Kemal Toprak Uçar

Predictive Analytics project contains 2 main applications which are 'Smart Recommendation' and 'Competitive Index Calculator' to extract inferences from previous campaigns and 3rd party data.

In Smart Recommendation, main campaign metrics are predicted for each platform to give some suggestions to the clients before the campaign launch. An extensive data analysis has been done for each platform and each KPI metrics to decide performance metrics and feature selection/engineering. A varied… Show more

Predictive Analytics project contains 2 main applications which are 'Smart Recommendation' and 'Competitive Index Calculator' to extract inferences from previous campaigns and 3rd party data.

In Smart Recommendation, main campaign metrics are predicted for each platform to give some suggestions to the clients before the campaign launch. An extensive data analysis has been done for each platform and each KPI metrics to decide performance metrics and feature selection/engineering. A varied complexity of AI models such as CatBoost, Distributed Random Forest, and Linear Regression are employed for model training and campaign prediction. 

According to the metrics of the previous campaigns, competitive inferences are extracted in the Competitive Index Calculator. These inferences are used as suggestions for the clients to set optimal configuration for the campaign launch. Descriptive analytics are used to generate these indices. Show less

A "marketplace" is an e-commerce medium where product and inventory information
is provided by varying third parties, whereas catalog service is hosted, and payments
are processed by the marketplace operator. As a result of increasing use of
marketplaces, e-commerce capabilities can now be accessed by everyone.
Consequently, both the number of merchants and products have been growing
exponentially. Such growth raises some problems including “Does product
description reflect… Show more

A "marketplace" is an e-commerce medium where product and inventory information
is provided by varying third parties, whereas catalog service is hosted, and payments
are processed by the marketplace operator. As a result of increasing use of
marketplaces, e-commerce capabilities can now be accessed by everyone.
Consequently, both the number of merchants and products have been growing
exponentially. Such growth raises some problems including “Does product
description reflect specifications of the real one?”, “Does the seller really own the
product?”, “Is this product legal for purchasing online”, “Is this product listed under
correct category”. These problems can lead to catastrophic results for e-commerce
companies, especially where most countries regulate e-commerce business. We
propose a methodology to detect an accurate product category from user-generated
content on e-commerce marketplaces. We devise an accurate system for automatic
categorization. In this work, we apply both Natural Language Processing (NLP) and Classification
methodologies. We transform unstructured text into vector representations of words
in our ML-ready dataset which includes non-moderated user-generated text including typos, special
punctuation, and abbreviations. Show less

A limited size grid world implementation whose size is given by user was implemented in Java. Whole three reinforcement learning methodologies were employed in the simulation. Formulas and algorithms were obtained from "Reinforcement Learning: An Introduction" by Andrew Barto and Richard S. Sutton.
Specifications:
-A basic GUI was implemented in order to visualize learning episode by episode.
-Monte Carlo was covered as First Visit method.
-A wind effect is implemented that can… Show more

A limited size grid world implementation whose size is given by user was implemented in Java. Whole three reinforcement learning methodologies were employed in the simulation. Formulas and algorithms were obtained from "Reinforcement Learning: An Introduction" by Andrew Barto and Richard S. Sutton.
Specifications:
-A basic GUI was implemented in order to visualize learning episode by episode.
-Monte Carlo was covered as First Visit method.
-A wind effect is implemented that can be enabled by user. (default is disabled) Show less

A simulator for an autonomous quadcopter was implemented where the Q-Learning algorithm was used for learning. However, the environment was discretized to make it possible to work with Q-Learning. Otherwise, if the environment would be considered as continuous, the state space would be infinitely large to converge to a feasible solution. The problem about continuous states will be overcome incorporating neural networks. Artificial Neural Networks (ANNs) are one the most efficient methods that… Show more

A simulator for an autonomous quadcopter was implemented where the Q-Learning algorithm was used for learning. However, the environment was discretized to make it possible to work with Q-Learning. Otherwise, if the environment would be considered as continuous, the state space would be infinitely large to converge to a feasible solution. The problem about continuous states will be overcome incorporating neural networks. Artificial Neural Networks (ANNs) are one the most efficient methods that can handle problems with continuous state space.
ANNs are inspired by biological neural networks. Parallel computing systems containing large number of simple processors are raked together with many interconnections. Imitating as a rather simplified version the human brain ANN models attempt to use some “organizational” principles. Learning process in the ANN context can be viewed as the problem of updating network architecture and connection weights so that a network can efficiently perform a specific task. Performance is improved over time by iteratively updating the weights in the network. ANNs' ability to automatically learn from examples makes them attractive and exciting.
An important pitfall of ANNs is that ANNs’ performance is deeply influenced by its the structure (i.e. how many processing nodes at each layer and the way they are connected). We employ NEAT to initiate an evolutionary search among various ANN structures to end up with one as fit as possible and satisfies the goals of the task. The advantage of NEAT compared with the other topology-based methods is that it does not have a bound on the complexity of the networks. NEAT can evolve networks of unbounded complexity from a minimal starting point which leads an increase on the efficiency of the search while minimizing the dimensions of the weight space.
Show less