Using graphs and random walks for discovering latent semantic relationships in text

Erkan, Gunes University of Michigan 2007 해외박사(DDOD)

We propose a graph-based representation of text collections where the nodes are textual units such as sentences or documents, and the edges represent the pairwise similarity function between these units. We show how random walks on such a graph can give us better approximations for the latent similarities between two natural language strings. We also derive algorithms based on random walk models to rank the nodes in a text similarity graph to address the text summarization problem in information retrieval. The similarity functions used in the graphs are intentionally chosen to be very simple and language-independent to make our methods as generic as possible, and to show that significant improvements can be achieved even by starting with such similarity functions. We put special emphasis on language modeling-based similarity functions since we use them for the first time on problems such as document clustering and classification, and get improved results compared to the classical similarity functions such as cosine. Our graph-based methods are applicable to a diverse set of problems including generic and focused summarization, document clustering, and text classification. The text summarization system we have developed has ranked as one of the top systems in Document Understanding Conferences over the past few years. In document clustering and classification, using language modeling functions performs consistently better than using the classical cosine measure reaching as high as 25% improvement in accuracy. Random walks on the similarity graph achieve additional significant improvements on top of this. We also revisit the nearest neighbor text classification methods and derive semi-supervised versions by using random walks that rival the state-of-the-art classification algorithms such as Suppor Vector Machines.

Decentralized enterprise risk management for global companies

Erkan, Hafize Gaye Princeton University 2006 해외박사(DDOD)

Enterprise risk management for large financial organizations has become a significant concern due to several emerging trends. Modern financial companies span a spectrum of domains and operate in global markets. Motivating these conglomerates is, among others, the diversification of risks. Unfortunately recent events have shown that financial organizations haven't been able to fully anticipate their overall risks, especially during stressful market conditions. European financial firms, for example, have lost a large portion of their market value, due to the convergence of risks rather than achieving diversification. Pension plans have lost their ability to pay the beneficiaries. Several problems arise with current risk management approaches: (1) historical data may misrepresent the future, especially regarding tail events, (2) little coordination exists in organizations regarding their performance under possible future scenarios, (3) regulations sometimes prevent organizations from mitigating enterprise risks, and (4) the theory of managing large organizations is limited when dealing with risk management. A global financial company will often possess a relatively large number of units (divisions), depending upon the organizational setup and regulatory constraints. Regulations across countries are often inconsistent, and generally unable to keep up with the surge in new products and services. Thus, decentralized risk management provides a practical framework for managing the risks of a large organization while improving the firm's profitability. My thesis encompasses theory and applications that employ concepts of large-scale optimization, financial engineering, corporate finance and operations management. The research focuses mainly on the design of integrated risk management systems for multinationals, both financial and non-financial companies, via decentralized stochastic programs and stochastic control problems. Previous work on decomposition involves solving stochastic programs. However these approaches in the literature haven't formed the basis for decentralized decision-making. My research applies techniques from the area of large-scale optimization to integrated risk management defining decentralized enterprise risk management at the divisional level. The developed theory is implemented in two well-defined domains: property/casualty insurance and supply chain management. The proposed decentralized risk management framework brings dynamic portfolio theory and the rebalancing notion into multinational companies, where the divisions can be thought as part of company's portfolio.

The Impact of the Item Types and Number of Solution Steps of Multiple-Choice Items on Item Difficulty and Discrimination and Test Reliability

Atalmis, Erkan Hasan University of Kansas 2014 해외박사(DDOD)

This study examined two multiple choice item-writing guidelines addressed by Haladyna, Downing, and Rodriguez (2002). One is related to using the "None of the Above (NOTA)" option, the other is about the plausible number of options for a multiple-choice item (MCI). These two guidelines were empirically tested using one-step and multi-step problems to identify their impact on item characteristics (item difficulty and item discrimination) and test characteristics (test reliability). Three forms with MCIs were generated and administered to approximately 1500 7th and 8th grade students in the United States and Turkey. Bi-factor Item Response Theory (IRT) was applied to assess dimensionality related to the number of solution steps of items. Multiple regression models were employed to determine the degree of impact these item-writing guidelines had on item and test characteristics for MCIs with one step solution (MCI with one-step solution) and those with more than one step solutions (MCIs with multi-step solution). The results show that item characteristics do not change significantly across the conventional MCIs with four options, MCIs with three options, and MCIs with NOTA option. The interaction between solution steps and the three MCI types had no significant impact on item characteristics. For the test with MCIs with a one-step solution, the findings demonstrate that four options are significantly more reliable than the NOTA options and not statistically different from three options. For the test with MCIs with multi-step solutions, four options are not statistically different from three and NOTA options. Compared to MCIs with four options, the results support that MCIs with NOTA options are preferable for MCIs with multi-step solutions while three options are desirable for both MCIs with one-step solutions and multi-step solutions.

Particle-based mesoscale modeling of flow and transport in complex fluids

Tuzel, Erkan University of Minnesota 2007 해외박사(DDOD)

The dynamic behavior of complex liquids and soft materials is of great importance in a wide range of disciplines. Computational studies of these phenomena are particularly demanding because of the presence of disparate length and energy scales, and the complicated coupling between the embedded objects and the hydrodynamic flow field. The goal of this dissertation is to contribute to the understanding of these systems through the development and application of robust, quantitative mesoscale simulation techniques which incorporate both hydrodynamic interactions and thermal fluctuations. The work involves the further development of a specific particle-based mesoscale algorithm---stochastic rotation dynamics---which solves the hydrodynamic equations by following the discrete time dynamics of particles with continuous coordinates and velocities, using efficient multi-particle collisions. A detailed study of the long length- and time-scale properties of the algorithm, which involves analytical derivations of hydrodynamic equations, Green-Kubo relations, and transport coefficients is presented. Extensive simulations are performed to verify these results. The original algorithm is generalized to model dense fluids and binary mixtures. The equation of state and analytical expressions for the transport coefficients are derived. It is also shown that the non-ideal model exhibits an order-disorder transition and caging in the limit of large collision frequencies. The phase diagram of the entropically driven de-mixing transition of the binary mixture is presented, the surface tension for a droplet is calculated, and a detailed analysis of the capillary wave spectrum is performed. Finally, the algorithm is extended to amphiphilic mixtures in order to be able to study microemulsions and micelle formation. We have also developed a constrained dynamics algorithm for modeling the dynamical behavior of wormlike chains embedded in a mesoscale solvent. Rigorously enforced bond-length constraints permit the use of longer time steps, resulting in increased computational efficiency. It also eliminates high frequency degrees of freedom which often complicate comparison with theory and experiment. Finally, in order to provide guidance for experiments, we have modeled the behavior of thermally driven microtubules and simulated sources of experimental error that affect curvature distribution measurements used to understand the physical basis of microtubule bending in living cells.

Experiential Learning Once More: The Role of Non-Credit Academic Activities in International Business Education

Kocas, Erkan Michigan State University ProQuest Dissertations & 2022 해외박사(DDOD)

Extensive research has proven that integrating experiential learning techniques with traditional teaching methods improves student learning, success, and outcomes in higher education institutions (Derous & Ryan, 2008; Dolan & Stevens, 2006; Eyler, 2009; Kuh & Ikenberry, 2009; McCarthy & McCarthy, 2006; Pierson & Troppe, 2010). The proven success of these methods has inspired many fields to introduce experiential learning activities. As a result, international business education has also become a field that benefits from teaching through experience (Hagan, 2012). Many researchers have studied the impact of these methods in international business education, mostly focusing on classrooms and other for-credit activities. However, the studies on the effects of experiential learning in non-credit academic activities (NCAAs) have been limited in the past. To situate my research in this space, I expand the examination of NCAAs and study former students who have participated in an NCAA during their college enrollment. I enrich the study with my experiences and offer a detailed profile of the aforementioned NCAA. My research seeks to understand the role of experiential learning in non-credit academic settings and examine student perceptions about what they believe they have learned as international business skills. To accomplish this, I used Kolb’s experiential learning theory as a framework that also guided data coding, interview questions to former students, and further analysis. Five themes were identified from the analysis of the interviews: learning through experience, impactful experiential learning activities, personal development and self-confidence, employment, and learning international business. In this work, I offer my findings categorized under these five themes, an extensive profile of the NCAA I studied, my analysis and discussion of the research findings, my recommendations, and implications for future research.

Late-stage exhumation of metamorphic core complexes and landscape development during orogenic collapse of the North American Cordillera

Toraman, Erkan University of Minnesota 2014 해외박사(DDOD)

Exogenic and endogenic processes control the exhumation of partially molten crust in extending orogens. Their relative contribution to total denudation is critical to evaluate different tectonic models. Therefore, assessing the timing, rates, style, and conditions of events from melt crystallization to cooling at near-surface temperatures is significantly important for understanding the thermo-mechanical evolution of orogenic crust and linking deep-to-shallow processes. Metamorphic core complexes (MCC) and gneiss domes located within the hinterland of orogenic belts expose a significant quantity of former partially-molten mid-to-lower crust in the form of migmatites. The Thor-Odin, Frenchman's Cap, and Okanogan domes are exposed in the biggest Cordilleran-style metamorphic core complex, the Shuswap MCC, where a series of migmatitic gneiss domes formed during collapse of the thickened crust in the Cenozoic. The Thor-Odin and Frenchman's Cap domes form the Shuswap MCC's narrow northern end where the present day topographic relief reaches up to 2.5 km with deeply incised valleys and ubiquitous glacial features. The Okanogan dome, on the other hand, represents the wider, lower-relief (≤ 1km) southern termination. In the Thor-Odin and Frenchman's Cap domes, zircon U-Th/He ages range from 45 to 37 Ma. Apatite fission track ages range between 48 to 14 Ma and increase with increasing sample elevation. Thermal modeling of samples from higher altitudes (~2100 to 1800 m) verify only rapid Eocene cooling, whereas the lower-elevation samples (~1800 to 500 m) reveal an additional Plio-Quaternary cooling event. The presence of the top of a fossil Eocene partial annealing zone at ~1800 m indicates that the migmatite dome reached near-surface depths (1-2 km) during its initial exhumation mainly by detachment tectonics. Apatite U-Th/He chronometry of these samples yields Miocene (26-5 Ma) ages. A number of low elevation (~500 m) samples collected from valleys reveal intra-sample single grain U-Th/He ages. Combined with the results of thermal modeling, these age variations indicate a rapid exhumation pulse at ca. 3 Ma, possibly related to continental glaciation. In the Okanogan dome, zircon U-Th/He ages range from 51 to 41 Ma and decrease towards the detachment fault zone, emphasizing up to 3.7 km/myr slip rate on the detachment zone. Apatite fission track and U-Th/He ages vary from 51 to 23 Ma, recording a very slow phase of erosional exhumation of the dome that removed ~ 2 km of rocks subsequently after its initial rapid ascent facilitated by detachment tectonics in the Eocene. Low-temperature data also document different cooling paths of rocks depending on their structural level; rocks closer to the detachment zone display rapid cooling rates (≥ 100 °C/myr), whereas deeper structural levels cool slowly (10-30 °C/myr). As in the North American Cordilleran hinterland, a series of migmatite-cored metamorphic core complexes is exposed in the Hellenides, where the geodynamic context of migmatitic dome formation is well known from previous research. Multiple low-temperature thermochronologic techniques combined with existing structural, geo- and thermochronologic, and petrologic data from Cordilleran and central Aegean migmatitic gneiss domes document rapid ascent of partially molten mid-to-lower crustal rocks, facilitating significant mass and heat advection to the shallow crust. Heat advection results in a high-geothermal gradient in shallow crust and shifts the brittle-ductile transition zone close to the surface. Percolation of surface-derived fluids through fault and fracture zones enables rapid cooling of rocks and enhances the brittle rheology.

Integration of bulk piezoelectric materials into microsystems

Aktakka, Ethem Erkan University of Michigan 2012 해외박사(DDOD)

Bulk piezoelectric ceramics, compared to deposited piezoelectric thin-films, provide greater electromechanical coupling and charge capacity, which are highly desirable in many MEMS applications. In this thesis, a technology platform is developed for wafer-level integration of bulk piezoelectric substrates on silicon, with a final film thickness of 5-100microm. The characterized processes include reliable low-temperature (200°C) AuIn diffusion bonding and parylene bonding of bulk-PZT on silicon, wafer-level lapping of bulk-PZT with high-uniformity (+/-0.5microm), and low-damage micro-machining of PZT films via dicing-saw patterning, laser ablation, and wet-etching. Preservation of ferroelectric and piezoelectric properties is confirmed with hysteresis and piezo-response measurements. The introduced technology offers higher material quality and unique advantages in fabrication flexibility over existing piezoelectric film deposition methods. In order to confirm the preserved bulk properties in the final film, diaphragm and cantilever beam actuators operating in the transverse-mode are designed, fabricated and tested. The diaphragm structure and electrode shapes/sizes are optimized for maximum deflection through finite-element simulations. During tests of fabricated devices, greater than 12microm PP displacement is obtained by actuation of a 1mm2 diaphragm at 111kHz with <7mW power consumption. The close match between test data and simulation results suggests that the piezoelectric properties of bulk-PZT5A are mostly preserved without any necessity of repolarization. Three generations of resonant vibration energy harvesters are designed, simulated and fabricated to demonstrate the competitive performance of the new fabrication process over traditional piezoelectric deposition systems. An unpackaged PZT/Si unimorph harvester with 27mm3 active device volume produces up to 205microW at 1.5g/154Hz. The prototypes have achieved the highest figure-of-merits (normalized-power-density x bandwidth) amongst previously reported inertial energy harvesters. The fabricated energy harvester is utilized to create an autonomous energy generation platform in 0.3cm3 by system-level integration of a 50-80% efficient power management IC, which incorporates a supply-independent bias circuitry, an active diode for low-dropout rectification, a bias-flip system for higher efficiency, and a trickle battery charger. The overall system does not require a pre-charged battery, and has power consumption of <1microW in active-mode (measured) and <5pA in sleep-mode (simulated). Under lg vibration at 155Hz, a 70mF ultra-capacitor is charged from OV to 1.85V in 50 minutes.

CLAZOMENAE: THE ARCHAIC SETTLEMENT (GREECE)

ERSOY, YASAR ERKAN BRYN MAWR COLLEGE 1993 국내박사