Utku Irmak

- Led multiple projects on communication relevance: (1) Email volume optimization by personalizing the email cadence, (2) Promotional content targeting by mining members’ historical activities, (3) Personalized ranking of app notifications. - Designed and developed offline modeling and scoring workflows to predict engagement in messages. Integrated the predictor into LinkedIn’s messaging platform. The solution is used by major LinkedIn products and provides significant engagement lifts.

- Designed and developed both offline modeling and real-time scoring components of the second-generation ad matching engine. Optimized the real-time scoring code and reduced the 95th percentile latency to less than 20ms. The model significantly improved the baseline eCPM and served 90% of LinkedIn Ads for two quarters. - Developed a seniority model to predict the member's experience level from the profile and the experience level required for a given job posting. Fully integrated the model into existing job recommendation engine. The A/B tests showed significant lift in job views and job applications; the model is deployed to 100%. - Developed a generic framework to identify similar entities for a given one (e.g., job title, industry, etc.) by leveraging the profile data. The solution is used in major Linkedin products.

- Contributed to the design and development of a large scale experimentation platform for the Behavioral Targeting (BT) advertising product. - Developed Pig and Map/Reduce programs that process user event logs on Hadoop and obtained important data insights. - Experimented with various new features and techniques that improve the BT models.

- Designed and developed efficient algorithms for the Contextual Shortcuts platform which recognizes named entities and concepts in text and creates intelligent hyperlinks to various Yahoo! services. - Designed and developed a machine-learning based ranking approach to identify the “key” entities and concepts in text by leveraging the historical click data (results published in ICDE 2009). - Designed and developed a novel text mining framework for the recognition of semi-structured named entities (e.g., phone numbers, date and time entities, etc.) in many languages, which reduced the editorial and engineering efforts significantly (results published in WWW 2010).

- Developed a novel semi-automatic wrapper generation system for the extraction of relational data on the web using active learning techniques (results published in WWW 2006). - Developed efficient remote file synchronization algorithms based on Erasure Codes and Set Reconciliation techniques which offer improvements over “rsync” (results published in INFOCOM 2005 and INFOCOM 2008). - Developed algorithms for efficient content distribution to low-bandwidth clients using value-based caching and delta encoding techniques (results published in WWW 2005).

- Developed efficient matching algorithms for “prospective” search engines which allow users to upload queries and run them on newly discovered pages in the future. Improved the algorithms using a clustering approach (results published in Usenix 2006).

Thesis: Algorithms for Information Extraction and Dissemination on the World-Wide Web Advisor: Prof. Torsten Suel