You are on page 1of 7

The Fourth International Conference on Electronic Business (ICEB2004) / Beijing 445

Towards Collaborative Travel Recommender Systems
Cane Wingki Leung, Stephen Chifai Chan, Korris Fulai Chung
Department of Computing, The Hong Kong Polytechnic University, Hong Kong, China
{cswkleung, csschan, cskchung}@comp.polyu.edu.hk

ABSTRACT
Collaborative filtering (CF) based recommender systems have been proven to be a promising solution to the problem of
information overload. Such systems provide personalized recommendations to users based on their previously
expressed preferences and that of other similar users. In the past decade, they have been successfully applied in various
domains, such as the recommendation of books and movies, where items are simple, independent and single units.
When applied in the tourism domain, however, CF falls short due to the simplicity of existing techniques and
complexity of tourism products. In view of this, a study was carried out to review the research problems and
opportunities. This paper details the results of the study, which includes a review on the recent developments in CF as
well as recommender systems in tourism, and suggests future research directions for personalized recommendation of
tourist destinations and products.

Keywords: Recommender System, Collaborative Filtering, Personalization, Data Mining, Travel & Tourism

1. INTRODUCTION recommended to a user can therefore be very different
(content-wise) from what the user has liked previously,
In the past decades, the rapid development of Internet overcoming a major limitation of content-based
technologies has led to severe information overload. recommender systems [31]. When compared to
The amount of information available on the web, and knowledge-based recommender systems, CF techniques
the amount added daily, is too much for us to process. are much simpler and easier to implement. While the
The use of recommender systems has therefore become former requires a domain knowledge engineering
increasingly popular. Such systems provide advices to process to build a knowledge base of items, the later can
users about items they might like to purchase or be fully automated. On the one hand, this allows CF to
examine, avoid them from being drown in the vast be easily applicable to any domain where a database of
amount of information [3]. user preferences is available. On the other hand, this
makes CF falls short when recommending complex
There are three major types of recommender systems: items such as travel purchases [4]. Nevertheless, CF has
content-based, knowledge-based and social- or become the most successful technology applied in
collaborative-filtering (CF) based. Content-based recommender systems. To date, it has already been
recommender systems establish users’ interest profiles studied and applied to various domains such as the
by analyzing the features of their preferred items. In recommendation of books, Usenet articles and jokes.
such systems, features of items are compared to those of
the preferred items of the user. The more relevant items, In recent years, CF has gained research attention in the
measured by features similarity, are recommended. tourism domain for several reasons. By observing and
Knowledge-based recommender systems make use of predicting user preferences, CF prevents users from
knowledge about users and products to generate being overwhelmed by vast amount of travel
recommendations. They use a reasoning process to information. This is particularly valuable in an
determine what products meet a user’s requirements. information-intensive domain like tourism, which
Social-filtering or CF based recommender systems comprises a large variety of products such as
provide personalized recommendations according to accommodations, attractions, restaurants, and many
user preferences. Such systems maintain data about others. Besides, one characteristic of the purchase of
users’ purchasing habits or interests and from there tourism products is that at the moment of decision-
identify groups of similar users. For a target user, making, only information about the product, not the
known as the active user, items liked by other similar product itself, is available to users [26][33]. Current
users are recommended. web-based tourist information portals can present
archives of pre-defined destinations, some with rich
In addition to providing personalized recommendations, multimedia content. While tourism product suppliers are
CF offers a number of advantages over the other two always in favor of their own businesses, other visitors’
techniques, making it one of the most promising experiences would probably provide more valuable and
solutions to information overload. As no content unbiased information about a destination. CF techniques
information is considered in the recommendation would help users obtain such information according to
process, items being filtered need not be amenable to their travel preferences [4].
parsing by a computer. Besides, recommendations
generated are based on tastes of users rather than more Yet, CF is not readily applicable to the recommendation
objective properties of the items themselves. The items of tourism products [4][28], which are more diversified

For example. ski in December. Lastly. If his/she wishes to three major steps. Besides.g. Lastly. To recommend tourism products. the part-of or component hierarchy may also exist if they are categorized according to their The underlying philosophy of CF is that each individual location (e. belongs to a larger group of like-minded individuals. In Section item or a list of recommended items. known as Top-N 2. Another special example of component hierarchy in CD-based systems therefore make personalized tourism products is tour package. techniques such as [14] and [30] already allow content information to be used in the recommendation process. From there some future caused by the complexity of tourism products.g. a visit to system.2 Recommending Tourism Products Related work in tourism is also described. similar users [7][20][27][31]. They include restaurants. namely memory-based and model-based as each item is uniquely identified (by a product code. a study was carried out to recommender system. For example. if a user is looking for a place to K-nearest neighbor (k-nn). In Section 3. tourism products have strong inter-item This paper details the results of the study. Context awareness is therefore objective to bridge the gap between CF and tourism another important factor in designing a travel products recommendation. the movie Shrek may be suggested Surfer’s Paradise is included in a 2-day tour to Gold to a user Sam. and the research directions are identified. recommend some places in Europe due to the availability of the desired activity at different locations At the similarity weighting stage. that of another user Bob. and many others. and the To more specifically describe various CF algorithms. and recommend products using statistical However. [2]. each user in the . 1. as well as tourist spots (e. model-based out irrelevant activities such as hiking. They are mainly recommend tourism products. several CF applications are introduced. which comprises recommendations to users based on the interests of multiple heterogeneous products such as flight tickets. was ski in June. different types of CF techniques are described and recommendation.3 Organization implicitly such as in purchase history. products can have different representations. after noticing that his taste is similar to Coast). Such classification mainly reflects how products are This. if a user is looking for a place to products using probabilistic methods. hotels. as long major types. CF algorithms may compute a predicted score (rating) of a particular The rest of this paper is organized as follows. Recommending these CF algorithms were traditionally classified into two products with pure CF techniques can be easy. With the and points in time.446 The Fourth International Conference on Electronic Business (ICEB2004) / Beijing and complex than products like books. and suggests relationships. who has liked Shrek. different product Users’ preferences provide the basis for making taxonomies should be taken into consideration. techniques they used to make recommendations. or may be purchased alone as a single unit of product. user database. tourism products are highly heterogeneous even 2. or 1.g. Section 5 special taxonomies they can have. identify the problems in recommending tourism products. In a movie recommender accommodation. In Section 4. a travel recommender system must be able to filter models can be constructed offline. however. can be more difficult in recommending recommended in real-time – either based on the entire tourism products due to product heterogeneity since database or models constructed and stored in memory. and to review the state-of-the-art in both areas. Surfer’s paradise is part of Gold Coast). It works in Australia as a suitable destination. the is-a travel recommender systems. theme parks. Memory-based algorithms operate over the entire for example) and no content information is considered. in filtering of content information may be based on this paper they are classified by the underlying different features for different products. these make less sense for tourism products as methods. Surfer’s Paradise is a 1. however. Since such ski.1 K-Nearest Neighbor (k-nn) essential context information such as availability and location. Furthermore. Model-based algorithms construct compact content information may also be essential in the filtering models from the user database. a travel advisory system may consider commonly used in early CF-based systems. Firstly. hierarchy may exist if products are categorized according to their type (e. concludes this paper. their strengths and weaknesses are commented on. CF TECHNIQUES at a single destination. For example. These products may exist in multiple packages.1 Collaborative Filtering (CF) beach). it will be more suitable to neighbor selection and prediction computation. collaborative recommendations. tourism products may also contain 2. also known as pure CF. for example. for the active user. Some existing algorithms usually improve online performance. and recommend process. They may also be arranged into different future research directions to work towards collaborative hierarchies or taxonomies. some CF techniques and related work introduced are Recommending tourism products poses additional further discussed with respect to their ability to challenges to existing CF researches. Such preferences can be expressed explicitly by means of ratings. namely similarity weighting of users.

Eigentaste [8]. resulted in improved domains where textual descriptions of products are not online performance over k-nn. As the number of items and their for each prediction. empirical 2. a number of k nearest neighbors. Then. data sparseness implies that given a very users are required to rate a gauge set of items. some partial information of the active user. This. Since the active user is only For the above reasons. has revealed a is improved. In real-time. that they have access to (that is. To do so. Among which the produces less personal recommendations. in which recommendations are produced by the similarity weighting stage. predictions are many groups (clusters) according to their similarity [2]. are then used to make predictions for the items. recommendations are resulted in poor prediction quality [29]. not meaningful. or Users can therefore be well correlated to preserve the first-rater problem. Since Furthermore. It can therefore be applied to has to be inspected in real-time. obtained by querying the pre-computed lookup tables. of predict the likeliness of items. score of an item is computed. the large database of items. as compared to the degrades when the numbers of users and items grow in number of users and ratings. [29] explored the item-based clustering applications. clusters users by their similarities reflected in the ratings Data sparseness arises as users can only rate the items. especially if the active user has rated a The Bayesian cluster model divides the user base into significant number of items. This might be due to the fact Some researches regarded CF as a classification that data sparseness was not addressed in the research. In k-nn. such as Bayesian similarity is reflected in the ratings they have given on network [2]. possible that the set of common items rated by the Lookup tables of recommendations for each cluster are active user and his neighbors is very small. Performance of the algorithms relationships are relatively static. as it requires real-time finding items that are similar to those the active user has computations to be performed over the entire database liked previously. k-nn was commonly used in compared to a controlled number of clusters. most essential ones are data sparseness.4 Association Rule Mining user would like or dislike it. user ratings are . This of their mean ratings. degraded due to its probabilistic nature. For two users to be comparable. user to the cluster containing the most similar users. both users have rated are counted. even if the built by sorting the non-gauge items in descending order selected neighbors are already the “nearest” ones. for each non-gauge item in each cluster is computed they cannot rate any unobserved items).1 User-based Clustering analyses [2][12] proved that it produces good predictions. there may not be a complete set ratings matrixes representing their preferences are dense.2 Classification prediction quality of this approach only shows slight improvement over k-nn. they have given on a gauge set of items. The Fourth International Conference on Electronic Business (ICEB2004) / Beijing 447 database is assigned a similarity weight using some first discretized. recommendations can be the system. K-nn has severe performance bottleneck in approach. conditioned on whether the active user are selected as predictors for items. it is based on the number of users who have rated that item. available.Like and measures (such as those analyzed in [2] and [27]). another user-based clustering technique. Besides. Similar to the classification technique. Despite its simplicity. cold start problem and scalability. Prediction quality. decision trees induction is done K-nn makes predictions that are ratings-based and offline. some other items have been liked previously. Please refer to [11] for details. who A Bayesian network makes use of decision trees to are the users having the highest similarity weights. The cold-start problem. categorized by any attribute. new data points can be easily added to the application. however. At the neighbor selection stage. scalability recommender systems. it number of challenges for researchers. the mean or part of the items. prediction it is not described here. Only the decision tree built for the target item content-independent. only items that active user.3. Then. of ratings across all items. 2. problem [1][21]. 2. Classification techniques.2 Item-based Clustering Scalability is an important issue especially in real-time Sarwar et al. However. a certain item can be made for the active user as no other users have rated it before [16]. Since Based on the interests of the selected neighbors and decision tree induction is beyond the scope of this paper. Furthermore. usually into two classes . it uses the preferences of users in that cluster to generate recommendations.3 Clustering can be applied as long as a database of users preferences on items is available. classification techniques predict whether the active 2. In this technique. arises when no prediction about prediction accuracy. This simple online process ensures scalability. Instead of predicting the rating of an item. As no pre-compilation of data is required. and 2. computed offline. Such Dislike. it uses statistical approaches that are easy to implement. generated by performing real-time computations over The recommendation process first assigns the active the database. or cannot be easily however. This avoids the performance bottleneck of k-nn and improves scalability.3.

predictive value for multiple clusters therefore could not Therefore. some to produce recommendations using the item-based preferences will be given to items that belong to Cj. it does not provide any ARM is usually used for market basket analysis. GroupLens [27][16] is a pioneering and well- consideration of such product taxonomy allows known project in automated CF. Amazon. In addition to between items by finding items frequently appeared the traditional similarity weighting measures. accuracy.448 The Fourth International Conference on Electronic Business (ICEB2004) / Beijing Association rule mining (ARM) is a data mining Personality Diagnosis (PD) [25] combined k-nn and technique for discovering interesting relationships clustering to improve prediction quality. CF has been proven successful in both research and When applied in CF.5 Hybrid Techniques [30]. Their technique uses liked item A previously. In A ⇒ B [20%. A major concern of MAR is no cold-start items. [30] combined information filtering (IF) might lead to either too many or too few rules due to and CF to provide better recommendations. An association rule a probabilistic interpretation of results to provide better is denoted as A ⇒ B. It can advantage over k-nn. such single cluster. and measured by two measures predictions. movies are clustered by their genre. however. etc. and the active user has clustering to improve scalability. Filtering agents were later integrated into the system to improve prediction quality 2. A few examples selected from different items that are not covered by the traditional. the minimum support and confidence have to be specified in advance. CF APPLICATIONS AND RELATED WORK therefore be used at different levels. the rule simplicity and adding new data points with ease. or both. [18]. Different minimum supports should 3. and its effectiveness was proved in [9]. It makes use of purchase histories of customers among Ci and another category Cj exists. in its prediction accuracy could not be Lin et al. They variations in user tastes and items’ popularities. For example if A ⇒ B is O’Conner & Herlocker [23] combined k-nn and interesting in a certain system. Items that might have significant algorithms mine rules for all items in the database. PD retains the advantages of k-nn such as known as support and confidence. It had used k-nn to recommending cold-start or less popular products. it employs together in transactional databases. tailored for CF. as its real-time prediction process only considers 2. Predictions are then if its support and confidence values are higher than the computed independently in each cluster. MAR can be applied on items that are other items in the same cluster. that higher support is likely to exist at higher level of abstraction. its prediction score is mining (MAR) to address data sparseness and the cold computed as a weighted sum of the ratings given to start problem. prediction quality and solve the cold start problem [17]. 90%] means that 90% of users who terms of accuracy.1 Adaptive Association Rule Mining items in a single cluster. integrated filtering agents. The technique. as compared to using a dataset with level of abstraction [10]. he/she will probably like item existing data partitioning and clustering techniques to B also. Firstly. In ICHM. Experimental user-specified minimum. many rules mined will not be relevant for a reflect real user preferences. single-level domains are introduced in this section. that evaluate and rate items using syntactic features to To address these issues. and automatically adjusts the minimum The Item-based Clustering Hybrid Method (ICHM) also support value to mine a user-specified number of rules.com [19] technique. In terms of time and space performance. and 20% of all including k-nn and the Bayesian network model users purchased both of them [11]. known as AR-CRS. director. the category Ci to which a is one of the most successful and well-known online preferred item i belongs is found. Its key feature is to mine rules at different reduced accuracy. actor.4. To do so. recommend Usenet articles. which Sarwar et al. Note that a rule is considered as interesting only cluster items based on user ratings. it outperforms several approaches purchased item A also purchased item B. items. This improves prediction support rule mining algorithm. given user. For example. Kim & Kim [15] applied multi-level association rules For a new item with no ratings.4. also be used to generate recommendations after interesting rules are minded.2 Multi-level Association Rule Mining information. AR-CRS mines rules for one target item at a time. items are clustered based on content 2. Secondly. Consistent improvement. MovieLens is another recommender system developed by the GroupLens team. results show that their technique is more scalable than k-nn. Experimental results organized in a hierarchical category structure. It recommends . such as show that this approach improves prediction accuracy. A possible reason for this is that their ARM algorithms are inefficient for collaborative algorithm restricts items to being exclusively in one recommendation for two reasons. known as filterbots. however. although it does not address data sparseness. If any association rule retailers. the classification of goods in a department store (is-a It is also able to recommend cold-start items with hierarchy). In a movie recommender. MAR can compute preferences for practice. provide a dense ratings set. integrated content information into CF to improve It can compute associations among users. pointed out that traditional observed. for example. they suggested an adaptive.

Pre- defined recommendations are then retrieved from the As aforementioned. is actually a pool of travel products and activities their work can provide reference for modeling different recommended by domain experts with respect to tourism products. although has been proven successful in maintains a knowledge base of destinations. Reuse. travel products and activities based on the Case-Based flexible and interoperable recommendation engines. and ranked according to methods has revealed some fundamental challenges the preferences of the active user. the system paradigm. purely CF algorithms may not be sufficient for travel retain the cases by updating the original ones. it is important to develop a data model for their Some related applications in the tourism domain were representation. Existing solutions. Although numerous approaches have been developed. a new case containing the 4.g. such which are previously adopted solutions and/or possible as hotels and restaurants. been experienced or rated by a user. When the number of registered users TripMatcher technology does not rely on CBR. known as TourML to represent objects and events. and universities participating in the Retrieve. developed various domains. single and independent units. based on the solutions modified by users in the Reuse Due to the presence of content and context information. only ratings/purchase data Dietorecs [4] and NutKing [22] are two similar and the unique identifiers of products are used in most applications developed by the eCommerce and Tourism algorithms. problems. or modified by the active user by systems. The retrieved cases concerning the scalability and prediction quality of CF can be accepted as-is. the Revise and Retain stages allow domain experts to. tourism products also studied. duration of travel). They recommend facilitate the development of more sophisticated. inter-related and may based technologies. which is a problem solving technique that solves the current problem by The Open Tourism Consortium (OTC) [24] is a adapting solutions for previously solved. To start generating recommendations. For each domain which poses additional challenges namely destination. their correlation is lost. it Although OTC is not doing any CF-related research. This is known as the In CBR. the widespread use of pure CF case base (the Retrieve stage). As mentioned before. management of the case base is an important user-based non-transitive association. solutions to new problems. The CBR cycle consists of 4 stages. The major concern problem. recommended products according to user preferences. source software to support tourism. CF is only used to rank the contain essential context and content information. therefore combined different . context awareness and the activities during different times of the year. (the Reuse stage). including association problem if two similar items have never the number of cases and their content. if two users recommendations. it was found that the processing a similar problem as it produce recommendations from a time of the system increased linearly until it crashed. falls short when applied in the tourism by professional researchers worldwide. Reasoning (CBR) technology. it faces increased. Although the recommendations. knowledge based of pre-defined trips. In Dietorecs and NutKing. such as them as new cases. are heterogeneous. Due to the complexity of tourism products. namely individuals. some simple. In fact. When it was first how these important steps have been addressed in developed. In contrast. however. In the CBR cycle. existing CF algorithms are all based on very simple data models. FUTURE RESEARCH DIRECTIONS active user’s profile and travel settings is created. However. the dominant paradigm for performing CF has been Ski Europe [32] recommends places for skiing in based on techniques regarding items in the domain as Europe based on the TripMatcher technology [4]. complex. In user-based methods. Jester to help maintain the quality of the case base. travel experts have ranked different product heterogeneity. products’ content Another problem in this paradigm is non-transitive information and users’ ratings to generate association [14]. CF researches have been focusing adding or removing elements in the recommended cases on developing more scalable and accurate algorithms. Recently. Thus they are very important steps TripMatcher [4]. revise the cases originally stored in the case base. However. Since then. using issue as it directly affects the correctness and quality of item-similarity instead of user similarity avoids this the solutions produced by a system. is unclear. have both experienced or rated similar items but not the same ones. Integrating data models in CF would Research Laboratory (eCTRL) [6]. government agencies. similar consortium of companies. TripMatcher existence of product taxonomies to CF. they developed an XML-based data exchange language A case base refers to a repository of representative cases. stage. Revise and Retain. it had used k-nn to generate Dietorecs and NutKing. but results in the item-based non-transitive here is the selection of representative cases. applies information filtering and text mining algorithms to such context information. of interest to tourists. The constant time algorithm Eigentaste was therefore developed and adopted. Detailed development of publicly available standards and open descriptions of each stage can be found in [34]. or store recommender systems. [8] is an online joke recommender. This Similar to DieTorecs and NutKing. different travel settings (e. They are mainly based on knowledge. The Fourth International Conference on Electronic Business (ICEB2004) / Beijing 449 movies to users using GroupLens’ technologies.

. H. because item similarities are already heterogeneous and complex. pp. W. and suggests future property which affects the generation and pruning of research directions to work towards collaborative travel associations rules. Applying this technique to the Jester dataset technique [29] behaves differently in the technical [13]. but should be aware of in future [5] Fesenmaier. its Information Systems. D. liked by the user. Heckerman. a In fact in most CF algorithms. “Knowledge-based recommender systems”. Oki. 232-241. 2002 hierarchies is unknown. J. Hirsh. They do not take into consideration are negative with an average -5. Whether this applies to other product systems”. Using recommendation process. However. 714- recommends products of a certain category.. E. In classical ratings-based CF... Consider a case where all ratings given by a user recommendations.. observed. minimum support value according to the desired pp. 1998 research on this is required. 2003 number of resulting rules. [3] Burke. D. 2000 The major concern in MAR has been the determination [4] Delgado. Schaumlechner. New challenges and shortcomings should also be addressed. ratings should first be transformed from a aspect. In other words. existing CF techniques fall short when This may lessen the problems of non-transitive recommending tourism products which are association in CF. Zanellai. In A. review the state-of-the-art in both areas. However. C. the algorithm in [15] National Conference on Artificial Intelligence. With the objective to implied in the hierarchy itself. “Recommendation infrequent. as classification: using social and content-based all of its parent items must also be frequent. a inter-item relationships and product taxonomies. however. Due to the information in recommendation”. In recent years. 43-52. MAR applied to the is-a hierarchy may not be problems in recommending tourism products. bridge the gap between CF and tourism products recommendation. York: Marcel Dekker. of a higher-level item must be equal to the sum of that of its child items. The adaptive minimum support method Wöber. F.. it does that it is liked by that user. This actually implies a loss of treating items as single units in a recommender system information in the user’s likeliness of items. In [15]. Proceeding of ENTER 2002. rules” based on items’ frequencies of appearance in Such technology has already addressed product transactions in a database. Considering other analysis of predictive algorithms for collaborative taxonomies in the recommendation process. R. R.. ARM/MAR requires transforming ratings data into class labels such as like and dislike.it A shortcoming of applying ARM/MAR in CF was [7] Goldberg. According to [15]. researches. “Knowledge bases and of minimum supports at different level of hierarchies user profiling in travel and hospitality recommender (is-a hierarchy). D. ratings can be positive or negative. Terry. K. In other words. the heterogeneity and context awareness using information appearance of an item in a user’s transaction implies retrieval and text mining algorithms. [6] eCommerce and Tourism Research Laboratory: http://ectrl. Obviously. filtering”. The item-based otherwise. not take into consideration product taxonomies in the however. Proceeding of the 15th presence of this property. Proceeding of ENTER 2003. In contrast. the support count recommender systems. product taxonomies in the recommendation process. Kent (ed. D. Proceeding of the 14th Conference on have not yet been studied in the literature of CF. have not CF has been a promising solution to the problem of yet been considered in travel recommender systems. pp. all of item child items must also be [1] Basu. if some 720.. This reveals a potential research problem. for example. as it can adjust the multiple decision styles”.. Further Uncertainty in Artificial Intelligence... CONCLUSION the is-a hierarchy to improve recommendations [15]. may be a possible solution to this problem. recommendations are certain rating given by a user is converted to 1 if it is generated based on the relationship between some greater than the average rating given by that user.. Ricci.. “Dietores: travel advisory for proposed in [18] may be considered.. Davidson. pp. 1-16. MAR has already shown that it is possible to integrate 5. which joke with rating -1 will also be considered as being are likely to be found in tourism products. a study was carried out to identify the Yet. or 0 attributes of items and users’ profiles. if a higher-level item REFERENCES is infrequent (with respect to the user-specified minimum support). S. information overload in the past decade. Cohen. C. R. Nichols. and other possible hierarchies. Encyclopedia of Library and To apply MAR in travel recommender systems. 1998 associations between that category and the category of a [2] Breese. This paper The is-a hierarchy exercises the upward closure details the results of the study. Such hierarchy.itc.. which is to it began to gain research attention in the tourism domain extend the existing techniques to incorporate different for its ability to produce personalized recommendations. Fuzzy logic is inappropriate in the tourism domain. if a lower-level item is frequent. and to readily extendable to other hierarchies such as part-of. Kadie. B.. Vol. “Empirical previously liked item exist. but still it is based on the relationship between continuous rating scale (-10 to 10) to like (1) and dislike items and users identified earlier to generate (0). Traditional ARM/MAR generate “interesting “Using collaborative filtering to weave an information . D. J. Supplement 32. C.). In such hierarchy.. 69.450 The Fourth International Conference on Electronic Business (ICEB2004) / Beijing techniques with CF to form a recommendation engine.

P. Perkins... Konstan.. lying Case-Based Reasoning: recommendations: item-to-item collaborative filtering”. S. 175-186. 2002 IEEE/WIC Intl... Borchers. pp. Conference on Web Intelligence. 1998 IEEE/WIC Intl.. Miller. Information on Uncertainty in Artificial Intelligence. I. pp. Technology and Competitive http://www. J. N. Riedl. J. M.. Riedl. “Collaborative filtering [12]Herlocker. Maes. Data Mining and Knowledge [33]Stabb. Proceeding of the 16th Conference collaborative filtering algorithms”. Proceeding of WWW10. 11(5): 798-804. collaborative recommender system addressing three Riedl. M... Pazzani. S. 2000 [25]Pennock. C. Roeder. Proceeding of ACM CHI ’95 [8] Goldberg.. 2003 Publishers. 2001 [21]Miyahara. “Efficient Proceeding of CHI’95. C. filtering with the simple Bayesian classifier”. J. Konstan. “Clustering app. Li. 2002 tourism”. Tourism.. 6(1): 83-105. 17 (6): 55-57. Conference on collaborative filtering with personal agents for better Artificial Intelligence.. 76-80. Communications of the ACM. Information Retrieval. 2001 Gordon... Suchak. “Eigentaste: a constant time collaborative filtering ACM Press.. Konstan. H. M. 1995 algorithm”... 285-295. 2002 [19]Linden. “Travel recommender systems”.. collaborative filtering of netnews”.com/ recommender systems”. “Collaborative [9] Good.. “Amazon.. J. “GroupLens: an open architecture for problems”. algorithms”. B.. Kim. U. F. J... Werther. Kim.. “Combining Proceeding of the 6th Pacific Rim Intl. Kim. Riedl. M.. W. Borchers. [18]Lin. A.. N.. J. “Intelligent systems for Discovery. C. 473-480. Fu.. D.berkeley.. 1999 Engineering. J. Herlocker..it:8080/dev/jsp/index. CAB Intl. pp.. Herlocker. pp. A. “Using filtering agents to 77-87. 439-446.. Conference on Web Intelligence [31]Shardanand.com [34]Watson. Proceeding of hybrid recommender system”. 2004 CSCW ’94. A.. Herlocker. L.M. “Mining multiple-level association collaborative filtering”. E. A. 2000 recommendations”. pp. Proceeding of the 16th National [22]NutKing: http://itr.ieor. Gupta D. J. M.. IEEE. 35(12): 61-70. Bergstrom.. Proceeding of the 8th Pacific Rim Intl.. Conference on Human Factors in Computing Systems. pp. roach for collaborative filtering system”. B. 33-38.. Ehrlich. Sarwar.. IEEE Intelligent Systems. roach”. Kim. The Fourth International Conference on Electronic Business (ICEB2004) / Beijing 451 tapestry”.. 2003 filtering: algorithms for automating ‘word of mouth’”. Smith. J. J. pp. 1999 [24]Open Tourism Consortium Knowledge Repository: [11]Han.and model- analysis of design choices in neighborhood-based based app. IEEE using multi-level association rules”..edu/~goldberg/jester-data/ Strategies.jsp Conference on Artificial Intelligence. Knowledge and Data Workshop..K. U. pp. pp. 524-527.opentourism.. to Usenet news”. M. J. Techniques for Enterprise Systems... 5. Riedl. 4(2): 133-151. Miller B. Schafer.. Communications of the ACM. B. D..M. J.. “Pointing the way: active 1992 collaborative filtering”. D. B. Karypis... Herlocker J. J. 2003 “Item-based collaborative filtering recommendation [16]Konstan J. [20]Maltz. Morgan Kaurmann Publishers.. 2002 2000 [13]Jester Data Download Page: [26]Poon. Kamber. 210-217. 287-310. pp. ort association rule mining for [32]Ski Europe: http://www. J. “Social information (WI’03). Alvarez. Retrieval.. Proceeding of the CSCW ’98. J.. Data Mining: Concepts and http://www. 679-689. Recommender Systems rules in large databases”. 1993 [14]Kim. 17(6): 53-55. J. 1995 adaptive-supp. 345-354. Horvitz. “Applying collaborative filtering [30]Sarwar. Maltz. [29]Sarwar. Y. Inc. Riedl. K. Proceeding of Conference on Artificial Intelligence. “Clustering items for [10]Han... K. B. K.phtml Techniques. Morgan Kaufmann Industry Report. G. J. 202-209. York. W. B.....ski-europe. G. 40(3): J. 1999 [23]O'Connor. J.. Wallingford. “An empirical by personality diagnosis: a hybrid memory. Ruiz.. Q.. “A new [27]Resnick.M. Iacovou. B. 1997 ..org/wiki/wiki. Proceeding of the Intelligent Systems. 1997 improve prediction quality in the GroupLens Research [17]Li.. B.. T. J. P. 1994 [15]Kim.. Q. Konstan. “A recommendation algorithm [28]Ricci.. J.itc. App.. P. pp. J.