CHAPTER 2 LITERATURE REVIEW
2.1 Travel recommendation system
This section reviews three topics. First, we briefly introduce and discuss popular recommendation technologies used in recommendation systems today, and position our research work through identifying their similarities and differences and examining the characteristics of tourism environment. Second, we overview the tourism expectation formation and define our target customers by clarifying their needs. Last, we introduce the relationship between color and emotion, color mixture, and their related works. From the literature review about colors and their relations to psychology and mathematics, we believe colors can serve as a modeling tool for representing and computing the images of the tourist expressive needs (psychologically dominant) and the images of destinations. In this research, we propose a model-based DRS that is based on color modeling and computing and emphasizes the emotional tourist service experience
2.1 Travel recommendation system
Recommendation systems (RS) are designed for comforting people searching for objects they want/need including but not limiting to products and services, or for navigating them with information in a complex environment where a large number of options exist. With characteristics actively or passively provided by users, recommendation systems can facilitate their decision making process in an individualized way (Burke, 2002; Göksedef and Gündüz-Ö güdücü, 2010; Zhang et al., 2009).
In practice, recommendation systems are widely applied in e-commerce, for example, Amazon.com successfully utilizes recommendation technology upon the principle ―people who bought X also bought Y‖ to suggest customers‘ favorable
‧
merchandises (Burke, 2002; Towle and Quinn, 2000; Wang et al., 2006). As one of e-commerce instances, the travel information service has additional features including wide-data, diverse types, and highly experience-related (OECD, 2010;
Zhang et al., 2009). Before discussing travel recommendation systems, we reviewed several recommendation technologies commonly used today in the following paragraphs.
Technologies used in recommendation systems can be generally categorized into content-based recommendation, memory-based (includes user-based and item-based) and model-based methods of collaborative filtering technology, knowledge-based recommendation, and hybrid recommendation systems composed by some of above technologies and other technologies as extensions (e.g. grid technology and semantic ontology) (Burke, 2002; Ricci, 2002; Towle and Quinn, 2000; Wang et al., 2006; Weng et al., 2009; Zhang et al., 2009).
Content-based recommendation utilizes item descriptions whose format can be either text-based or attribute-value based. Basically, it retrieves items to match user needs, preferences, and constraints represented in provided languages (e.g.
attributes) (Bridge et al., 2005; Ricci, 2002). In advance, similar to item-based approach emphasizing on rating data, Burke (2002) and Huang (2008) declare that content-based approach performs the recommendation based on the similarities between the content of a particular object and the contents of those objects in the current user‘s selection history. The most significant shortcoming of content-based method is the difficulty to analyze the multimedia content of items (Huang, 2008;
Zhang et al., 2009). Another issue may arise owing to the necessity of constructing categories with a set of feature variables for each kind of item. This is also the reason why many travel recommendation systems only focus on the suggestion of
‧
destinations, which are comparatively stable and reusable concepts (Ricci, 2002).
Instead of dealing with item contents, collaborative filtering approaches usually concentrate on rating data enabling the ability to represent complex objects such as music and movies. They are often divided into two categories:
memory-based (which includes user-based and item-based method) and model-based (Burke, 2002; Wang et al., 2006). In memory-based approach, a user-item matrix will form in the recommendation system through collecting scoring information explicitly or implicitly from users. With this matrix, user-based collaborative filtering predicts the scores of the current user‘s unrated items according to the historical rating data of his ―nearest neighbors‖, a team of users having similar preferences with the current user; while pure item-based collaborative filtering predicts the score of a particular item by averaging the current user‘s rating data of similar items in the past. However, there often exists sparsity problem due to that only rated items can be recommended, or a user must have rating records to be a candidate of ―nearest neighbors‖ (Bridge et al., 2005;
Burke, 2002; Göksedef and Gündüz-Ö güdücü, 2010; Wang et al., 2006; Zhang et al., 2009). In certain cases, item-based approach is a better choice because the item quantity is comparatively stable. Generally, there is an option to pursue efficiency － unifying user-based and item-based approaches brings the expansibility against the limited accuracy caused by a small portion usage of the user-item matrix (Wang et al., 2006; Zhang et al., 2009).
Model-based recommendation uses training examples (often historical rating data) with various learning technique (e.g. neural networks (Jennings and Higuchi, 1993), latent semantic indexing (Foltz, 1990), and Bayesian networks (Condliff et al., 1999)) to derive a model or to find patterns to predict item ratings for the
‧
problem to a certain degree (while knowledge-based approach is considered as a better solution to this problem but more difficult to realize). Although numerous parameters are needed to be tuned and the model construction is time-consuming (which can be implemented off-line), it is still worthwhile to adopt this approach because the quality of the recommendation results are usually found satisfying (Burke, 2002; Wang et al., 2006; Zhang et al., 2009).Knowledge-based recommendation infers the prediction with pre-existing functional knowledge describing how favorable items can meet user needs. While being one of solutions to the challenges of memory-based collaborative filtering there is still a tough point of this approach, the acquisition of knowledge. It requires knowledge about objects, knowledge about mappings between user needs and objects, and user knowledge. Google search engine is a good example using knowledge about the relations between Web pages to infer the desired query results (Burke, 2002). Furthermore, once the development of semantic ontology is mature, the vision of the accurate, automatic, and intelligent recommendation systems will come true (Burke, 2002; Hatala and Wakkary, 2005; Towle and Quinn, 2000; Zhang et al., 2009).
Hybrid recommendation systems combine two or more recommendation techniques to construct a module for better performance (Burke, 2002; Göksedef and Gündüz-Ö güdücü, 2010). Hybridization methods are various including Weighted, Mixed, Switching, Feature Combination, Cascade, Feature augmentation, and Meta-level. To adapt the dynamic e-commerce environment and various user needs, hybrid recommendation technologies are suitable solutions if adequate research work is conducted to define the tradeoffs (Burke,
‧
Overall, recommendation technologies we have discussed are all function-oriented which can be demonstrated in the following facts: 1) Content-based recommendation retrieves items from their descriptions. All the elements in this process including the queries from users, the matching method, and the item contents mainly manipulate functional item features such as a title and an author name of a book. 2) Collaborative filtering accounts on rating information from users. 3) Knowledge-based approach grounds functional knowledge related to items, users, and the relationship between the items and the user needs (Burke, 2002). However, since tourism is an industry which is highly experience-related, it is beneficial for travel recommendation systems to take emotional elements of travel products and psychological emotional needs of tourists into account.
According to the review of tourism executed by OECD (2010), experience economy is a trend and innovating product development and service delivery in tourism. SMEs within a travel destination provide services as value propositions and co-create service experiences with their customers, where a service experience is defined by B. Edvardsson (2005) as ―a service process that creates the customer‘s cognitive, emotional, and behavioral responses, resulting in a mental mark, a memory.‖ Finally, tourists evaluate the consumed service experiences according to the individual and situational filter, which results in the so-called value-in-use (Sandström et al., 2008).
Service experience creates value. However, Sandström et al. (2008) claimed that ―insufficient attention has been paid to the total service experience, including the emotional dimensions.‖ Besides, experiences are memorable, revealed, sensed by individual (Pine and Gilmore, 1998). That is, we cannot conclude that any two
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
travelers‘ trips are the same even though they went together (Ricci, 2002).
The comparison among different kinds of recommendation technology including ours is shown as the table 1.
‧
Table 1. The Comparison among Recommendation Technologies
Technique Content-based Collaborative Filtering Knowledge-based Color Imagery
Algorithm Retrieving Nearest Neighbors,
In the previous section, we learned that most of recommendation systems do not catch users‘ expectations of suggested objects (preference-driven systems) or