Wirtschaftswissenschaften
Topic models such as latent Dirichlet allocation (LDA) aim to identify latent topics within text corpora. However, although LDA-type models fall into the category of Natural Language Processing, the actual model input is heavily modified from the original natural language. Among other things, this is typically done by removing specific terms, which arguably might also remove information. In this paper, an extension to LDA is proposed called uLDA, which seeks to incorporate some of these formerly eliminated terms -- namely stop words -- to match natural topics more closely. After developing and evaluating the new extension on established fit measures, uLDA is then tasked with approximating human-perceived topics. For this, a ground truth for topic labels is generated using a human-based experiment. These values are then used as a reference to be matched by the model output. Results show that the new extension outperforms traditional topic models regarding out-of-sample fit across all data sets and regarding human topic approximation for most data sets. These findings demonstrate that the novel extension can extract valuable information from the additional data conveyed by stop words and shows potential for better modeling natural language in the future.
Recommending products that are helpful to customers and tailored to their needs is of pivotal importance for successful online retailing. Online purchase data is typically used to generate such recommendations. This dissertation studies two topic models that use purchase data to make product recommendations. The Author Topic Model (ATM) and Sticky Author Topic Model (Sticky ATM) are applied to the purchase data of an online retailer of animal health products, and their predictive performances are contrasted with those of the benchmark methods Unigram, Bigram, and Collaborative Filtering (CF). This work focuses on the generation of new product recommendations. To increase novelty in recommendations, a new pre-processing approach is presented. The data is prepared prior to model application such that more novel products are included in the recommendations. A total of six data preparation variants are tested. The key finding is that topic models are very competitive with the benchmark methods and outperform them with the data preparation variant, where repetitively purchased items (repeat items) and customers with one item transaction (single-item customers) are eliminated from the data. Marketing practitioners should consider this pre-processing when implementing topic models as recommender models in their online shops.