Personaliesd search.

It is our ambition to fundamentally move forward the state-of-the-art in personalised search - with a focus on e-commerce, e-news and video - by studying and developing new personalised search algorithms taking into account both the searched keywords and the full picture of the user's on- site and in-store (offline) behaviour. We will address the following research questions. First, in the context of personalised search, how can we measure and evaluate success? Personalised search is a relatively young research domain and as such there is not yet a standardised framework or benchmark dataset for evaluating performance, as there is in learning-to-rank or recommender systems. It is our goal to develop such a standardised framework and create a benchmark dataset that can be used across experiments. Additionally, given this project's unique position on the border between research and industry, we can not only measure the performance of the algorithms offline, but also online, with Froomle's existing clients. It is our expectation that clients in different industries will have different measures of success, e.g. clients in media may want to keep users engaged, whereas clients in retail might want to shorten the path to a purchase. Hence, we aim to identify these KPIs and lay down a framework for evaluation for each. Concretely, our goal is to do a live test in retail, in video and in news, evaluating the results with the KPI's developed specifically for the corresponding domain. Second, how can personal and search relevance be combined to determine an optimal ranking of items personalised to the individual? In order to provide the user with relevant search results ranked to their personal tastes, one needs to establish a means of combining (at least) two measures of relevance: relevance to the query and relevance to the person. Both measures can again be composites of multiple "features", e.g. pageviews, purchases, etc. for personal relevance and query match-score, authority and recency for search relevance. Here, we aim to identify which features can be relevant in delivering an optimal personalised search experience, e.g. pageviews and recency, but not authority and purchases. Then, we address the problem of combining these scores. This problem is anything but trivial and a static combination of personal and search relevance do not suffice. To solve this problem, we will develop at least one ranking algorithm that can transform multiple inputs into an optimal ranking, personalised to the individual. This requires that we will define at least one new learning objective that takes into account this personal aspect of the optimal ranking. Furthermore we will measure the corresponding performance improvement on at least one live application according to the principles and methodology derived by research question 1. Third, can we build an integrating ranking solution that approaches the problem of personalised search as a problem of optimally inferring the user's intent, rather than a problem of optimally combining the user's query with his historical behaviour? From this then builds the final research question. Rather than optimally combining query-based relevance with behaviour-based relevance, can we instead approach search as a recommendation problem, where a search query is merely an extra tool in our tool belt that will help us determine the user's current intent? Our goal is to develop at least one such algorithm and measure the corresponding performance improvement on at least one live application. Developing these new algorithms for personalised search and a framework for evaluation will allow Froomle to add personalised search to their current offering of advanced recommender systems. This will be an important step in bridging the gap between the giants of technology and other, traditionally offline businesses with a focus on e-commerce, e-news and video.

Keywords:DATA SCIENCE, ARTIFICIAL INTELLIGENCE (AI)

Disciplines:Machine learning and decision making

Project type:Collaboration project