Recommender systems have cemented themselves in the daily online activities
of most people, especially in the usage of streaming services, music services etc.
Recommender systems today usually make use of either or both collaborative
filtering and content-based filtering. However, they have yet to make the big
breakthrough in the more complex area of ’Narrative Driven Recommendation’,
an area where users often already have a vague or specific idea about the desired
items they want to be recommended. Most traditional recommender algorithms
struggle to correctly address these vaguely or specifically defined requirements,
which may vary a lot in their specificity.
We perform a literature review showing that even though Narrative Driven Recommendation (NDR) is not as common as the regular type of recommendation
that we see today, it is not just a ’niche problem’ either. We show that if there
were any NDR systems available today, we would see more users trying to express
their complex needs and preferences to such systems.
In this thesis, we present different ways of creating narratives from two datasets,
a dataset containing book-requests and a dataset containing POI-requests. We
use these narratives to create queries, and later we issue these queries against
indices of items using Elasticsearch. After receiving the results, we then rerank
the returned documents using BM25 scoring, a language model, and a language
model with query expansion.
To evaluate our experiments we look at how the narrative recommendations perform by themselves, and how the different datasets impact the results of the
queries. We also take a look at some reranking methods and compare these to
see if they work the same across the different datasets.