《网络购物推荐系统（英） 毕业论文.doc》由会员分享，可在线阅读，更多相关《网络购物推荐系统（英） 毕业论文.doc（11页珍藏版）》请在三一办公上搜索。
1、A Knowledge-based Recommender System for Customized Online Shopping网络购物推荐系统AbstractThe concept of personalization has long been advocated to be one of the edges to improve the stickiness of on-line stores. By enabling an on-line store with adequate knowledge about the preference characteristics of d
2、ifferent customers, it is possible to provide customized services to further raise the customer satisfaction level. In this paper, we describe in details how to implement a knowledge-based recommender system for supporting such an adaptive store. Our proposed conceptual framework is characterized by
3、 a user profiling and product characterization module, a matching engine, an intelligent gift finder, and a backend subsystem for content management. A prototype of an on-line furnishing company has been built for idea illustration. Limitations and future extensions of the proposed system are also d
4、iscussed.Keywords: On-line Shopping, Personalization, Recommender Systems, Knowledge-based Systems1 INTRODUCTIONThe development of Web technologies has brought a lot of advantages to merchants for moving their business on line. Within the past few years, a large variety of on-line stores has been st
5、arted in the cyberspace. However, the survival rate is just around 50%, where some recognized dom-com like B, K, MVP.com are included 1. We believe that one important factor determining the success of on-line stores is whether the on-line shopping experience can be enhanced to such an extent that so
6、me customers choose to and continue to shop on-line. Along this direction, the concept of personalization has long been advocated as one of the edges to improve the stickiness of on-line stores. A survey, recently conducted by Cyber Dialogue, reveals that customers are more likely to purchase from a
7、 site that allows personalization, and register at a site that allows personalization or content customization 2. To achieve that, an on-line store needs to be enabled with adequate knowledge about customers preference characteristics and use it effectively to provide personalized services with high
8、 precision. A typical example of personalized services is the use of recommender systems. Recommender systems have been implemented by many big Web retailers, such as A and CDN. Typically, they use an intelligent engine to mine the customers rating records and then create predictive user models for
9、product recommendation. Software products of recommender systems are now available from various companies like NetPerception, Andromedia, Manna, etc. Based on the underlying technology, recommender systems can be broadly categorized as: Knowledge-based 3 where user models are created explicitly via
10、a knowledge acquisition process. Content-based 4 where user models are created implicitly by applying machine learning or information retrieval techniques to user preference ratings and features extracted from product description, and Collaborative 5 where user models are created solely by utilizing
11、 overlap of user preference ratings.In the literature, there exist a lot of works on content-based and collaborative recommender systems. One of their common characteristics is that a substantial amount of good user preference ratings is required before precise recommendations can be provided. Howev
12、er, if a company is lacking such ratings information or it has new items arrived constantly, these two approaches will fail. Here we argue that before such ratings information can be collected, the knowledge-based approach should provide a good complementary solution. With a similar rationale, Ardis
13、sono et al. 6 proposed a knowledge-based system using for tailoring the interaction users using a shell called SETA for adaptive Web stores, where stereographical information is also used for user modeling. Sen et al. 7 proposed an intelligent buyer agent which aims to educate the user to be a more
14、informed customer by understanding the user query and providing alternatives using a pre-built domain-specific knowledge base, which is based on propositional logic representation. For automatic rule generation, Kim et al. 8 have built a prototype system where the decision tree induction algorithm i
15、s applied to personalize advertisements. As there is always a trade-off between personalization and privacy, what kind knowledge needed to be acquired for exchanging personalized services is definitely an important concern of on-line customers. So, the question becomes: how can the user information
16、requirement be minimized while an acceptable level of recommendation service can still be provided?. In this paper, we restrict the user information needed to only demographic information and describe in details how a related knowledge-based system can be built to support an adaptive on-line store i
17、n providing customized recommendation services. Our proposed conceptual framework is characterized by a user profiling and product characterization module, a matching engine, an intelligent gift finder, and a backend management system. A prototype of an on-line furnishing company has been built and
18、is used throughout the paper for idea illustration. The limitations and future extensions of the proposed framework will also be discussed.2 SYSTEM OVERVIEWKnowledge-based systems are characterized by the fact that its two important components, namely the knowledge base and the inference engine (som
19、etimes also called the shell in expert systems) are separated. A typical example is the rule-based system where the knowledge base is represented in the form of a set of if-then rules and forward-chaining reasoning is used in the inference engine. The knowledge engineer can keep on expanding the kno
20、wledge base by acquiring more domain knowledge with the inference engine being unchanged at all.In this project, instead of using the rule-based syntax, a feature vector-based representation is adopted. Also, we assume a conventional 2-tier architecture, where domain knowledge is stored in a relatio
21、nal database and all the functional modules of the inference engine are run on the web server. The knowledge required to be acquired and stored in the database for driving this customized on-line store include: Generic products information, e.g., product name, price, manufacturing country, etc. Prod
22、uct characteristics, e.g., degrees of reliability, design style, etc. User demographic information, e.g., sex, age, occupation, and User preference profiles, e.g. preferences on reliability, dressing style, etc.The inference engine contains the following functional modules: User profiling module whi
23、ch acquires the user demographic information via a simple questionnaire during membership registration and transform the information to create a preference profile for supporting the subsequent matching. Matching engine which computes the similarity score between user preference profiles and product
24、 characteristics to support personalized product ranking shown in the catalog or as special product recommendations. Intelligent gift finder which can assist the customer via a wizard interface to identify possible gifts for a particular recipient. Back-end management system for managing the content
25、s for supporting the above modules, which is important as adding adaptability to an on-line store greatly increases its complexity and the store can easily become unmanageable.To provide personalized product recommendations to customers based on their preferences, one needs to first create the repre
26、sentations for user preferences and product characteristics, and then define a measure for computing the similarity between them (see Section 4).3 PRODUCT CHARACTERIZATION & USER PROFILING3.1 Generic RepresentationA set of discriminative features F:=f1, f2, , fN has first to be identified based on d
27、omain knowledge. Then, the user preference can be represented as a vector of preference values on those feature representation u=u1, u2, u3, , uN ui Umin,Umax and the product can be characterized as a vector of values revealing the extent to which it possesses those features, denoted as p=p1, p2, ,
28、pN pi Pmin, Pmax. Product CharacterizationBased on a chosen set of features, product characteristic vectors p have to be created for all the products. Unless for the cases where each product comes with a detailed product description so that some information extraction techniques can be applied, huma
29、n effort for the creation of p is inevitable.User ProfilingFor acquiring user preference profiles u, it can be achieved by filling in a questionnaire during the registration process. However, in practice, requiring the users to provide preference values for a long list of features is infeasible as t
30、he required effort may simply scare them from continuing to shop in your store. So, the questionnaire for a newly registered user has to be reasonable short and the questions should be easy enough for the user to provide answers. Typical examples are the demographic data like gender, age and occupat
31、ion, here denote as d=d1, d2, , dM di Di, where Di is a set of possible stereotypical categories for di.1 However, such a simple representation contradicts the requirement for a discriminative set of features. One solution is using domain knowledge to transform the demographic information d user int
32、o a preference profile representation u containing a rich set of features via a transformation fu(d): D Umin,UmaxN where D:= D1D2 DM. The precision of the preference profile thus highly relies on that of the transformation.Another issue related to user profile representation is about the importance
33、of each individual feature. Under the aforementioned feature vector representation, user preferences on the features are assumed to be equally important. However, this is not the case in practice. Some users may consider color to be a more important feature than durability while some may find it the
34、 other way round. The situation can be even worse as this kind of information is usually unconscious for users and hard to be provided precisely. In our system, we model the relative importance of the feature with a weighting vector w=w1,w2, ,wN wi 0,1 and . Also, we introduce one more transformatio
35、n fw(d): D 0,1N. This transformation can be interpreted as the relative importance of the features for different combinations of demographic categories. It is hoped that this can free up the user from providing subjective weighting values. Again, the precision of the transformation is crucial to the
36、 success of weighting application.Obtaining the transformations that can effectively reflect the interests of the different demographic categories is by no means straightforward. Some possible objective means include conducting marketing surveys or analyzing past transaction records. Regarding their
37、 implementations, the input dimensions of the two transformations are equal to . Creating them directly may result in large storage requirement as well as tedious work in creating and managing them. By assuming the effect of each elements in d on the overall transformation to be independent, the tra
38、nsformation for preference can be decomposed into a set of transformations , each corresponds to a particular element in d. The storage requirement can then be reduced from to . With the decomposition, the preference profile is then computed asu = where denotes the ith element of fs output. The rang
39、e of value for each element in u remains to be Umin,Umax. Similarly, the transformation for weighting can be decomposed as . More details about the use of the weighting vector are described in Section 4. 3.2 An On-line Furnishing Company PrototypeTo provide a concrete example for explaining the repr
40、esentation issue, we have built an on-line furnishing company prototype for idea illustration.2 The furniture items include tables, sofa, beds, quilts, etc. For user profiling and product characterization, the set of features F we have used is shown in Figure 1 and the range of value for each elemen
41、t in both representations is set to be Umin=Pmin=-1 and Umax=Pmax=+1. Products with softness=“-1” means that they are extremely hard whereas those with softness=“1” means that the product is very soft. For demographic information d, 3 attributes - gender, year-of-birth and occupation are adopted (i.
42、e., M=3). For the creation of the transformation functions fu(dj) and fw(dj) (see Figure 2 and Figure 3) as well as the product feature vectors p, it is done manually based on domain knowledge.After a user registers with our system, his/her basic personal demographic information will automatically b
43、e stored. If he/she logs onto the system again, a personal preference profile will be created based on the methodology previously described. Recommendation services can thus be provided.Product IDColorfulEssentialExoticEasy tocleanDurableSafeSoftModern10002-0.40.80.4-0.10.50.60.40.7100230.6-0.40.3-0
44、.20.40.30.10.6100450.30.40.8-0.2-0.10.40.80.1Figure 1: Examples of product feature vectors, pCategoryColorfulEssentialExoticEasy tocleanDurableSafeSoftModern35-45 yr old-0.30.70.10.80.80.30.1-0.4Female0.1-0.20.20.5-0.20.50.2-0.1Housewife-0.50.8-0.30.9-0.40.1-0.2-0.1Figure 2: Examples of preference t
45、ransformation, fu(dj)CategoryColorfulEssentialExoticEasy to cleanDurableSafeSoftModern35-45 yr old0.030.180.060.330.240.080.040.04Female0.20.080.110.150.080.090.090.2Housewife0.040.170.080.310.220.10.040.04Figure 3: Examples of weighting transformation, fw(dj)4 MATCHING ENGINE Given the user prefere
46、nce profile u, the product characteristics p and the range of preference values, a similarity measure can then be defined. In our prototype, as the preference value range is -1,1, one obvious measure is the dot product between u and p weighted by w, given as with the output equal -1,1.3 Based on the
47、 similarity scores computed, personalized product ranking can be achieved. It can also be used to customize the catalog for browsing with the hope that the user can identify their intended products with less mouse clicks. The keyword search engine can also benefit by ranking the search results based
48、 on the scores so as to improve the chance that the intended items are put on the first few pages of the search results. Besides, when there is a list of new products, personalized recommendation services can be provided to further improve the quality of customer services.5 INTELLIGENT GIFT FINDER5.1 Profiling Gift RecipientsFor on-line shopping customization, we used to focus on how to acquire the intere
