A semantic complement to enhance electronic market

Expert Systems with Applications 36 (2009) 5768–5774 Contents lists available at ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa A semantic complement to enhance electronic market Sharmin Moosavi *, Mohammadali Nematbakhsh, Hadi Khosravi Farsani Computer Engineering Department, University of Isfahan, Hezarjerib.st, Isfahan, Iran a r t i c l e i n f o Keywords: Complementary product Ontology Similarity Recommender system . a b s t r a c t Consumers prefer to purchase bundled and related products to use them together to perform a task or satisfy a need. In this paper, we propose a complementary association for bundling products to enhance promotions, recommendations and selling strategies in marketplaces such as combinatorial auction. We propose an ontology based model and define a Need association to determine complement of product classes. Using this type of association, we develop a mathematical model to relatively measure complementary degree of classes and the latest purchased products to recommend Top-N products. We experiment this approach with a recommender system utilizing complementary products. Experimental results on the dataset of Building Equipment Company show superiority in terms of performance and precision. Ó 2008 Elsevier Ltd. All rights reserved. 1. Introduction Today, most companies use their own critical data to extract valuable information. The information of product associations and their measurement degree play an important role to improve selling strategy, investment, marketing, management policies and product catalog (Khosravi Farsani & Nematbakhsh, 2007; Webb, 2003). Association rule using data mining techniques shows the purchasing association that customers generate on their purchasing behavior and determines the products that are usually accompany each other in each purchase transaction. There have been several reports about drawbacks to this type of association, mainly low precision and performance (Webb, 2003). Another type of product association is substitution that has been given especial attention in the last few years. Substitution is a kind of relation that represents the degree of similarity between items (Resnik, 1999). There have been several techniques to the corresponding degree of similarity and are mostly based on finding well-known properties of the items (Ganjifar, Abolhasani, et al., 2006; Knappe, 2005; Resnik, 1999). Complement is another kind of association that has been defined in economics and has many applications (Economides & Brian, 2006; Schmidtke, 2006). For example, Recommending a similar car or bicycle to a buyer of a Mercedes car is not very useful; but recommending an especial media player or burgeler-protection is more useful. As another example, complement products bring additional marketing power for a company. It allows vendor lock-in as it increases the switching cost and when a seller decides to raise the capacity of sale, he advertises the complementary products. Complements-in-consumption is a type of complement that we define in this paper. It is described as a product that should be consumed with another and the customer’s satisfaction of consuming both goods is greater than when they are consumed separately. This means that, if items A and B are complementary, the more an item is being consumed, the more B item will be consumed. An example of complements-in-consumption is burger and hamburger buns. Because the two are usually used together, increasing the volume of burger makes more hamburger buns to be sold. Car and gasoline are another example of complementary items. In this paper, we refer to complements-in-consumption as complementary attribute and we use ontology technique to determine the complement products. We model product classes in OWL ontology language and define an attribute called Need for each of the determined classes. Our method uses this attribute to compute the degree of complementary to one product. The remainder of the paper is as follows: degree of complementary between two products will be presented in Section 2. Our suggested association will be implemented in a new recommender system in Section 3. The suggested recommender system will be compared by association rule mining. Other usage of complementary measure will be presented in Section 4. Finally, Section 5 will be devoted to conclusions derived from this study. 2. Complementary procedure * Corresponding author. Fax: +98 3117932670. E-mail addresses: [email protected] (S. Moosavi), nematbakhsh @eng.ui.ac.ir (M. Nematbakhsh), [email protected] (H.K. Farsani). 0957-4174/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2008.06.137 In this paper, we suggest a procedure to find complementariness of products. The proposed procedure consists of three phases 5769 S. Moosavi et al. / Expert Systems with Applications 36 (2009) 5768–5774 as shown in Fig. 1. The procedure should be initiated by designing product catalog of the electronic marketplace. Attributes of the products and value of them have been determined in phase 2. In phase 3, we determine complementary degree of each two pair of classes. This degree is computed using ontology techniques as you will see in the future sections. Designing catalog of the marketplace is the initial phase of the complementary procedure. This phase is very important from two points, from both user and system points. Users can observe their interesting products easily since products have been classified in the catalog and the system can easily manage its products in the groups. Our complementary procedure uses the catalog to determine complementary degree between products. We suggest catalog UNSPSC (2001) structure for product classification. The classes and relationships between them have been modeled by OWL Ontology tags. Each class of the catalog is represented by OWL:Class tag of corresponding ontology language. Classes of the products may be divided into several subclasses. This property of the catalog has been implemented by OWL:Subclass tag of this ontology language. Moreover, all of the classes have a name attribute which represents name of the product class. A part of the catalog which belongs to transportation system has been shown in Fig. 2. Classes of the catalog have a need property which is very important for our complementary procedure. This property exists in each class and is being modeled using attribute definition of OWL ontology language. It may have several values which represent the name of the other classes that are required when this product is being consumed. For example, a car product requires gasoline when the customer uses car. So, need property of a car will include gasoline product. As another example, consider to the pen and pencil. When the customer uses pencil, he/she requires paper and eraser. So, need property of pencil will include paper and eraser (Fig. 4). Classification hierarchy is built using IS-A property of each class. This property is being modeled using OWL subClass of tag. When class X is sub-class of class Y, X inherits all of the properties of Y. In other words, IS-A property contains the name of the super class, and in our classification system it has only one value. For example, IS-A property of car is Passenger motor vehicles. From among the product attributes, Needs and IS-A are important. Products in our catalog have other properties, such as Consumption interval, Wordnet, UNSPC, ECL@SS. Consumption interval is another attribute of each class of the product catalog. 1. Designing Product Catalog 2. Determining attributes of each class in the catalog 3. Determining complementary of two classes WordNet Needs IS-A Classi ECL@SS Con.Interval UNSPSC Fig. 3. Attributes of each class. This property has a single time value that shows the maximum time of consuming a product. WordNet attribute is used to distinguish purpose of the class. UNSPSC and ECL@SS are used for matching the classes of these classification system and standard classification systems (Khosravi Farsani & Nematbakhsh, 2007). Based on previous discussion, each class of the classification system has Needs, IS-A, Con.Interval, etc. properties (Fig. 3). 2.1. Determining complementary degree of two classes If Needs property of class A compromise class B, these two classes have a high complementary degree. For example, complementary degree between pencil and eraser is very high since eraser exists in the Needs property of the pencil. If Needs property of class A do not compromise class B, our algorithm applies the above procedure with A and super class of B. It means that our algorithm examines whether super class of B exists in Needs property of A or not? This procedure is repeated again while an ancestor of class B is found in Needs property of class A. In each step of procedure implementation, the complementary degree between A and B is reduced. Our algorithm uses IS-A Property of class B, if class B does not exist in the need property of class A. Suppose, we want to compute complementary degree between pencil and graph paper. As you can see in Figs. 4 and 5), pencil requires writing paper, pencil sharpener, and eraser. Graph paper is not a member of pencil needs, but its super class exists in the pencil needs. Therefore, graph paper is complementary of pencil indirectly. More formally, the complementary products problem can be formulated as follows: Let A.N is the set of all Needs of class A and A.H is the super class of A. Com(A,B) is a complementary function that measures the complementary degree of item A rather than item B. this function is defined as follows: 8 1 > < k  comðA; B:HÞ B R A:N comðA; BÞ ¼ 1 B 2 A:N > : 0 B:H ¼ / ð1Þ In this function, k is a predefined factor larger than 1 and indicates decreasing rate of complemen when moving along the classification system. In this paper, we assume k is equal to 2. We illustrate this function using the following two examples: Example 1. Complementary degree of Pencil and Eraser according to Figs. 4 and 5 is calculated as follows: Fig. 1. Procedure of determining complementary of two products. Commercial and Military and Private Vehicles and their Accessories and Components Passenger motor vehicles Automobiles or cars Busses Motor vehicles Definition 44121702 WordNet ECL@SS Pencil UNSPSC Sports car Fig. 2. Parts of the classes of transportation system in the catalog. 44121702 Eraser, Pencil sharpener, Writing paper Needs IS-A Con.Interval One month Fig. 4. Properties of pencil. Writing instruments 5770 S. Moosavi et al. / Expert Systems with Applications 36 (2009) 5768–5774 Office Equipment and Accessories and Supplies Office supplies Writing instruments Desk supplies Correction media Erasers Manual pencil sharpener Pen or pencil sets Fig. 5. Parts of the classes of office equipment in the catalog. Pencil.N = {Eraser, Pencil sharpener, printing and Writing paper} Eraser.H = correction media Eraser e Pencil.N therefore Com(Pencil, Eraser) = 1 Example 2. Complementary degree between Graph paper and Pencil according to Figs. 4 and 6 is calculated as follows: Pencil.N = {Eraser, Pencil sharpener, printing and Writing paper} Graph paper.H = printing and writing paper Graph paper R Pencil.N therefore Com(Pencil, Graph paper) = 1/2 * Com (Pencil, Printing and writing paper) Printing and writing paper e Pencil.N therefore Com(Pencil, Printing and writing paper) = 1 therefore Com(Pencil, Graph paper) = 1/2 Complementary products have transitive property which means that if B exists in A’s Needs and C exists in B’s Needs, then C is a complementary of A. For example, pencil sharpener requires a pencil and pencil requires an eraser, thus eraser is complementary of pencil sharpener. Complementary degree of pencil and pencil sharpener is higher than complementary degree between eraser and pencil sharpener. Our method uses transitive property, if complementary degree between class A and class B is based on expression (1), and it is computed to zero. Complementary degree between two products using transitive property is calculated as follows: 8 1 > < k0  comðA:N; BÞ B R A:N comðA; BÞ ¼ 1 B R A:N > : 0 B:H ¼ / ð2Þ In this function, k0 is a predefined factor larger than 1 and indicates decreasing rate of complementation along with the transitive sequence (we value of k0 set to 2).We illustrate this function using the following two examples: Example 3. According to Fig. 5, complementary degree between Pencil sharpener and Eraser can be calculated as follows: Pencil sharpener.N = {Pencil} Eraser.H = Correction media Eraser R Pencil sharpener.N therefore Com(Pencil sharpener, Eraser) = 1/2 * Com(Pencil sharpener, Correction media) Paper Materials and Products Paper products Printing and writing paper Writing paper Graph paper Notebook filler Fig. 6. parts of the classes of paper products in the catalog. Correction media R Pencil sharpener.N therefore Correction media.H = Office Equipment and Accessories and Supplies Com(Pencil sharpener, Correction media) = 1/2 * Com(Pencil sharpener, Office Equipment and Accessories and Supplies) Office Equipment and Accessories and Supplies R Pencil sharpener.N therefore Office Equipment and Accessories and Supplies.H = u therefore Com(Pencil sharpener, Office Equipment and Accessories and Supplies) = 0 therefore Com(Pencil sharpener, Correction media) = 0 therefore Com(Pencil sharpener, Eraser) = 0 therefore Eraser R Pencil sharpener.N therefore Com(Pencil sharpener, Eraser) = 1/2 * Com(Pencil, Erase) Pencil.N = {Eraser, Pencil sharpener, Writing paper} Eraser e Pencil.N therefore Com(Pencil, Eraser) = 1 therefore Com(Pencil sharpener, Eraser) = 0.5 For example, if a user purchased any kind of pencil sharpener, then he will probably purchase an eraser with complementary degree of 0.5. 3. Case study: complementary-based recommender system Recommender systems aim to tackle information overflow in the electronic commerce environment. These systems customize electronic catalog for each customer and are required to respond to several challenges such as type of the recommended products, time of recommendation, etc. In this section, we present an algorithm which considers the interval for recommendation using complementary. If the consumption interval of the previous purchased product has expired, our algorithm recommends a similar product; otherwise it finds a complementary for its previously purchased products and recommends Top-N products to the customer. Recommender systems are generally two methods to formulate recommendations both depending on the type of items to be recommended and the way that user models are constructed (Allen, 1990). The two different approaches are content based (Balabanovic & Sholam, 1997; Kalles, Papagelis, & Zaroliagis, 2003) and collaborative filtering (Herlocker, Konstan, Terveen, & Riedl, 2004; Hofmann, 2003), while additional hybrid techniques have been proposed as well (Balabanovic & Sholam, 1997). Content based algorithms are principally used when documents are to be recommended, such as web pages, publications, jokes or news. This type of recommender systems uses one of the above similarity algorithms to recommend a similar item to the user. Similar item is very close to the previous interested items by the user. Recommending a similar product to the user is not very efficient. Since users do not need two products which is the same in some of their properties. For example, Car and bicycle is very similar but recommending a car when user has bought a bicycle is not very good. Collaborative-filtering algorithms aim to identify users that have common interests and preferences by calculating similarities and dissimilarities between user profiles (Herlocker et al., 2004). The idea behind this method is that, it classifies users based on previous purchased products and recommend N product which is purchased by other members of the group but not purchased by the user. The success of these recommender systems strongly depend on classification algorithm and selection algorithm which selects top-N product of that group. In fact, these algorithms create similar group of user which has similar previous purchased products. S. Moosavi et al. / Expert Systems with Applications 36 (2009) 5768–5774 5771 Fig. 7. State diagram for the customer in the electronic marketplace. Collaborative faltering is divided in the two approaches; userbased and item-based. User-based technique is limited by sparsity1 (Huang, Chen, & Zeng, 2004), cold-start (Melville, Mooney, & Nagaragan, 2002; Schein, Popescul, & Ungar, 2002), scalability2 (Cosley, Lawrence, & Pennock, 2002) Item-based technique solved scalability problem but that still has other problems. We propose a new kind of content based algorithm which uses complementary and similarity product for recommendation. The proposed recommender system has high accuracy and solves scalability, sparsity and cold-start problems. Fixing features of the product classes is one of the drawbacks of the proposed system which can be solved by future researches. In our approach, complementary products are classified in two categories based on product recommendation interval as follows.  Complement Product Over Purchase Time (CPOPT): the user will be recommended items complementary to the ones he selects at the current session.  Complementary Product After Purchase Time (CPAPT): the user will be recommended items complementary to the ones that has been purchased previously. Our approach in CPOPT decides to recommend complementary products during the purchase. When the customer enters to the electronic marketplace, he selects some products and adds them to the shopping cart. Our recommender system always traverses the shopping cart to recommend the complementary of the selected products. For example, when the customer adds hamburger to the shopping cart, our algorithm checks it and presents complementary of hamburger for the customer in order to purchase. The proposed recommender system uses CPAPT after the customer enters to the electronic marketplace. In CPAPT algorithm, when the customer enters to the electronic marketplace, the system considers his/her purchases in previous transaction and then recommends complementary of the purchased products. But there is a problem. Recommending complement of a product to the customer may depend on the purchasing time of the main product. For example, if the customer purchased a hamburger in Wednesday, recommending hamburger buns as complement of it in the next week is not benefit since customer consumes product and complement of it jointly. In fact consumption time of purchased product has been over and in this situation our algorithm recommend similar product to the customer. 3.1. Recommendation procedure Entering of the customer to the electronic marketplace, our recommendation system gets customer’s background information. This information includes his previously purchased products. If the customer has purchased products some time ago, our algorithm determines the consumption interval for these products. If the consumption interval of these products has ended, our system recommends similar product; otherwise it recommends complementary product (CPAPT). If a customer purchases products in current time, our method recommends products complementary to the purchased products (CPOPT). State diagram for the customer in the electronic marketplace is shown in Fig. 7. In the next section, our proposed method is investigated using some evaluation tests. Our proposed algorithm also considers properties of the recommended products based on customer interesting. Suppose our algorithm decides to recommend pencil, it does another important stage to specify parameters of pencil such as color, size, manufacture, etc. Our algorithm is able to use the following two approaches for this problem:  Customer preferences: we model parameters of the recommended product as a vector (Recommended Vector). The elements of this vector are product attributes such as color, size, manufacture, etc. In addition, each purchased product is modeled by another vector (Purchase Vector). We compare these two vectors and determine cosine degree between them. Each product which has high cosine degree may be selected as a desired product. Value of attributes of the recommended product is selected from value of attributes of the desired product. For example, customer has purchased pencil that has red, large size, earaserCom for the color, size and manufacture. Suppose our algorithm has decided to recommend eraser. This customer prefers to use product of earaserCom with red color and so on.  Preferences of other customers: in this approach, we model attributes of the recommended product as a vector. Each value of its attribute is selected from preferences of other customers. Suppose our algorithm has decided to recommend eraser, it examine to find the best value for eraser parameters from those customers who has bought eraser. 4. Proof and evaluation 4.1. Dataset 1 Number of ratings already obtained is usually very small compared to the number of ratings that need to be predicted. 2 Recommender systems require computations that are very expensive and grow non-linearly with the number of users and items in a database. Therefore, in order to bring recommendation algorithms successfully on the web, and succeed in providing recommendations with acceptable delay, sophisticated data structures and advanced, scalable architectures are required. We used the data of Building Equipment Company collected from 2005/04/22 to 2006/08/01. This dataset has 2266 customers, 2581 products and 21662 sale transactions. Products of company have been classified into 18 groups and an ontology is constructed. Tile, sanitary china, cement, head stones, etc. are sample of these groups. 5772 S. Moosavi et al. / Expert Systems with Applications 36 (2009) 5768–5774 4.2. Evaluation metric CPAPT 0.44 0.42 F1 metric To evaluate our algorithm, we utilized widely used metrics in information retrieval (IR) community, namely precision and Recall (Khosravi Farsani & Nematbakhsh, 2006; Kowalski, 1997). In the first step, the algorithm works on the training portion of data where top-N set will be generated as the set of recommendations (for each customer). The main goal is to look into the test set and match the newly recommended products with previously generated top-N set. Products that appear in both sets are members of a special set, which is called the hit set. Precision and Recall metrics defined in this context as below: 0.4 0.38 0.36 0.34 0.32 0.3 10 20 30 40 50 60 70 80 90 TOP-N Fig. 8. F1 values in CPAPT algorithm.  Precision. Which is defined as the ratio of hit set size to the topN set size, i.e., precision = size of hit set/size of top-N set.  Recall. We define recall as the ratio of hit set size to the test set size, i.e., recall = size of hit set/size of test set. CPOPT 0.54  F1 ¼ 2 Recall Precision=ðRecall þ PrecisionÞ 0.52 F1 metric These two measures are often opposite in nature. For instance, increase in the number of N tends to increase Recall but decrease Precision. The fact that both are critical to the quality judgment leads to using a combination of them. In particular, this paper uses the standard F1 Metric (Khosravi Farsani & Nematbakhsh, 2006, 2007) that considers them equal in importance and is computed as 0.5 0.48 0.46 0.44 0.42 ð3Þ 10 20 30 40 50 60 70 80 90 100 130 Top-N 4.3. Experimental results In this section, we present and discuss the evaluation results for the proposed recommender system. The algorithm utilizes the above data set as input to evaluate two types of recommendation algorithm. CPAPT algorithm is the first type and the second type is CPOPT algorithm. In order to evaluate our algorithm, we compute complementary degree between each pair of products and record them in a table. Purchased products of each customer have been divided into two sets namely Training set and Test set. We have randomly selected 80% of products in to training set and 20% in to test set. We apply our algorithm to Training set and produce Top-N products which are complementary of the previous purchased product. We measure F1 for each customer and compute the average of this parameter as the final F1. This procedure has been applied 10 times to reduce the influence of the selected sets on the result. In other words, our algorithm has been executed each time with different values of Train and Test set. To evaluate CPAPT algorithm, items of each factor are divided into two sets namely Train set and Test set. Our algorithm finds complementary of Train set and produces Top-N product. These Top-N products are matched with test set and computed F1 as described above. Final results demonstrate that F1 changes from 0.36 to 0.42 (Fig. 8). To evaluate CPOPT algorithm, we classify the purchased products of each customer according to the time of purchase. We divide each class into Train and Test set. We apply our algorithm to train set and produce Top-N product. These Top-N products have been matched to the test set and have computed F1 as described above. F1 changes from 0.46 to 0.53. As you can see in Fig. 9 this type of recommendation algorithm has better results in comparison with previous algorithm. At present, a few marketplaces have the ability to recommend product when the customer browse product in the electronic market. As you can see in Fig. 9, this type of recommendation has better results in comparison with other type of recommendation. In fact, behavior of the seller in the real world has been modeled by Fig. 9. F1 values in CPOPT algorithm. CPOPT algorithm. When customer selects a product in the physical store, Seller recommends complement of the selected product and tries to convince customer to buy another product. CPOPT algorithm model the behavior of the seller in the electronic market. Association rule is a popular approach to acquire complement of one product. Suppose a customer has purchased product A in the previous and returns to the electronic marketplace for a new purchase. Based on association rule mining, we find number of the transaction which contain A and Bi (i is the number of product) and compute confidence of the rules A ? Bi (for all i). Bj is complement of A, if confidence factor of the rule A ? Bj is greater than a consensus threshold. Our algorithm recommends Bj as complemented of previous purchased product. Confidence of A ! B ¼ ðnumber of transactions contain A and BÞ=ðnumber of transactions contain AÞ ð4Þ We compare our algorithm with association rule to understand efficiency of the proposed algorithm. We divide data set in to Train set and Test set. We determine complement of all products as described in expression (4) and maintain them in a table. For each customer, we generate some recommendation, compare them with Test set and compute F1 metric as you can see in Fig. 10. Its results are weaker than our two suggested algorithms. This figure demonstrates that F1 changes from .13 to .17. An important aspect of an effective recommender system is run time to recommend Top-N products. We compare our suggested approach with association rule from run time viewpoint. Our approach computes degree of complementary in a few second, but association rule consumes several hours to compute degree of relationship between products since it takes many time to analyze market baskets. Fig. 11 compared run time of association rule method and our approach. So, complementary products recommending has a good accuracy and high speed. This approach is more efficient than association rule and solves cold-start and scalability problem. S. Moosavi et al. / Expert Systems with Applications 36 (2009) 5768–5774 5773 tary product and base product together. For example, advertising car and media player together, leads to more sell of media player. CPAPT 0.45 F1 metric 0.4 0.35 0.3 5.3. Non-excludable public product 0.25 In economics, a kind of products is public product that is a nonexcludable and non-rival in consumption product. Non-rival means that, consumption of the product by a person does not reduce the amount of the product available for consumption by others. For example, breathing air does not reduce the amount of air available to others. Non-excludable means that companies are not able to sell the products. An example of public products is Open Source Software. Today, companies such as IBM invest in Open Source Software. This poses immediately the question, why companies invest in public products. One reason is companies expect to benefit from some proprietary products that are complementary of the public product. Even though the companies cannot directly capture the benefit of a public product, they can capture benefit indirectly through selling more complementary proprietary goods at a potentially higher price. 0.2 0.15 0.1 0.05 0 10 20 30 40 50 60 70 80 90 TOP-N result of complemenary product result of association rule Fig. 10. F1 values for association rule. Time 80000 Time(second) 70000 60000 6. Conclusion 50000 40000 30000 20000 10000 0 1 6 11 16 21 26 31 36 41 46 pairs of products: 1 = (motar, sand) , 2 = (motar , plaster) , ... time of association rule time of complementary products Fig. 11. Run time comparison of our algorithm with association rule. 5. Other usage of complementary measure Complementary concept may be used in several other applications. In this section, we review some of its usage for other decision of electronic marketplace. In this paper, we proposed a new association between products namely complementary product that is very effective compared to other existing associations. We specified an attribute for each class of the catalog and called it Need property. This property contains other classes which may be consumed with this class. In order to find complement of the classes, we modeled electronic catalog using OWL ontology language and found complement of the classes based on suggested procedure. Complementary products will be used in several applications such as recommender systems, promotion systems, selling strategies, etc. In this paper, we utilized complementary product for a new recommender system. We compare our complementary-based recommender system with association rule mining and provided some results. Experimental results have shown our suggested recommender system will increase precision and performance. References 5.1. Pricing An important factor to measure quantity demanded of a typical product is the price of its complementary. Economics have defined a new metric which called the cross price elasticity of demand. This metric measure the influence of quantity demanded of a product to a change in the price of another product. For example, if price of fuel product increases 10% and the quantity of new cars decreased by 20%, then cross elasticity of demand would be 20%/10% = 2. When two products are complementary, with increase the price of base product, the demand for its complementary will be decrease. Therefore the cross price elasticity of demand is negative. In the example above, fuel and cars are complementary. This property of complementary products gives additional market power to companies. The market power is an ability of a company to change the market price of a product. 5.2. Advertising One of the best strategies for selling a product is advertising complementary products. Economics believe that this strategy has very efficiency compare to other approach. This strategy leads to customer to become more familiar with different kind of her/his required products and to be encouraged for purchase complemen- Allen, R. B. (1990). User models: Theory, method, and practice. International Journal of Man-Machine Studies, 32(5), 511–543. Balabanovic, M., & Sholam, Y. (1997). Combining content-based and collaborative recommendation. Communications of the ACM, 40(3), 66–72. Cosley, D., Lawrence, S., & Pennock, D.M. (2002). REFEREE: An open framework for practical testing of recommender systems using research index. In Proceedings of the 28th very large data bases conference (pp. 35–46). Hong Kong, China. Economides, N., & Brian, V. (2006). Pricing of complements and network effects. JELclassification code: L12;L13;C63;D42;D43. Ganjifar, Y., Abolhasani, H., et al. (2006). A similarity measure for OWL-S annotated web services, Web Intelligence, 2006, WI 2006. In IEEE/WIC/ACM international conference, Hong Kong (pp. 621–624). Herlocker, J., Konstan, J. A., Terveen, L., & Riedl, J. (2004). Evaluating collaborative filtering recommender systems. ACM Transactions on Information Systems (TOIS), 22(1), 5–53. Huang, Z., Chen, H., & Zeng, D. (2004). Applying associative retrieval techniques to alleviate the sparsity problem in collaborative filtering. ACM Transactions on Information Systems, 22(1). Hofmann, T. (2003). Collaborative filtering via Gaussian probabilistic latent semantic analysis. In Proceedings of the 26th annual international ACM SIGIR conference on research and development in information retrieval (pp. 259–266). Toronto, Canada. Kalles, D., Papagelis, A., & Zaroliagis, C. (2003). Algorithmic aspects of web intelligent systems. In N. Zhong, J. Liu, & Y. Yao (Eds.), Web intelligence. Berlin: Springer (pp. 323–345). Khosravi Farsani, H., & Nematbakhsh, M. (2007). Designing a catalog management system an ontology approach. Malaysian Journal of Computer Science, 20(1). Khosravi Farsani, H., & Nematbakhsh, M. (2006). A semantic recommendation procedure for electronic catalog. Transaction on Engineering, Computing and Technology, 6, 350–355. 5774 S. Moosavi et al. / Expert Systems with Applications 36 (2009) 5768–5774 Knappe, R. (2005). Measures of semantic similarity and relatedness for use in ontologybased information retrieval. Thesis of doctor. Denmark: Roskilde University. Kowalski, G. (1997). Information retrieval systems: Theory and implementation. Kluwer Academic Publisher. Melville, P., Mooney, R.J., & Nagaragan, R. (2002). Content-boosted collaborative filtering for improved recommendations. In Proceedings of the national conference of the American association artificial intelligence (pp. 187–192). Edmonton, Alberta, Canada. Resnik, P. (1999) Semantic Similarity in a taxonomy: An information-based measure and its application to problems of ambiguity in natural language. Journal of Artificial Intelligence, (pp. 95-130). Schein, A.I., Popescul, A., & Ungar, L.H. (2002). Methods and metrics for cold-start recommendations. In Proceedings of the 25th annual international ACM SIGIR conference on research and development in information retrieval (pp. 253–260). Tampere, Finland. Schmidtke, R. (2006). Private provision of a complementary public good. JELclassification numbers: C72, L13, L86. UNSPSC (2001). Why coding and classifying products is critical to success in electronic commerce, using the UNSPSC, White Paper, Granada Research. Webb, G.I. (2003). Preliminary investigations into statistically valid exploratory rule discovery. In Proceedings of the Australasian data mining workshop AUSDM03 (pp. 1–9). University of Technology, Sydney.