D.Sailaja, K.Nasaramma, M.Sumender Roy, Venkateswarlu Bondu

Transcription

1 Predictive Modeling of Customers in Personalization Alications with Context D.Sailaja, K.Nasaramma, M.Sumender Roy, Venkateswarlu Bondu Nasaramma.K is currently ursuing her M.Tech in Godavari Institute of Engineering and Technology. Rajahmundry, India Sailaja.D is currently ursuing her M.Tech in Avanti Grou of Colleges, Visakhaatnam, India. Sumender Roy is currently working as Professor in Godavari Institute of Engineering and Technology. Rajahmundry, India. Venkateswarlu Bondu ursuing Ph.D in Comuter Science in Andhra University. Abstract--- The idea that context is imortant when redicting customer behavior has been maintained by scholars in marketing and data mining. However, no systematic study measuring how much the contextual information really matters in building customer models in ersonalization alications has been done before. In this roject, we show how imortant the contextual information is when redicting customer behavior and how to use it when building customer models. It is done by conducting an emirical study across a wide range of exerimental conditions. The exerimental results show that context does matter when modeling the behavior of individual customers. These findings have significant imlications for data miners and marketers. They show that contextual information does matter in ersonalization and comanies have different oortunities to make context valuable for imroving redictive erformance of customers behavior. Index Terms Personalization, Context, Data Mining, User Modeling, Predictive Modeling.. INTRODUCTION:. Introduction to Project: Contextual information indeed makes a significant difference in building better customer models in marketing and e-commerce alications. In this roject, we address the question of whether this additional contextual information matters, i.e., does it lead to building better ersonalized models of customer behavior, where by better we assume suerior redictive erformance. This roblem is not trivial because it entails a tradeoff between transaction homogeneity and data sarsity: by roviding contextual information, customer transactions ertaining to this articular context are reduced, making fewer data oints to fit the model, while homogeneity of these transactions increases, making it easier to redict more accurately customer behavior in similar contexts. In data mining terms, this roblem is related to the well-known biasvariance tradeoff, i.e., given contextual information, which effect dominates the other: decreased bias due to the homogeneity of transactions associated with the secified context or increased variance due to insufficient data associated with this context. Therefore the research question that we just described can be summarized as follows: Does context matter for building better models to redicting customer behavior? In this aer, we answer the question emirically by conducting an emirical study on data set across a wide range of exerimental conditions. To answer the question, we built two alternative customer models, one including contextual information and the other one not, and comared their redictive erformances. This study makes the following contributions to studying context in ersonalization alications. First, we demonstrate that context indeed matters when redicting customer behavior for whole or small homogenous grous of customers and gets diluted during the rocess of aggregating customers data. Finally, the context is taking externally, and then used for redicting customer s behavior. We show that the resulting model significantly outerforms the basic uncontextual model..2 Introduction to modules:.2. Collecting dataset externally..2.2 Clustering the data..2.3 Aly redictive modeling on datasets and clusters. 34 P a g e

2 .2. Collecting dataset externally: The exeriment has been conducted on e-retailer dataset which is created by us. For each customer, the following demograhic data were added: age, revious studies, marital status, comosition of the family, lace of living, hobbies, and whether the customer owned a car. The transactional data include: item urchased, rice, day, time, session duration, number of clicks er connection, and the time elased for the web age..2.2 Clustering the data: Aly K-mean clustering algorithm on dataset and divide into clusters: cluster, cluster2 and cluster Aly redictive modeling on dataset and clusters: Aly Naive-Bayes classification algorithm to redict customer behavior on whole dataset and each clusters are cluster, cluster2 and cluster3. 2. PROBLEM FORMULATION: In our roject, first exlain what we mean by context, then how we model customer behavior, and finally, the methodology for comaring contextual and uncontextual models. 2. What Is Context? Many definitions of context can be found in the literature deending on the field of alication, enabling technologies, and the available customer data. The Webster s dictionary defines context as conditions or circumstances which affect something. In the data mining community, context is defined as those events that characterize the life of a customer and can determine a change in his/her references, status, and value for a comany. Examles of context include a new job, the birth of a son, marriage, divorce, and retirement. In the contextaware systems literature, context was initially defined as the location of the user, the identity of eole near the user, the objects around, and the changes in these elements. Other factors have been added to the revious definition. For instance, includes the date, the season, and the temerature. Add the hysical and concetual statuses of interest for a user and include the user s emotional status and broaden the definition to any information that can characterize and is relevant to the interaction between a user and an alication. Some associate the context with the user, while others emhasize how context relates to the alication? Context has temoral (when to deliver), satial (where), and technological (how) dimensions. Context is usually referred to the resent situation, but sometimes the history of ast is considered as well. In this aer, context is defined as the intent of a urchase made by a customer in an e-commerce alication. Different urchasing intents may lead to different tyes of behavior. For examle, the same customer may buy from the same online account different roducts for different reasons: a book for imroving her ersonal work skills, a book as a gift, or an electronic device for her hobby. In general, the context in which a customer erforms a transaction is defined with a set of contextual attributes K that can have a comlicated structure reflecting the comlex nature of this information. Each contextual attribute K in K is defined by a set of q attributes: K = (K;...;Kq) contextual attribute K secifying the intent of a urchasing transaction in an e-retailer alication, considered. K is ersonal context: ersonal urchase made for the work-related or other uroses. Similarly, the Gift value for K can be slit into a gift for a artner or a friend and a gift for arents or others. Thus, K= {PersonalWork; PersonalOther; GiftPartner/Friend; Gift Parent/ Other}. Finally, attribute K to be taken. 2.2 Customer Modeling: Let C be the customer base reresented by N customers. Each customer C i is defined by the set of m demograhic attributes A ={ A ;A 2 ;...;A m }, and a set of r transactions Trans(C i )={ TR i ; TR i2 ;... ; TR ir }, where each transaction TR ij erformed by customer C i is defined by a set of transactional attributes T ={T ; T 2 ;... ; T }. In addition, we also have contextual information K associated with each transaction TR ir, in Fig. reresents a fragment of the customer table containing demograhic, transactional, and contextual information about the customer C i. For examle, customer C i can be defined by demograhic attributes A= {IDuser; Name; Age; Income}, by five transactions Trans(C i )={TR ; TR 2 ; TR 3 ; TR 4 ; TR 5 }, each transaction defined by the transactional attributes: T ={ ProductID; StoreID; Price; TransactionTime}. In general, however, we suort multile contextual attributes. Finally, the customer base C can be artitioned into several segments by comuting summary statistic S for customer C i over the transactions Trans(C i )={TR i ; TR i2 ;... ; TR ik } made by that customer using statistical aggregation and moment function, such as mean. For instance, for the transactions made by the customers in the revious examle, the statistic can be S = {Average rice}. This means, among other things, that each customer C i has a unique summary statistic S and that a customer is reresented with a unique oint in the sace of these summary statistics. After generating such a data oint er customer in the 35 P a g e

3 sace of statistics S, customers can be clustered into grous (segments) in that sace using the K-Means clustering technique. Given segment α, of k customers C ;... ; C k, and their resective demograhic A i ={ A i ; A i2 ;...;A im } and transactional data Trans(C i )={ TR i ; TR i2 ;... ; TR ik } for customers i in α, we want to build a single redictive model M α on this segment of customers α : Y =f(x ;X 2 ;...;X );..(); where deendent variable Y is one of the transactional attributes T j, and indeendent variables X ;X 2 ;..;X are all the transactional and demograhic variables, excet variable T j, i.e., they form the set T A T j. The erformance of model M α can be measured using some fitness function f maing the data of this grou of customers α into real, i.e., f( α ) Є R. For examle, model M α can be a decision tree built on data α of customers C ;... ; C k, for the urose of redicting T j variable time of urchase using all the transactional and demograhic variables, excet variable T j as indeendent variables. The fitness function f of model M α can be its redictive accuracy on the outof-samle data. The redictive models do not assume any contextual information since the contextual variable K is not a art of the model. Therefore, we call the model of this tye uncontextual. We define contextual counterarts of redictive model (), where the model takes the following form: Demograhic Attributes A Transactional Attributes T Y =f Kq=a (X ;X 2 ;...;X ); (2) Context K A A m T T K K q TR j A j A jm T j. TR j2 A j A jm T j2. TR jr A j A jm T jr. T j. T j2. T jr.... transactions associated with a articular value of the context attribute K q =α are used for building the model. In this case, the contextual information is used as a label for filtering customer transactions and then droed variable, such as the demograhic and transactional attributes X ;X 2 ;..;X. This means that it is used as one of the attributes for redicting Y.One interesting question when building contextual models is where to lace urchasing transactions of customer C when she bought a gift for customer C 2 : should such a transaction be associated with the urchasing history of customer C or C 2? In this aer, we associate such urchases with customer C and not C 2 for the following reasons: First, these urchases reflect ercetions of customer C about what customer C 2 needs, not the real traits and needs of customer C 2. Second, even though the user may want to interret exectations and references of another individual, it would be very unusual to model behavior of a erson by observing the behavior of another individual. Third, when building a model for customer C, the demograhical and transactional data used in this behavioral model are those related to customer C. One way to handle this roblem of gifts is to define an aroriate context of gifts and urchases for others and treat such urchasing behavior in these contexts. This rovides for extra flexibility because we can treat such urchasing transactions differently in different contexts. For answering the research question (Does context matter?), a comarison between erformance results of redictive models is erformed for the uncontextual () and the contextual (2) models across a wide range of exerimental conditions can be secified by relacing the deendent variable Y in () with the context variable K as K q =f (X ;X 2 ;...;X ); (3). In model (3), the deendent variable is the contextual information. One contextual attribute is inferred at a time. For model (3), f is a redictive function learned via naïve bayes model. 2.3 Figures: Demograhic, transactional, and contextual information about the customers and their transactions.: Fig. Where the model 2 constitute the way of creating a contextual model. Model 2 indicates that only Grahical reresentation of the redictive model: 36 P a g e

4 Fig Results: 3. Does Context Matter? To give a flavor of the results, given the exeriment settings (contextual information, data set, one classifier, two deendent variables)., the below figures resents three grahs generated by lotting the values of the e-retailer data set for contextual information and uncontextual. The grahs are resented in the order of rogressively more refined contextual information. Moreover, most of the contextual models show a better redictive erformance comared to the uncontextual, excet for some cases where the difference is very small. Although for these charts the curves are not always monotonic, the redictive erformance of the contextual models is usually higher than that of the uncontextual model. Analyzing and comaring the erformance curves, the first evident result shows the uncontextual model is always below the other two contextual models. It means that, indeendently from the level of analysis, and from the exerimental settings, gathering contextual information gives better results in terms of customers behavior redictions. Another interesting oint is the monotonic shae of each erformance curve. The results shows better results in context comared to uncontextual in three different stores (store00, store0, and store02). In three stores shows context is better comared to uncontext in two tyes of charts (curve, bar). 37 P a g e

5 Is it necessary to acquire contextual information or is it ossible to infer it from the data? How do we exloit the inferred contextual information for modeling customer behavior? The exeriments will be conducted redicting different transactional variable Y, using more classifier algorithms f, considering more levels of market granularity through finer segments, and generating more contextual models using the further sublevels of the ersonal and gift shoing mode variables. 5. BIBOLOGRAPHY:. Using Context to Imrove Predictive Modeling of Customers in Personalization Alications. Cosimo Palmisano, Alexander Tuzhilin, Michele Gorgoglione. IEEE Transaction on Knowledge and Data Engineering. 4. CONCLUSION & FUTURE WORK: 4. Conclusion: This work has been conducted a comarative study of contextual and un-contextual models across multile dimension of analysis such as the context variable, classifier algorithm and different tyes of metrics. This analysis has been erformed with the aim of demonstrating the relevance of the context in building customer rofiles. Our results show that, in different exerimental settings, the contextual model, in general, outerforms the un-contextual one; this overcome is more relevant. This underlines the fact that the context matters in building customers urchasing rofiles and it should be of great effectiveness if alied for building ersonalized recommendation in the e-commerce environment. 2. Data Mining Concets and Techniques.- Han & Kamber. 3. Increasing Customer Value by Integrating DataMining and Camaign Management Software, Direct Marketing Magazine, Feb, Data Mining: Introductory and Advanced Toics Margaret Dunham. Prentice Hall, Mining the e-commerce data to analyze the target customer behavior. Yuantao Jiang, Siqin Yu. 6. Java Comlete Reference. Herbert Schildt. 7. Unified Modeling Language. Grady Booch. 8. htt:// Further research will be held making more comlex the exerimental settings in order to give more emirical evidence of the results. And will rove two more questions for better redicting 38 P a g e