Price Determinants of Airbnb Apartments : An Approach with Deep Language Representations
Piispanen, Antti (2021)
2021
Master's Programme in Computational Big Data Analytics
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2021-05-05
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202104012805
https://urn.fi/URN:NBN:fi:tuni-202104012805
Tiivistelmä
Airbnb is a company that enables private persons to rent their apartments for short-term tourist accommodation. It facilitates the transaction between the buyer and the seller by hosting an online marketplace. The company has experienced rapid growth and currently mediates millions of stays annually around the globe. It has been considered a disruptive innovator because of its novel business model and its effect on local, conventional accommodation industry. As costs are a major factor affecting consumer choice in the hospitality industry and even more so with Airbnb, the attributes affecting the price are worth investigating.
This study aimed to make a contribution to the scientific literature on this current topic with a new methodological approach. This thesis applied natural language processing methods, more precisely, deep language representations. The aim was to determine if the listing-level textual data contains price-relevant semantic content. The precise research question this thesis tried to answer was: does the text contain relevant information for price determination even after all the other listing attributes have been taken into account. The two main approaches being employed were static word embeddings from pre-trained models and contextualized representations generated by fine-tuned models. The best-performing deep learning model on the chosen dataset was sequence representations from ALBERTlarge followed by a feed-forward network. The extracted feature was technically a price prediction – a summary feature of the price-relevant content of the text. The predictions of that model were incorporated into a linear and a non-linear model: a least-squares linear regression and a gradient boosted decision tree ensemble. The performance of the models was evaluated on a test set, both with and without the feature included. The extracted feature proved to be statistically significant, had an effect of non-trivial magnitude on the target variable and was useful in predicting the price. Thus, it can be considered to be a major price determinant. However, the non-linear model faired almost as well even without the feature included. On the other hand, the text content just by itself contains much of the information to model the price accurately. The models had better performance than in much of the previous research.
The conclusions of this thesis are twofold. Firstly, a Transformer-based model can be used for a regression tasks with reasonable success even though there is a disconnect in the pre-training task and the fine-tuning task. Secondly, the textual data of Airbnb listings can be considered to have some additional information pertaining to things customers find valuable and thus affect price besides the previously known numerical attributes. However, identifying said latent factors was, unfortunately, not successful.
This study aimed to make a contribution to the scientific literature on this current topic with a new methodological approach. This thesis applied natural language processing methods, more precisely, deep language representations. The aim was to determine if the listing-level textual data contains price-relevant semantic content. The precise research question this thesis tried to answer was: does the text contain relevant information for price determination even after all the other listing attributes have been taken into account. The two main approaches being employed were static word embeddings from pre-trained models and contextualized representations generated by fine-tuned models. The best-performing deep learning model on the chosen dataset was sequence representations from ALBERTlarge followed by a feed-forward network. The extracted feature was technically a price prediction – a summary feature of the price-relevant content of the text. The predictions of that model were incorporated into a linear and a non-linear model: a least-squares linear regression and a gradient boosted decision tree ensemble. The performance of the models was evaluated on a test set, both with and without the feature included. The extracted feature proved to be statistically significant, had an effect of non-trivial magnitude on the target variable and was useful in predicting the price. Thus, it can be considered to be a major price determinant. However, the non-linear model faired almost as well even without the feature included. On the other hand, the text content just by itself contains much of the information to model the price accurately. The models had better performance than in much of the previous research.
The conclusions of this thesis are twofold. Firstly, a Transformer-based model can be used for a regression tasks with reasonable success even though there is a disconnect in the pre-training task and the fine-tuning task. Secondly, the textual data of Airbnb listings can be considered to have some additional information pertaining to things customers find valuable and thus affect price besides the previously known numerical attributes. However, identifying said latent factors was, unfortunately, not successful.