Abstract
Due to the growing number of digital products, recommendation systems are becomingincreasingly important in order to provide users with personalized recommendations for
items and thus support them in the selection of items. The sparsity problem in the
area of recommendation systems arises from the small number of users who rate items.
With a small amount of feedback for items, it is difficult for recommendation systems to
create relevant personalized recommendations for users. In addition, short-lived items
pose a challenge, as they can only be recommended within a certain period of time.
This work shows strategies to overcome the sparsity problem by integrating additional information about users into the original interest values. For this purpose, implicit feedback is transformed into a numerical representation of interest. Furthermore,
machine learning methods are used that can predict interest values for other categories
based on known interests of a user. The methods used to predict the interest information
for users are correlation coefficients, matrix factorization, singular value decomposition,
association rules, linear regression, decision trees, gradient boosting and random forest.
In addition, a method for creating recommendations for users where no information
about their interests is known is explained.
In addition to an offline evaluation, an online evaluation is performed to evaluate
the strategies by testing the developed strategies on French users. Two A/B tests with
a duration of 14 days each were carried out for this purpose. In the first test, a distinction is made between the standard item sequence in group A and a random item
sequence in group B. The second test creates recommendations in group A based on the
information collected from users and transformed into interest values, while in group
B these interest values are enriched with additional categories in order to identify new
interesting categories for the user. A significant increase in the average number of clicks
on items was achieved by the developed strategies compared to standard or randomized
recommendations. In addition, by enriching the interests of users, an increased number
of clicks by 5 % can be obtained.
| Date of Award | 2024 |
|---|---|
| Original language | German (Austria) |
| Supervisor | Ulrich Bodenhofer (Supervisor) |