Statistical and machine learning methods for recommender systems in a complex, high dimensional, online marketplace.

This project is a cooperation between University of Oslo and FINN.no. It is concerned with developing novel statistical and machine learning methodologies in the area of recommendation systems: matching relevant content (items) to users in an online marketplace, where there is a large number of items and the information about each item and user is very limited. Marketplaces are platforms where users buy and sell various types of items. The items can range from low-value ones such as books and clothes to high-value ones such as cars and real estate properties. Sellers can also post non-tangible items such as job openings and services. Many marketplace sellers are non-professional individuals selling used items. A marketplace is similar to a e-commerce platform, but has a very large number of unique items across multiple categories from a very large and fragmented seller group. The project is focusing on how recommender algorithms can effectively explore what items a users wants, how the algorithm can propose a diverse set of items to the user, how to assign delayed user feedback into the algorithms and how to effectively use content information such as images, text and structured data in order to give the most relevant recommendation to the user. The algorithms will be developed and tested both in offline and online experiments. The project have successfully A/B tested around 10 different machine learning algorithms in FINN.no's recommendations products. Multiple of these have been successful and replace existing recommender systems. There is ongoing work on academic writing of some of these results. During the last period the project has, in addition to testing different machine learning algorithms for the recommendations products in FINN.no, published, made available and promoted a data set that is more suited for recommendations products for an online marketplace. The data set is more rich on information that could be beneficial to know when calculating what is the most relevant item. This type of information has previously not been publicly available.

-

This project is concerned with developing novel statistical and machine learning methodologies in the area of recommendation systems: matching relevant content (items) to users in an online marketplace, where there is a large number of items and the information about each item and user is very limited. Marketplaces are platforms where users buy and sell various types of items. The items can range from low-value ones such as books and clothes to high-value ones such as cars and real estate properties. Sellers can also post non-tangible items such as job openings and services. Many marketplace sellers are non-professional individuals selling used items, therefore marketplaces can be viewed as a special type of e-commerce that involves a very large number of unique items across multiple categories from a very large and fragmented seller group. This thesis will develop new statistical and machine learning methods to solve some of the important open problems for recommender system in marketplaces, building models, inferential and predictive procedures, and computational codes. The goal is to build and test these models in a real world recommender system. The project is a co-operation between multiple participants: Finn.no with the industrial, computational, algorithmical, data management experience; UiO with a strong experience in statistical models for big data and uncertainty quantification; the University of Lancaster with world leading experts in statistical learning, decision making and stochastic game theory.