Ahmet Bulut

SWAT: hierarchical stream summarization in large networks

The problem of statistics and aggregate maintenance over data streams has gained popularity in recent years especially in telecommunications network monitoring, trend-related analysis, Web-click streams, stock tickers, and other time-variant data. The amount of data generated in such applications can become too large to store, or if stored too large to scan multiple times. We consider queries over data streams that are biased towards the more recent values. We develop a technique that summarizes a dynamic stream incrementally at multiple resolutions. This approximation can be used to answer point queries, range queries, and inner product queries. Moreover, the precision of answers can be changed adoptively by a client. Later, we extend the above technique to work in a distributed setting, specifically in a large network where a central site summarizes the stream and clients ask queries. We minimize the message overhead by deciding what and where to replicate by using an adaptive replication scheme. We maintain a hierarchy of approximations that change adoptively based on the query and update rates. We show experimentally that our technique performs better than existing techniques: up to 50 times better in terms of approximation quality, up to four orders of magnitude times better in response time, and up to five times better in terms of message complexity.

Real-time nondestructive structural health monitoring using support vector machines and wavelets

We present an alternative to visual inspection for detecting damage to civil infrastructure. We describe a real-time decision support system for nondestructive health monitoring. The system is instrumented by an integrated network of wireless sensors mounted on civil infrastructures such as bridges, highways, and commercial and industrial facilities. To address scalability and power consumption issues related to sensor networks, we propose a three-tier system that uses wavelets to adaptively reduce the streaming data spatially and temporally. At the sensor level, measurement data is temporally compressed before being sent upstream to intermediate communication nodes. There, correlated data from multiple sensors is combined and sent to the operation center for further reduction and interpretation. At each level, the compression ratio can be adaptively changed via wavelets. This multi-resolution approach is useful in optimizing total resources in the system. At the operation center, Support Vector Machines (SVMs) are used to detect the location of potential damage from the reduced data. We demonstrate that the SVM is a robust classifier in the presence of noise and that wavelet-based compression gracefully degrades its classification accuracy. We validate the effectiveness of our approach using a finite element model of the Humboldt Bay Bridge. We envision that our approach will prove novel and useful in the design of scalable nondestructive health monitoring systems.

Cloud-based SVM for food categorization

As people across the globe are becoming more interested in watching their weight, eating more healthily, and avoiding obesity, a system that can measure calories and nutrition in everyday meals can be very useful. Recently, due to ubiquity of mobile devices such as smart phones, the health monitoring applications are accessible by the patients practically all the time. We have created a semi-automatic food calorie and nutrition measurement system via mobile that can help patients and dietitians to measure and manage daily food intake. While segmentation and recognition are the two main steps of a food calorie measurement system, in this paper we have focused on the recognition part and mainly the training phase of the classification algorithm. This paper presents a cloud-based Support Vector Machine (SVM) method for classifying objects in cluster. We propose a method for food recognition application that is referred to as the Cloud SVM training mechanism in a cloud computing environment with Map Reduce technique for distributed machine learning. The results show that by using cloud computing system in classification phase and updating the database periodically, the accuracy of the recognition step has increased in single food portion, non-mixed and mixed plate of food compared to LIBSVM.

Distributed data streams indexing using content-based routing paradigm

In recent years, we have seen a dramatic increase in the use of data-centric distributed systems such as global grid infrastructures, sensor networks, network monitoring, and various publish-subscribe systems. The realization of this potential requires adequate support from middleware that could be used to deploy and support such systems. In this regard, we propose an integrated distributed indexing architecture that supports scalable handling of intense dynamic information flows. The architecture is geared towards providing timely responses to queries of different types while minimizing the use of network and computational resources. The underlying communication framework ensures scalability and load balancing of communication as well as adaptivity in presence of dynamic changes. We elaborate on database and content-based routing methodologies used in the integrated solution as well as non-trivial interaction between them, and thereby provide a valuable feedback to the designers of these techniques. We demonstrate the effectiveness of our architecture with performance results that we obtained using our prototype implementation on top of the Chord system simulator.

Optimization Techniques for Reactive Network Monitoring

We develop a framework for minimizing the communication overhead of monitoring global system parameters in IP networks and sensor networks. A global system predicate is defined as a conjunction of the local properties of different network elements. A typical example is to identify the time windows when the outbound traffic from each network element exceeds a predefined threshold. Our main idea is to optimize the scheduling of local event reporting across network elements for a given network traffic load and local event frequencies. The system architecture consists of N distributed network elements coordinated by a central monitoring station. Each network element monitors a set of local properties and the central station is responsible for identifying the status of global parameters registered in the system. We design an optimal algorithm, the partition and rank (PAR) scheme, when the local events are independent; whereas, when they are dependent, we show that the problem is NP-complete and develop two efficient heuristics: the PAR for dependent events (PAR-D) and adaptive (Ada) algorithms, which adapt well to changing network conditions, and outperform the current state of the art techniques in terms of communication cost.

An Adaptive and Scalable Middleware for Distributed Indexing of Data Streams

We are witnessing a dramatic increase in the use of data-centric distributed systems such as global grid infrastructures, sensor networks, network monitoring, and various publish-subscribe systems. The visions of massive demand-driven data dissemination, intensive processing, and intelligent fusion in order to build dynamic knowledge bases that seemed infeasible just a few years ago are about to come true. However, the realization of this potential demands adequate support from middleware that could be used to deploy and support such systems.

Erratum to: Cloud-based SVM for food categorization

The original version of this article did not include a reference for section III-part (b) and Figure 2 to “Catak FO, Balaban ME (2013) CloudSVM: Training an SVM Classifier in Cloud Computing Systems. In: Zu Q, Hu B, Elci A (eds) Pervasive Computing and the Networked World. Joint International Conference, ICPCA/SWS 2012, Istanbul, Turkey, November 28-30, 2012, Revised Selected Papers. Lecture Notes in Computer Science, vol 7719. Springer, Berlin Heidelberg, p 57–68”. There is also a missing reference in section IX-Part (d)Subsection(2) to “https://code.google.com/p/cascadesvmand” and “Graf HP, Cosatto E, Bottou L, Dourdanovic I, Vapnik V (2004) Parallel Support Vector Machines: The Cascade SVM. Advances in Neural Information Processing Systems, p 521–528”. Multimed Tools Appl (2015) 74:5261 DOI 10.1007/s11042-015-2666-6

TopicMachine: Conversion Prediction in Search Advertising Using Latent Topic Models

Search Engine Marketing (SEM) agencies manage thousands of search keywords for their clients. The campaign management dashboards provided by advertisement brokers have interfaces to change search campaign attributes. Using these dashboards, advertisers create test variants for various bid choices, keyword ideas, and advertisement text options. Later on, they conduct controlled experiments for selecting the best performing variants. Given a large keyword portfolio and many variants to consider, campaign management can easily become a burden on even experienced advertisers. In order to target users in need of a particular service, advertisers have to determine the purchase intents or information needs of target users. Once the target intents are determined, advertisers can target those users with relevant search keywords. In order to formulate information needs and to scale campaign management with increasing number of keywords, we propose a framework called Topic Machine, where we learn the latent topics hidden in the available search terms reports. Our hypothesis is that these topics correspond to the set of information needs that best match-make a given client with users. In our experiments, Topic Machine outperformed its closest competitor by 41 percent on predicting total user subscriptions.

AdScope: Search Campaign Scoping Using Relevance Feedback

In a search ad campaign, the host search network provides information on expenses and revenue, user clicks, conversions, and search queries issued pre-click. An experienced advertiser goes through these queries and identifies the relevant ones to sharpen and the irrelevant ones to shrink to improve the campaigns reach and scope. With the right scope, the budget can be spent to target relevant users. AdScope ranks user queries with respect to relevance and recommends to advertisers the topmost queries for inclusion in a campaign and the bottommost queries for exclusion. It does this by combining the feedback collected from both users and advertisers to improve the ranking. The authors measured AdScopes performance with relevance classification and found that it achieved the highest classification accuracy (89.3) percent for queries that contain at least two terms.

Çok boyutlu veriden rastgele ormanlar ile düşük karmaşıklıklı güdümsüz öğrenme

With the ever increasing rate of digital information available from online sources, information has gone from being scarce to being abundant. Big data analytics require low complexity and distributed computing techniques. We propose the use of randomized decision trees and their ensemble in the form of a forest for unsupervised learning. Random probing of good attributes reduces the computational complexity making learning feasible on high-dimensional big data. Using an ensemble of trees improves the learning. We propose a new splitting measure for tree construction and an aggregation mechanism for predictive learning (unsupervised classification). The experiments on standard datasets show that our proposed proposed method outperforms the state-of-the-art.