Leveraging multi-criteria customer feedback for satisfaction analysis and improved recommendations

Abstract

Travel websites and online booking platforms represent today’s major sources for customers when gathering information before a trip. In particular, community-provided customer reviews and ratings of various tourism services represent a valuable source of information for trip planning. With respect to customer ratings, many modern travel and tourism platforms—in contrast to several other e-commerce domains—allow customers to rate objects along multiple dimensions and thus to provide more fine-granular post-trip feedback on the booked accommodation or travel package. In this paper, we first show how this multi-criteria rating information can help to obtain a better understanding of factors driving customer satisfaction for different segments. For this purpose, we performed a Penalty-Reward contrast analysis on a data set from a major tourism platform, which reveals that customer segments significantly differ in the way the formation of overall satisfaction can be explained. Beyond the pure identification of segment-specific satisfaction factors, we furthermore show how this fine-granular rating information can be exploited to improve the accuracy of rating-based recommender systems. In particular, we propose to utilize user- and object-specific factor relevance weights which can be learned through linear regression. An empirical evaluation on datasets from different domains finally shows that our method helps us to predict the customer preferences more accurately and thus to develop better online recommendation services.

Appendix: parameter optimization for weighted prediction model

The goal of the weight optimization process described in Sect. 3.3 is to find weight parameters $w_u*$ and $w_i*$ that minimize the prediction error on the training data and at the same time do not overfit the model to the data. The optimization goal is given in Eq. 3, where $K$ corresponds to the user-item rating tuples in the training set and lambda is the penalty factor.

$$\begin{aligned} \min\limits_{{w_i*, w_u*}} \sum\limits_{{(u,i) \in K}} \left(r_{u,i} - (w_{u} * \hat{r}_{u,i}^{user} + w_{i} * \hat{r}_{u,i}^{item})\right)^2 + \lambda \left(\sum\limits_{{u}} w_{u}^2 + \sum\limits_{{i}} w_{i}^2\right) \end{aligned}$$

(3)

Algorithm 1 shows our procedure to iteratively optimize the weights similar to the gradient descent approach from Koren (2010) and other recent works. The algorithm starts with randomly chosen initial weights and iterates over all ratings in the training set. It generates predictions with the current weights and compares them with the true ratings. Based on the observed error, the weights are then slightly adjusted. This procedure is repeated for a pre-defined number of iterations (e.g., 50). The parameters $\gamma$ and $\lambda$ determine the step size for the weight adaptation and a penalty factor for overfitting (Jannach et al. 2012).