korrelaatio löydettiin arvioidun syvänunenaikaisen aikaikkunan (SWS segment) RRi:n keskihajonnan ja ortostaattisen testin seisomavaiheen keskimääräisen RRi:n välillä. Korrelaatio laskettiin parametrittomalla Kendallin Tau testillä ja keskimäärin se oli pieni (τ = 0.26 ± 0.26). Käytössä olleen data-aineiston 11 osallistujasta kahdeksalla tämä korrelaatio oli kuitenkin tilastollisesti merkittävä (p < 0.05). Unenaikaisen RRi:n keskihajonta ei sellaisenaan sovellu korvaamaan ortostaattista testiä. Lisätutkimusta unenaikaisen RRi:n keskihajonnan käyttäytymisestä suhteessa ortostaattisen testin tuloksiin tarvitaan. Lisäksi suurien yksilöiden välisten erojen vuoksi, unenaikaisen RRin: keskihajonnan sopivuus tulisi varmistaa yksilökohtaisesti.

As end-to-end encryption on the Internet is becoming more prevalent, techniques such as deep packet inspection (DPI) can no longer be expected to be able to classify traffic. In many cellular networks a large fraction of all traffic is video traffic, and being able to divide flows in the network into video and non-video can provide considerable traffic engineering benefits. In this study we examine machine learning based flow classification using features that are available also for encrypted flows. Using a data set of several several billion packets from a live cellular network we examine the obtainable classification performance for two different ensemble-based classifiers. Further, we contrast the classification performance of a statistical-based feature set with a less computationally demanding alternate feature set. To also examine the runtime aspects of the problem, we export the trained models and use a tailor-made C implementation to evaluate the runtime performance. The results quantify the trade-off between classification and runtime performance, and show that up to 1 million classifications per second can be achieved for a single core. Considering that only the subset of flows reaching some minimum flow length will need to be classified, the results are promising with regards to deployment also in scenarios with very high flow arrival rates., HITS

Flow classification is an important tool to enable efficient network resource usage, support traffic engineering, and aid QoS mechanisms. As traffic is increasingly becoming encrypted by default, flow classification is turning towards the use of machine learning methods employing features that are also available for encrypted traffic. In this work we evaluate flow features that capture the distributional properties of in-flow per-packet metrics such as packet size and inter-arrival time. The characteristics of such distributions are often captured with general statistical measures such as standard deviation, variance, etc. We instead propose a Kolmogorov-Smirnov discretization (KSD) algorithm to perform histogram bin construction based on the distributional properties observed in the data. This allows for a richer, histogram based, representation which also requires less resources for feature computation than higher order statistical moments. A comprehensive evaluation using synthetic data from Gaussian and Beta mixtures show that the KSD approach provides Jensen-Shannon distance results surpassing those of uniform binning and probabilistic binning. An empirical evaluation using live traffic traces from a cellular network further shows that when coupled with a random forest classifier the KSD-constructed features improve classification performance compared to general statistical features based on higher order moments, or alternative bin placement approaches., HITS

Using machine learning in high-speed networks for tasks such as flow classification typically requires either very resource efficient classification approaches, large amounts of computational resources, or specialized hardware. Here we provide a sketch of the discretize-optimize (DISCO) approach which can construct an extremely efficient classifier for low dimensional problems by combining feature selection, efficient discretization, novel bin placement, and lookup. As feature selection and discretization parameters are crucial, appropriate combinatorial optimization is an important aspect of the approach. A performance evaluation is performed for a YouTube classification task using a cellular traffic data set. The initial evaluation results show that the DISCO approach can move the Pareto boundary in the classification performance versus runtime trade-off by up to an order of magnitude compared to runtime optimized random forest and decision tree classifiers., HITS, 4707