Tibebu Biru verfügbar

Tibebu Biru

Senior Data Scientist, Data Analytics, Data Mining, Machine Learning, Database, Java

Profilbild von Tibebu Biru Senior Data Scientist, Data Analytics, Data Mining, Machine Learning, Database, Java aus Hamburg
  • 20095 Hamburg Freelancer in
  • Abschluss: MSc Informatik
  • Stunden-/Tagessatz:
  • Sprachkenntnisse: deutsch (gut) | englisch (Muttersprache) | französisch (gut)
  • Letztes Update: 28.10.2020
Profil (CV)

Diese Anzeige ist nur für angemeldete Nutzer möglich.

  • Data Analytics: Design, Exploratory Analysis and Implementation of custom data models for data management and representation
  • Visual Analytics: Presentation/visualization of data using ggplot and Matplotlib
  • Data Mining: Regression, Pattern Classification and Association-rule learning
  • Machine Learning: Decision Trees, SVM, Clustering and Neural Networks
  • Reinforcement Learning: Application of RL algorithms
  • (Q-Learning, SARSA, DQN, DRQN, PPO)
  • Database Programs: PostgreSQL, MySQL, MongoDB
  • Machine learning: Python, Scikit-learn, Tensorflow, PyTorch
  • Web Technology & Tools for Visual Analytics: HTML,CSS, JavaScript, Node.js, D3.js
  • MSc. in Informatics: Engineering and Management (Johannes Kepler, University of Linz, Austria, February 2020)
  • BSc. in Electrical and Computer Engineering (Jacobs University, Bremen, Germany, June 2017)
  • 08/2019 - 02/2020

  • Implementation & Evaluation of Reinforcement Learning Algorithms On a Mobile Development Board
  • Project Focus                        

    • As part of the implementation and evaluation of Reinforcement Learning(RL) algorithms for robot navigation, the performance of classical tabular-based methods i.e. Q-Learning and SARSA are investigated. Moreover, neural-based reinforcement learning techniques are developed and implemented to compare the performance with tabular methods. The robot operating system(ROS) is used to control the robot and a compatible simulation environment, Gazebo, is used as a robotics simulator. NVIDIA Jetson TX1 module, development board with GPU computing, is chosen to run different simulation experiments. The main objective of the project was to demonstrate the performance of various Reinforcement Learning algorithms and provide proof-of-concept to encourage future investigations on RL for mobile robot navigation.


    Main Project activities


    • Implementation of deep Q-Network(DQN) architecture for sensor-based mobile robot navigation.
    • Development of deep recurrent Q-Network(DRQN) which incorporates convolutional layers and Long. Short-Term Memory(LSTM) for vision-based navigation
    • Monitoring and analysis of the performance of Proximal Policy Optimization(PPO).
    • Comparison of the performance of sensor-based and vision-based robot navigation
    • Evaluation of implemented RL algorithms based on cumulative reward obtained training in different simulation environments.

  • 03/2019 - 06/2019

  • Learning From User-generated Data: Session-based and Context-aware Recommendation System
  • Project Focus                        

    • The goal of the project was to develop a session-based and context-aware recommender system that adapts a list of accommodations according to the needs of the user. The training and test set data  contained user actions(filter usage, search refinements, item interactions, item searches, item click-outs) and metadata for accomodations. Overall, the data contained about 16 million user-item interactions, 927,000 items metadata and test set of approximately 4 million user-item interactions. Then, the final prediction is to be made on about 254,000 items clicked by users. The ratings which have been incorporated with the users‘ interaction with the items(clicks) are used to build the model for the recommender system. In general, the number of clicks varies dramatically among users and this sparsity of data influenced the predictions that were made. In order to overcome the cold start problem, where a user appears for the first time, a popularity based on geolocation is used to provide recommendations.


    Main Project activities


    • Building a baseline model that predicts the most popular items to the user.
    • Associative-rule learning: It is used to explore items that frequently occur together.
    • Development and Implementation of recommendation system using collaborative filtering approach based on matrix factorization method.
    • Implementation of item-based collaborative filtering combined with popularity-based recommendation.
    • Implementation of recommendation system based on deep learning technique using gated recurrent unit(GRU).
    • Evaluation of each recommendation system implemented using mean reciprocal rank(MRR).
    • Writing a scientific paper that showed detailed procedure applied to make recommendations and choice of the best final model.

  • 11/2018 - 01/2019

  • Pattern Classification: Music and Speech Detection in Radio Broadcasts
  • Project Focus                        


    • The project is motivated by a real-world problem from the media world. The main focus of the work conducted is to develop a system that automatically classifies from beginning to end in radio streams where music is played. The training data comes from a large number of radio stations that are continuously monitored and archived. Moreover, the amount of royalties will be calculated that must be paid by the radio stations to the national Performance Rights Organisation (PRO) based on the amount of music detected. In general, the PRO is interested in passages with pure music where music is not overlaid with any other sounds particularly speech. In the context of this project, the final goal was to build a system that detects music in radio audio streams and that maximizes the expected revenue(income) for the PRO by optimizing a cost/gain matrix.


    Main Project activities


    • Visual analytics to discover trends, features related to each target class and subset of features highly correlated in radio stream data.
    • Rigorous experimentation with Random Forest, SVM, and Neural Networks with systematic evaluation of different parameter settings.
    • Building classifier models  for music and speech separately and then combining them to get a final classifier model.
    • Selection of strategy for cost-sensitive learning with    ensemble methods via voting.
    • Analysis and documentation of classification accuracy and relevant parameters to detect music in complex audio signals.
    • Submission of predictions for an independent test set data with the expected amount of revenue to be collected using the best classifier model.

  • 01/2017 - 06/2017

  • Implementation of Mesh-networks With LoRa Radios
  • Project Focus                        


    • Principally, this project is conducted to propose a method that aims to develop foundation for mesh-network using LoPy, which are micro-python programmable development boards, based on LoRa protocol. It also demonstrates the possibility to implement packet forward scheme successfully using LoPy. In addition, the concept of finite-state machines is applied on LoPy  to transition into node and gateway mode interchangeably. Finally, an evaluation to assess the performance of routing based on distance is implemented in MATLAB using stationary nodes.


    Main Project activities


    • Implementation and improvement of the addressing scheme to create communication between LoPy devices(nodes).
    • Design and implementation of a finite state machine to make a LoPy behave as sender and receiver at different times of message flow.
    • Integration of temperature sensors to test sending and receiving of data packets by using LoRa protocol.
    • Evaluation of the performance of mesh-network architecture based on received signal strength using simulation in MATLAB.