Individualized Garment Recommendation System for Online Shopping

Transcription

1 Proceedings of the 8th WSEAS International Conference on SYSEM SCIENCE and SIMULAION in ENGINEERING Individualized Garment Recommendation System for Online Shopping ERUJI SEKOZAWA Department of Information Systems Creation, Kanagaa University, Yokohama , JAPAN Abstract: -his study proposes a system that supports online purchases of garments ith functions to select and recommend garments on the basis of the customer s tastes. he system has three basic capabilities: () It analyzes the customer s tastes using the analytical hierarchy process (AHP) and selects and suggests clothes. (2) It designates a priority for suggesting the garments using correlations found by clustering. () It suggests second purchases of items that have been chosen by previous customers ith similar tastes ho bought a garment that the current online customer has just decided to buy. he second purchases are selected using market basket analysis. Key Words: - Analytical hierarchy process, Online shopping, One-to-one, Recommendation system, Coordination, Market basket analysis Introduction he only ay the customer can judge an item on display is to read the descriptions on the computer monitor (for example: type of apparel; color; price; brand; quality label; size) and look at the image of the item. Customers ho do not have full knoledge about fashion often find themselves unable to find any among the great number of possible purchases that match ell ith their on tastes. Others find it impossible to make up their minds among the choices they like and give up on the purchase. Others make purchases, but are still uncertain hether they have bought something that really suits them. he aim of this study is to offer basic functions to resolve the above issues in online shopping. Specifically, the objectives are to () analyze a buyer s tastes using AHP, and select and offer pertinent clothes items; (2) create a cluster of apparel based on their correlation ith a buyer s tastes, and prioritize the items in the cluster; and () suggest garments bought by other buyers ith similar tastes, on the basis of market basket analysis (). 2 Outline of One-to-one Recommendation System Figure is a diagram of the system. As already described, the essential procedures in the function of this system are: () AHP; (2) cluster analysis; and () market basket analysis. hese ere used in combination. Also, databases for (a) apparel attributes and (b) simultaneous purchases ere used ith the three procedures above. Databases (a) and (b) in Fig. are described in Section 2, and examples of the application of procedures (), (2) and () above are described in Sections, 4 and, respectively. Questionnaire table of customer Clothes performance data by professional () Preference analysis by AHP Clothes matching the are retrieved (a) Clothing attributes database Generalized evaluation value of customer (2) Result of analysis of Cluster analysis Ranking of proposal Customer's combination purchase data Fig. : Overall structure of online system Coordination (combination) proposal () Market basket analysis (b) Simultaneous purchase database Association rule 2. Databases required by the system (a) Database of apparel (product) attributes his is shon in (a) of Fig.. Each apparel type as entered ith its digitized assessment scores before the experiment. One standard of these criteria, for example, as Design. he score as entered in absolute values as follos:, Very Good;, Good;, Average;, Bad;, Very Bad. hus, the database as constructed ith nine-level rankings of all the clothes according to their attributes. (b) Simultaneous purchase database A diagram of this database is presented in Fig. (b). ISSN: 0-26 ISBN:

2 Proceedings of the 8th WSEAS International Conference on SYSEM SCIENCE and SIMULAION in ENGINEERING his database is a tabulated record of all apparel items that ere bought together on any date, after analysis of the apparel items. he database can be used to calculate the incidence of purchase of items and the confidence that they ill be bought ith something else. Association rules can be constructed to anticipate hat combinations customers are most likely to select. Clothes of customer preference System Sensitivity Design Silhouette 2.2 Outline and procedure of analysis ()Grasping the customer s tastes ith AHP and selecting apparel An apparel (product) database is created for digitized characterizations of clothing. he system surveys customers, analyzes their tastes ith AHP procedures and searches for clothes popular ith them. system system system Avant-garde Contemporary Conservative Elegant Cut and sen Shirt Pants Simple Fig. 2: Hierarchical structure in AHP Plump build Jacket Slim build (2) Cluster analysis for customer tastes K-means clustering is applied in cluster analysis by grouping items in the database that have strong correlations ith each other in order to classify popular apparel in clusters. () Recommending apparel combinations A simultaneous purchase database is created for use by the fashion advisor. Market basket analysis is used ith the clustered apparel items to recommend apparel ith a high degree of confidence that the customer ill like, hile the customer is buying a part of a combination. Digitization of customer tastes ith AHP One feature of AHP is its capability to quantify the vague elements in human perceptions and assign numerical values for decisions. In this research, AHP as applied for digitization of human tastes and objective judgments.. Hierarchical structure for searching for popular attire AHP allos the user to assess apparel from a ealth of different angles, and on this basis, to construct a system capable of searching for the optimal attire. Figure 2 shos the hierarchical structure that as created. Level is the general Apparel customers like objective, and Levels 2 and include assessment standards. he loest, Level 4, has suggestions for alternative purchases, including Cut and Sen, Shirts, Pants, and Jackets..2 Analysis of priority of assessment standards by AHP he priority of each element in the assessment standards is determined. First, the essential assessment standards on Level 2 of the hierarchical structure are compared ith each other and analyzed. Customers are surveyed for their on priorities among the four assessment standards on Level 2, and the results are analyzed. he paired-comparison matrix in able shos the results of the survey after multiplication by eighting vector W. For this customer, the most important factor as Factor System (0.8), folloed by Sensitivity (0.26), Design (0.22) and Silhouette (0.0). he highest-valued eigenvalue of the paired comparison matrix for the standards provided by the survey as λ max = 4.8, and the consistency index able : Paired comparison of assessment standards (prioritization) System Sensitivity Design Silhouette λmax System / / / 4.8 Sensitivity / / Design / 0.0 (CI) as 0.0 < 0.. his value as significant. CI is defined in Eq.(). λ max n C.I. = () n C.I Silhouette (Weight) ISSN: ISBN:

3 Proceedings of the 8th WSEAS International Conference on SYSEM SCIENCE and SIMULAION in ENGINEERING able 2 shos customers evaluations hen there ere three assessment standards for System. he reader can see that this customer s rating of the factors as,,, in diminishing order of importance. his evaluation had a CI under 0., so it can be considered valid. Let us analyze a survey of customers for the importance of the third-level factors in Sensitivity. he factors in Design and Silhouette ere also prioritized. able 2: Paired analysis of three factors in System (evaluating priority) System λmax / /.0 / C.I 0.0. Procedure for absolute evaluations Let us turn to the set of articles from hich alternative purchases could be suggested, the fourth level of hierarchy. When the CI exceeded 0., the values in the paired comparison matrix had to be re-confirmed. Hoever, it is difficult to identify hich pair violates consistency if the set of clothes (to put it another ay, the set of possible recommendations) contains 60 articles. In addition, hen another item is added to the database, the paired comparison must be re-run. hus, there are problems in the conventional method for comparative assessment. Absolute assessments ere employed instead of comparative assessments. his method compares standards, rather than comparing specific clothing articles according to the standards. Procedure: Setting absolute assessment standards he absolute standards are as follos: Very good; good; average; bad; and very bad. able shos these standards. Very good Good Bad Very bad λmax able : Absolute assessment standard Very good / / / /.24 Good / / / / / C.I Bad / 0.06 Very bad (Weight) (Weight) he folloing evaluation vector as normalized to the maximum eigenvalues of the assessment standard comparison matrix: u = (0.0,0.260,0.4,0.068,0.0 ) (2) Procedure 2: Assessment matrix S ij shos assessed value for alternate garment j ith respect to standard i. S ij is defined as the assessed value a ij for alternate garment j ith respect to standard i is divided by the maximum assessed value a imax. a ij(garment j value by standard i) Sij = () a (maximum value under standard i) imax Procedure : Overall assessed value he survey results shon in <.2> ere processed ith AHP and indicated the folloing eighting vectors for the assessment standards (four elements), System (three elements), Sensitivity (three elements), Design (three elements) and Silhouette (three elements). 2 4 = (0.8, 0.26, 0.22, 0.0 ) = (0.6, 0.260, 0.06 ) = (0.6, 0.260, 0.06 ) = (0.6, 0.260, 0.06 ) = (0.0, 0.2, 0.64 ) corresponds to the second level of the hierarchical structure, and 2 correspond to the third level. he above shos ho the relative eights ere assigned beteen assessment standards and to the standards themselves. he assessment matrix as found in Procedure 2, so the overall evaluation scores for the clothing articles to be suggested are calculated as follos: Ei = Sij W () In absolute comparisons, the evaluation vector u is identified, the attire attributes database is established, and the assessment matrix is found ith Eq.() above. his is calculated using the assessment matrix instead of the relative comparison eighting matrix. Figure provides an example of the results of this kind of analysis. It is graphs of three clothing articles selected by eighting for cut and sen. he M-400 in the graph means that it is item #400 from Cut and sen. he reader can see from this figure that all three samples have high eights in System, a standard important to this customer. hese (4) ISSN: 0-26 ISBN:

4 Proceedings of the 8th WSEAS International Conference on SYSEM SCIENCE and SIMULAION in ENGINEERING Clothes number % Criterion 2Sensitivity 4Silhouette System Design Avant-garde Contemporary Conservative Elegant Simple Plump build Slim build Fig. Examples of overall scoring of three clothing articles have high eights in, Avant-garde, Elegant Design, and Slim. 4 Classification of tastes and recommending attire on the basis of cluster analysis Clothes items ere grouped into clusters using the clothing attributes database. he intent of this clustering as to assemble groups of clothes that are likely to appeal to a ide variety of people s tastes. Clothes suiting customers likes (clothes ith high overall scores) are classified into clusters on the basis of the items suggested by AHP. Cluster analysis also makes it easier to apply market basket analysis effectively, as described belo. his generates suggestions for garments coordinated ith other purchases that also meet the customer s tastes. 4. Clustering using K-means his study employed K-means clustering, a partitioned optimization procedure ithin nonconfigurational classifications. In K-means clustering (), the cluster centroid c i is considered the representative point and the set of data is divided into k clusters by minimizing the folloing evaluation equation: k ( i i = x c J = D ( x, c )) (6) i x: Article of clothing classified into cluster D(x,c i ): Distance beteen centroid c i and attire x Clustering results Clustering analysis is performed ith the clothes attributes database in order to group articles of clothing that can correlate ith each other in any of many different ays. he examples shon here ere obtained using clustering about the System criterion and setting the number of clusters at 2. he cluster numbers are set automatically by the system in the order of increasing sample number, and the system can be set to create clusters of uniform size. he number in the table shos to hich cluster the item belongs. Each cluster is displayed in a simple manner, and it is easy to display hat characteristic a cluster has in common. Clothes items ith similar attributes of each type can be combined into a single large display, as shon in able 4. he table presents items clustered for similarities in, Avant-garde, Simple design and Slim build. he M-420 cut and sen, S-200 Pants, G-20 Shirt and V-20 Jacket are grouped here because of their similarity. able 4: Combinations of similar clothing items Cluster Number No. of samples C-M-4 C-S-2 C-G-2 C-V Indivi duality Avant Contem Conser Plump Slim Elegant Simple garde porary vative build build ISSN: ISBN:

5 Proceedings of the 8th WSEAS International Conference on SYSEM SCIENCE and SIMULAION in ENGINEERING 4. Prioritizing merchandise for recommendation to the customer Clothes items bearing certain similarities ere grouped into clusters, as described in the previous section, as a ay of organizing them in accordance ith a ide variety of customers tastes. he items in each cluster ere then prioritized to plan the appropriate order to present them to the customer. Ordinarily, hoever, the first natural groups for selecting clothes are Cut and Sen, Shirts, Pants and Jackets. Here, in order to avoid the bother of selecting names to apply to ne clusters and keep things easy for the customers to understand, the clothes are displayed using the above classifications. he system can recommend articles selected to meet the customer s tastes in order of priority, as shon in able. he priority is assigned, beginning ith the item having the highest eighting. In the above example, the recommended items ere the M-00 cut and sen, the S-00 Pants, the V-00 Jacket and the G-2020 Shirt. able : Priorities set for clothes items Customer s priority Cut and sen Pants Jacket Shirt st place M-00 S-00 V-00 G nd place M-2020 S-00 V-040 G-040 rd place M-040 S-400 V-2020 G-00 Coordination of attire ith market basket analysis In market basket analysis, the fashion advisor recommends items to the current customer that other customers ith similar tastes have bought. his form of recommendation is an essential role of the human sales staff in a brick-and-mortar store. A virtual sales girl on an online shopping site can perform the same task, using market basket analysis. Here, to attain the goal of coordinating attire, the clothes are previously sorted into the above four clusters of Cut and Sen, Shirts, Pants, and Jackets. It is investigated ho to suggest clothes in combinations that ill suit the client.. Issues and Policies in the creation of purchase databases An attire attributes database is used to create a database that allos calculation of the probability that to items ill be bought at the same time, and this is essential hen applying market basket analysis. Performing market basket analysis ith the simultaneous purchase database allos the seller to suggest attire besides the item that as just purchased that has a high potential of sale. Association rules are constructed to calculate the rate of appearance of an item in the purchase database and the confidence that it ill be bought along ith something else..2 Construction of association rules Association rules are constructed using clothes items that have been clustered by many applicable characteristics and gathered into a simultaneous purchase table in order to calculate the probability of simultaneous purchases. An intermediate product such as a simultaneous purchase table can provide information about hich combinations of garments appear most often in transactions. he folloing shos ho a simultaneous purchase table and association rules are created. Confidence in able 6 describes ho high the probability is that the customer ill buy any of the pants (S-00 00) after buying V-00 and G-00. able 6: Simultaneous purchase table (partial) and confidence levels Preceding garment sold V-00 V-00 V-00 Combination G-00 G-00 G-00 Folloing garment sold S-00 S-020 S-00 Incidence of preceding garment and folloing garment Incidence of preceding garment Confidence Incidence is the quantity obtained by dividing the total number of purchases of the given item simultaneously ith some other item by the total number of items purchased at the same time as any other garment. Incidence of sales of the preceding garment means the incidence of sales of the combination of V-00 and G-00. Incidence of simultaneous sales of the preceding and folloing garment means the incidence of sales of all three items, V-00, G-00 and S he confidence is found in the folloing calculation: ISSN: 0-26 ISBN:

6 Proceedings of the 8th WSEAS International Conference on SYSEM SCIENCE and SIMULAION in ENGINEERING Confidence = Incidence of simultaneo us sales of the preceding and folloing Incidence of sales of the preceding garment garment () he calculated confidences for all combinations of preceding garment purchases ere used to establish the folloing example of the association rules: (i) V-00 & G-00 S-00 S-00 (ii) V-00 & S-00 M-00 G-020 (iii) V-020 & S-020 M-00 G-00 It is possible to put the confidence values in numerical order as ell, so a rule as ritten to recommend the most likely (highest confidence value) and, if necessary, the second most likely item for additional sale. 6 Conclusions he authors rote system to select and recommend garments on the basis of the customer s tastes, and proposed a solution to the problems on the conventional online system. Currently, a customer ishing to purchase a garment online must read the descriptions and look at the images of clothes and rely completely on her on subjective judgment to find and select the garment. his system aids the customer in selecting garments and also makes recommendations. ) Digitization of tastes using AHP People s tastes and perceptions of clothes must be expressed in the form of numbers. AHP as used to accomplish this. An AHP-based data search system as used to create a database of the attributes of garments, and this as used to perceive individual customers tastes and to suggest garments they ere inclined to like. 2) Classification of tastes and recommendations of clothes on the basis of cluster analysis Clustering of the data in the clothes attributes database allos the system to respond appropriately to customers of many different. Clothes fitting customers tastes can also be classified into clusters consisting of the items recommended by AHP. By using K-means clustering, the number of items to be examined can be reduced to a manageable level. Alternative garments can be suggested in order of priority, ithin the cluster of garments meeting the customer s. ) Coordination of garments by market basket analysis Market basket analysis allos the system to suggest clothes to a customer on the basis of hat has been bought by other customers ho have similar tastes. A complete database must be prepared of garments clustered on the basis of their characteristics. his is used to calculate the probability of simultaneous purchase of pairs of garments, and this calculation is applied to suggest potential second sales to customers. References: []ien-chin Wang, Ying-Hsiu Chen: Selection of the MAS by applying fuzzy preference relation, WSEAS ransactions on Business and Economics, Vol., 2006, pp [2]Ching-Fu Chen: Applying the Analytical Hierarchy Process (AHP) Approach to Convention Site Selection, Journal of ravel Research Vol. 4 Issue 2, 2006, pp.6-4 []Balan S, Vrat Prem, Kumar Pradeep: A strategic decision model for the justification of supply chain as a means to improve national development index, International Journal of echnology Management Vol.40 Issue-, 200, pp.6-86 [4]Chunhao Li, Yonghe Sun, Yanhui Jia, Hui Li: An improved ranking approach to AHP alternatives based on variable eights, IEEE Proceeding on Intelligent Control and Automation, 2008, pp []Hou Junming, Su Chong, Wang Wanshan: AHP Methodology for Partner Selection in Collaborative Design, IEEE International Symposium on Computer Science and Computational echnology, 2008, pp.64-6 [6]Wei-Hsin Chou: Using AHP to Assess a Plan of raining the Adolescent Golf Player, IEEE Fifth International Conference on Fuzzy Systems and Knoledge Discovery, 2008, pp.- []Sloane E.B., Liberatore M.J., Nydick R.L., Luo W., Chung Q.B.: Clinical engineering technology assessment decision support: a case study using the analytic hierarchy process (AHP), IEEE Engineering in Medicine and Biology, Vol., 2002, pp.0- ISSN: ISBN: