Ömer Aynur , Kayseri, Türkiye

Extraction Phrase Chunking using syntaxnet Turkish Dependency Parser:
Extraction of phrase chunks from Turkish language using Google's syntaxnet dependency parser.Used nltk
tool kit for processing language.Due to this, we can extract meaningful groups of words in Turkish texts.

Sentiment Analysis For Customer Resource Services and Chat Bots:
Developed sentiment analysis system for industry compaines for their customer services using nltk, sklearn
and anaconda platform. Got above %90 accuracy and other evaluation metrics.The core of the model is
SVM and used more than 300.000 samples of domain specified data.

Telecom Fraud Detection :
Working on anomaly detection for preventing telecom frauds.Currently using couple of cluster based outlier
detection algorithms such as IForest, HBOS.Also trying ensemble methods to get the model working
better.Using several open source librares optimized to work in outlier detection well.Also using numpy,
pandas, sklearn to make things more efficient.


Telecom Fraud Detection with Autoencoders :
Using the same telecom fraud data, created autoencoder core for the detection to telecom frauds.In
encoder layer, dimensions of data is decreased.In decoder laye , dimensions of data is reconstructed as it
was before encoding operations.In that case, we have reconstruction errors.With proper threshold for
reconstruction errors, we are able to eliminate which is fraud calls or not.To build it, got help from Keras
library to construct autoencoder core.


Text Generation with Language Modelling:
Trained with 100.000 Turkish sentences using LSTM network to create N-gram language model aiming to
have Turkish text generator.Loss function used in this project is categorical crossentropy and size of
embedding is 20 which seemly most fitting.Used Keras and nltk library in Anaconda platform.


Traditional Recommendation System:
Using traditional method collaborative filtering is applied with neural dense layers.More than 300.000 user
and almost 1M item is used in training process.After training process is done, we can make
recommendations using either user-based collaborative filtering or item-based collaborative filtering.As
items are categorical, categorical cross entropy function is used as target function.Structure of dense layer
is based on utility matrix method in which rows are user, columns are items and data is ratings that users
rate.In this work, keras and pandas libraries is used intensely.


Session Based Recommendations System:
Using same data, extraction of sequential relations over ordered selected items in a session is aimed.In that
way, we make recommendations only based on items’ sequential relations.LSTM network is used to extract
sequential features of items.Size of sessions is up to average duration of purchases.In this work, size of
session is taken as 100 second.Keras and pandas libraries are used excessively.This technique is very similar
to forecasting of next best action for clients or users and is very used in retail industry.