Deal or No Deal: Catering to User PreferencesKartik Muralidharan, Swapna Gottipati, Rajesh Krishna Balan

Singapore Management University

{kartikm.2010,swapnag.2010,rajesh}@smu.edu.sg

Abstract—A common problem in large urban cities is the hugenumber of retail options available. In response, a number ofshopping assistance applications have been created for mobilephones. However, these applications mostly allow users to knowwhere stores are or find promotions on specific items. What ismissing is a system that factors in a user’s shopping preferencesand automatically tells them which stores are of their interest.A key challenge in this system is building a matching algorithmthat can combine user preferences with fairly unstructured dealsand store information to generate a final rank ordered list. Inthis work, we present the first prototype of myDeal, a systemthat automatically ranks deals according to user preferences andpresents them to the user on their mobile device.

I. INTRODUCTION

“Jill is walking down the street and notices a new mall occupying a

formerly vacant lot. Similar to other malls in her country, this mall has 100 or

more stores inside. Jill loves to shop but the thought of having to walk through

this new mall and identify the shops that are of specific interest to her is not

that appealing. In addition, even for her popular stores, she really only wants to

visit them if they are having a sale or other interesting promotions. She wishes

that there was an application on her mobile phone that would automatically

tell her of stores and promotions of interest to her everytime she encountered

a mall or other shopping area.”

The above scenario happens frequently to consumers inlarge urban Asian cities, such as Tokyo, Bangalore, Hanoi, Sin-gapore, etc., where commercial shopping malls are interspersedwith regular office and residential buildings. In addition, inthese Asian cities, it is not uncommon for new malls to appearand disappear fairly regularly. The store selection in existingmalls also frequently change in response to market pressuresand shopping trends. Finally, the density of malls in these citiestends to be quite high (malls can be just tens of meters apartfrom each other in the worst cases).

All of these factors make it incredibly hard for consumersto identify stores of interest to them in any particular mall.However, it is not sufficient to just identify interesting storesby name as consumers also want to identify stores that arehaving deals of specific interest to them [1]. These promotionscome in two forms — 1) promotions that are offered by thestore regardless of payment implement (50% off storewide salefor example), and 2) promotions that are offered if specificpayment or discount implements are used (for example, 20%off when using a VISA card, or mileage points if a frequentflyer card is shown, etc.). Finally, consumers would prefer tofind the most interesting stores, with best deals from theirperspective, from the ease of their mobile device.

Existing representative mobile advertising i.e., location-based advertising, is limited in effective targeting. This ismainly because it delivers promotions of near places justdepending on a user’s current location without considering anyother user context. For example, deals for a nearby restaurantpromoting group dining offers do not attract people who aredining alone. Other applications like LiveCompare [2] catertoward on-the-fly price comparisons for mobile users, whileretail or card specific applications such as Citi Shopper [3]

allow users to browse and view card specific deals. Howeverneither application group shows deals that are independent ofany product or payment card.

In this paper, we present the myDeal system that attemptsto correct the deficiencies of previous solutions and provideconsumers with a context-based mobile shopping application.Here, in addition to location, we allow the user to specifyother context such as preferences as well as payment cardsowned - all important pieces of information needed to improvepromotion relevance. The key complexity in providing thisassistance arises from having to combine both structured(easy to understand numeric discounts (“5% off”) etc.) andunstructured (free-form text (“A free teddy bear”) etc.) pro-motion information (1). The algorithm also needs to factorin consumer preferences such as “prefer discounts over freegifts” etc (2). Also, most consumers own multiple paymentand discount cards and this multiplies the complexity as eachindividual card (and combinations of) has its own set ofpromotions and deals (3). Our final algorithm adapts naturallanguage semantic techniques and combines all three factorsto provide users with the most relevant promotions and showthose immediately — with the remaining promotions hiddenaway from view until the consumer goes and actively looksfor the less interesting promotions.

II. RELATED WORK

At one end we have stores with promotions for theirproducts and services. At the other end are consumers withtheir needs. While a number of researchers [4], [5], [6] havesuggested the use of user context to provide a higher degreeof ad personalization, most research has focused on using onlylocation to increase the relevancy of advertisements delivered.In the AdNext system [7], authors propose to use mobilitypatterns to predict which store the user is likely to visit next,and show him advertisements related to that store. CommercialLBAs such as Shopkick [8] also exploit current user locationto push relevant advertisements. More recently SmartAds [9]deliver contextual ads to mobile apps by taking into accountthe content of the page the ads are displayed on. These worksare orthogonal to myDeal; myDeal allows users to explicitlystate their preferences and pull relevant promotions onto theirmobile device whenever required. The key focus of myDealis being able to match these user preferences with fairlyunstructured ad content.

Other competitors fall into two broad categories. In the firstcategory, applications like Froogle [10] cater towards on-the-fly price comparisons for mobile users. In the second categoryare deal information services, like Citibank’s Card InformationService [3] and mobile applications like Mobiqpons [11], thatattempt to inform consumers about the best deals available.However, these applications fail in two main areas; first, theyare usually targeted to just certain malls and/or certain paymentcards (only Citibank cards for example), and second, they donot take into consideration user preferences - factors that areimportant in getting the best deal! In all the cases we observe

2014 IEEE International Conference on Pervasive Computing and Communications Work in Progress

978-1-4799-2736-4/14/$31.00 ©2014 IEEE 199

Fig. 1. myDeal System Architecture.

that the burden of finding the best deal is left to the consumers— the applications just provide different ways to find lists ofdeals matching certain broad criteria.

Collaborative filtering(CF) techniques have also been usedfor targeted advertising [12], [13]. CF requires users to activelyparticipate and express his or her preference by rating items inthe system. These inputs are used to build user-item (or user-rating) matrices and recommend similar promotions for userswith similar taste. We believe however, that user interests arehighly situational and hence require myDeal users to explicitlystate their preference every time the system is used. However,myDeal could incorporate such additional signals to furtherincrease promotion relevance.

III. THE MYDEAL SHOPPING ASSISTANT

A major limitation with all existing shopping assistanceprograms (that we know off) is that they do not really takeinto account user preferences. At best, the user is allowedto specify categories or keywords and the programs sort theresults based on that. However, a single mall can have hundredsof stores with many hundreds of deals between them. Assuch, even keyword searches and categories break down inthis type of rich data environment. This problem is made worsewhen we consider that there are multiple malls usually withinwalking distance and that the consumer has multiple paymentimplements (which usually offer promotions on top of thosealready offered by the stores themselves)1.

A key consideration for us was to therefore integrate userpreference into our system so that only the most relevantpromotions and stores are brought to the consumer’s immediateattention. However, this is a non-trivial process as promo-tions are stated using a combination of structured (easy tounderstand numeric discounts. e.g.“5% off”) and unstructured(free-form text. e.g.“Free teddy bear”) components. A naiveapproach would be to just rank promotions based on theeasy to sort structured components. This is not good as manypromotion details, of high interest to the consumer, tend to belocated in the unstructured components.

A. Finding the Best DealKeeping in mind this design consideration, the first building

block needed for myDeal is a way to describe deals. Our sys-tem architecture is shown in Figure 1. We manually inspecteda few hundred deals, from various retailers and payment cards,and identified just four components that were used to createall deals - every deal we encountered was some combinationof these four components with different values and attributes

1A pre-study demographics survey showed that several users carried be-tween 4-6 cards in order to avail such offers.

assigned to each component. We used a simple XML-basedschema to describe deals comprising of these four components:

1) Cash back: These are specific cash refund. For exam-ple, 3% cash back of entire bill. These discounts arein the form of percentages or fixed values ($5 cashback).

2) Discounts: These are specific cash discounts. Forexample, 5% cash discount of entire bill. Thesediscounts are in the form of percentages or fixedvalues ($10 discount).

3) Vouchers: These are vouchers that can be accruedand then exchanged later for either cash or products.Frequent flyer miles, store loyalty points, etc. aresome examples.

4) Rewards: These are deals in the form of real products.For example, get a teddy bear free with every $10purchase.

The second part of myDeal is a matching and scoringsubsystem that rank orders the best deals available for the user.These ranks determine how good a deal the user would get ifthey shopped at a particular retailer, possibly for a particularproduct, using particular payment and discount cards. Rankordering the best deals involves two major steps: 1) Matchingdeals preferred by the user to those offered by the retailer andcard issuers and 2) assigning scores for each of these validcombinations.

The matching algorithm is dependent on the parametersof four entities involved in most shopping scenarios; namelythe deals, the cards carried by the user, user preferences andlocation. The four entities are mapped in two steps. First,users are mapped to retailers based on their location and thekind of deals they are looking for (e.g. Dining) and their dealpreference. Second, retailers are mapped to the cards carriedby the user. The matching algorithm filters out those retailoutlets that do not match the above criteria.

In order to score any deal, we first need to extract valuesfor the following four categories from the deal description:Discount, CashBack, Voucher and Reward. For each categorywe use regular expressions to match and extract the corre-sponding values. For example, if the deal description is “Geta 20% Discount on the Total Bill and enjoy a complimentaryVoucher of $10” we extract the following values Discount =20 and Voucher = 10.

We then use the following formula to score a deal:

Score = α · Discount + β · CashBack + γ · Voucher + θ · Reward,

where α, β, γ and θ are weights to adjust the importanceof each deal category. The value of these weights are specifiedby the user as deal preferences.

However, deriving a value for a reward is slightly lesstrivial. Consider the following deal description: “Enjoy a FreeIce Cream or Cake with every meal purchased”. To applya score to this deal we must effectively assign a value toboth reward items Ice Cream and Cake. We propose a 2-stepmachine learning algorithm that uses semantics to determinethe corresponding value of the reward.

Step 1: Get a list of all reward items and their corre-sponding value (if specified) from all deals. This is done usingstandard NLP techniques such as part of speech (POS) tagging.For our system we use the Brill POS tagger from CST [14].In our example, the rewards Ice Cream and Cake are extractedand added to the list.

2014 IEEE International Conference on Pervasive Computing and Communications Work in Progress

200

Fig. 2. myDeal Usage Sequence on the Windows Phone 8

Step 2: Using a semantic similarity method described byLin et. al [15] and Pedersen et. al [16], we cluster rewardsin the list into a semantic space in which rewards that areclosely associated are placed in the same cluster. Severalexisting algorithms compute relatedness only by traversing thehypernymy taxonomy and find that Ice Cream and Cake arerelatively unrelated. However, WordNet provides other typesof semantic links in addition to hypernymy, such as meronymy(part/whole relationships), antonymy, and verb entailment, aswell as implicit links defined by overlap in the text of defi-nitional glosses. These links can provide valuable relatednessinformation. If we assume that relatedness is transitive acrossa wide variety of such links, then it is natural to follow pathssuch as ice cream-frozen dessert-dessert, sweet-dessert and finda higher degree of relatedness between Ice Cream and Cake.

Lin’s similarity measure uses the information content (IC)of the words/concepts, and the least common subsumer (LCS)of the concepts in the WordNet taxonomy. LCS is the commonancestor of two concepts which has the maximum informationcontent. The similarity measure between concepts wi, wj isdefined as follows.

Sim(wi, wj) =2× IC(LCS(wi, wj))

IC(wi) + IC(wj)(1)

whereIC(c) = −log(P (c)) (2)

LCS(wi, wj) is a common subsumer of wi, wj , IC(c)is the information content of the concept c and P (c) is theprobability of c. In our example, the rewards Ice Cream andCake are likely to be clustered together as ’Dessert’. Themedian value of the cluster is then assigned to rewards thatcurrently do not have any value associated with them. Goingahead with our example, if the median value of the cluster’Dessert’ is $5, the rewards Ice Cream and Cake are alsoassigned the same value.

B. End-to-End SystemOur system works as follows: The user enters a shopping

mall and loads the myDeal application (Figure 2a). myDealpresents several options that the user can navigate through.The user can either choose to view all deals offered at thatlocation or may choose to specify a particular category ofinterest (e.g.Dining, Essential, etc.) as shown in (Figure 2b).The user then proceeds to input additional optional details liketype of deal preferred, keywords if any, and desired amountto spend and so on as shown in (Figure 2c). myDeal willextract user’s card details from the secure storage area of thephone, append any additional user input and location, and

Scenario Weights (%) Rank Error

No. C D V R Diff. Mag.

1 100 0 0 0 0.0 (0.0) 0.0 (0.0)

2 0 100 0 0 1.2 (1.03) 2.0 (2.0)

3 0 0 100 0 0.1 (0.0) 1.0 (0.0)

4 33.3 33.3 33.3 0 1.0 (1.05) 2.0 (0.0)

5 50 50 0 0 1.0 (1.05) 2.0 (0.0)

6 50 0 50 0 0.0 (0.0) 0.0 (0.0)

7 0 50 50 0 0.5 (0.7) 1.7 (0.5)

8 0 0 0 100 2.5 (2.0) 3.6 (1.3)

9 25 25 25 25 1.6 (1.26) 2.0 (1.07)

10 25 25 0 50 1.8 (1.8) 2.6 (1.6)

11 25 0 25 50 5.1 (4.2) 5.1 (4.2)

12 0 25 25 50 1.0 (1.4) 2.0 (1.4)

C=Cash Back, D=Discount, V=Vouchers, R=Rewards. Values in parenthe-ses are standard deviations.

TABLE I. ACCURACY OF ALGORITHM RELATIVE TO EXPERT

send them to the backend service over an existing wirelesscommunication channel. The ranking service will compute ascore which are then ordered and sent to the user’s mobileapplication (Figure 2d).

One question we faced when building myDeal was decidingwhere to perform the ranking. In particular, we could performthe matching and scoring on the user’s mobile phone orperform the ranking using an external service. Each of theseoptions had its strengths and weaknesses. Rank ordering on theuser’s mobile phone provides the highest amount of privacy— no card information is sent anywhere. However, the user’smobile phone is computationally limited and requires accessto deal information across multiple retailers and card issuers.While we are concerned about resource utilization and privacy,as these issues are not the primary focus of this work, wedecided to use a backend service to perform the filtering andscoring, at the risk of users card information(only card name)revealed to the hosting site.

IV. EXPERIMENTAL RESULTS

A. Results: The Algorithm is AccurateOur dataset for the algorithm evaluation consisted of real-

world deals manually extracted from multiple sources (10shopping malls and 4 major credit card providers). We useda total of 842 deals from 610 restaurant covering 7 cuisinetypes in the user study. Also, we observed from a detailedbreakdown of deal component combinations (cash back, dis-count, vouchers and rewards) 5% of the deals offered cashdeals while 25% of the deals offered rewards— validating ourdecision to specifically handle unstructured free-form rewardsin our ranking algorithm.

We evaluated the ranking algorithm by comparing its scoresfor the top 10 deals versus the scores of an expert using twelvedifferent scenarios. In each scenario, we changed the weightsallocated to each of the four deal components. Table I showsthe average difference in rank position and the average errormagnitude (for all the errors that were made, what was theaverage error) for the top 10 deals ordered by an expert withthat of our algorithm for all twelve scenarios. The total timetaken by our algorithm to score all 842 deals used in the userstudy was 312ms - well within reasonable limits.

The values in the “Rank Difference” columns indicate theaverage difference in the score ranks assigned by each entity

2014 IEEE International Conference on Pervasive Computing and Communications Work in Progress

201

for each set of deals while the numbers in the second columnshow that in the event of a rank variation what the averagedifference would. A lower number in both columns indicate ahigher level of agreement between the two entities in terms ofscoring.

The first seven scenarios were evaluated without consid-ering the unstructured data part of the deal (Rewards in theform of free text). The results show that the algorithm waseffectively able to extract values from the structured data partof the deal. In particular, the largest average difference betweenthe expert and the algorithm was just 1.2 (scenario 2) and evenin that case, the average magnitude of the error was just 2scoring positions (i.e., if the expert ranked a deal 3rd, if thealgorithm made a mistake, it would, on average, rank that deal1st or 5th).

In the last 5 scenarios (scenarios 8 to 12), the deal scoresincluded the free-form text of the deal. The average rankdifference and error magnitude in this case is relatively higher.This result was expected as scoring would involve assigning avalue to the free text which is subjective. The algorithm usessimilarity matching to assign this value while the expert usesheuristics to do the same. Overall, the average difference isnot large enough to invalidate the algorithm accuracy.

B. Current limitations of the Ranking AlgorithmA key part of the algorithm complexity lies in extracting

both the structured and unstructured information parts of thedeal. As the description of deals are not standard, extractionalgorithms are limited by the current state-of-the-art in naturallanguage processing rules and technologies. In our currentprototype, we cannot handle deals that include elaborate free-form conditions to satisfy the deal requirement. For example,several deals require purchasing additional items in order toredeem the primary offer. Any additional purchase should ef-fectively reduce the overall score and the algorithm would needto determine the value of this additional item and reduce theoverall score by the appropriate factor. Our system currentlycannot handle these types of deals in a general way.

V. FUTURE WORK AND CONCLUSION

A common conundrum in pervasive computing is decidinghow to present information to users of these systems. Onone hand, users should be provided as much information aspossible so that they can make better decisions. However,providing this much information tends to overload the userand have negative usability impacts. A common technique toreduce this information overload is to use automation (in theform of AI or similar techniques) to process the data and thenprovide the user with only the pertinent information.

In this work, we attempted to find the sweet spot betweenautomation and user intervention in the specific context of find-ing the best deal that matched the user’s criterion. We createdmyDeal, a system that automatically ranked deals according touser preferences and presented them to the user on their mobiledevice. Preliminary results show that the ranking algorithmwas able to combine both structured numeric discounts andunstructured (free-form text) promotion information accuratelyinto a single score while factoring in user preferences.

However, a true test of the system lies in validating whethermyDeal allows real users to find the best deal with highconfidence. We are currently in the midst of conducting fieldtests at a real shopping mall. We hope to understand fromthe study whether users are able to find the most relevant deal

when using our system and if they like myDeal and find it easyand quick to use. We also hope to validate the effectivenessof the UI component of myDeal and show that our UI ismore accurate (users can find the most interesting promotionsbetter) and faster to use than two other common UI designs(1. all promotions just listed alphabetically, and 2. promotionalinformation hidden under various categories) for these typesof shopping applications. We further plan to investigate theuse of different scoring algorithms and develop mechanismsto preserve user privacy even when using an external serviceto perform the matching.

REFERENCES

[1] S. Moriarty, N. Mitchell, and W. Wells, Advertising: Principles andPractices. Upper Saddle, New Jersey: Pearson Education, Canada,Ltd, 2009.

[2] L. Deng and L. P. Cox, “LiveCompare: grocery bargain hunting throughparticipatory sensing,” in Proceedings of the 10th Workshop on MobileComputing Systems and Applications (HotMobile), Santa Cruz, CA,Feb. 2009.

[3] Citi Shopper (SM), http://citishopper.citi.com/, Citibank Singpore Ltd.

[4] C. Narayanaswami, D. Coffman, M. C. Lee, Y. S. Moon, J. H. Han,H. K. Jang, S. McFaddin, Y. S. Paik, J. H. Kim, J. K. Lee, J. W. Park,and D. Soroker, “Pervasive symbiotic advertising,” in Proceedings ofthe 9th workshop on Mobile computing systems and applications,ser. HotMobile ’08. New York, NY, USA: ACM, 2008, pp. 80–85.[Online]. Available: http://doi.acm.org/10.1145/1411759.1411781

[5] M. C. Sala, K. Partridge, L. Jacobson, and J. B. Begole, “An explorationinto activity-informed physical advertising using pest,” in Proceedingsof the 5th international conference on Pervasive computing, ser.PERVASIVE’07. Berlin, Heidelberg: Springer-Verlag, 2007, pp. 73–90.[Online]. Available: http://dl.acm.org/citation.cfm?id=1758156.1758163

[6] J. Yan, N. Liu, G. Wang, W. Zhang, Y. Jiang, and Z. Chen, “Howmuch can behavioral targeting help online advertising?” in Proceedingsof the 18th international conference on World wide web, ser. WWW’09. New York, NY, USA: ACM, 2009, pp. 261–270. [Online].Available: http://doi.acm.org/10.1145/1526709.1526745

[7] B. Kim, J.-Y. Ha, S. Lee, S. Kang, Y. Lee, Y. Rhee, L. Nachman, andJ. Song, “Adnext: a visit-pattern-aware mobile advertising system forurban commercial complexes,” in Proceedings of the 12th Workshopon Mobile Computing Systems and Applications, ser. HotMobile ’11.New York, NY, USA: ACM, 2011, pp. 7–12. [Online]. Available:http://doi.acm.org/10.1145/2184489.2184492

[8] http://www.shopkick.com/, Shopkick.

[9] S. Nath, F. X. Lin, L. Ravindranath, and J. Padhye, “Smartads: bringingcontextual ads to mobile apps,” in Proceeding of the 11th annualinternational conference on Mobile systems, applications, and services,ser. MobiSys ’13. New York, NY, USA: ACM, 2013, pp. 111–124.[Online]. Available: http://doi.acm.org/10.1145/2462456.2464452

[10] Froogle, http://www.google.com/products#t:0, Google Limited.

[11] Mobile Coupons, http://www.mobiqpons.com/, Mobiqpons.

[12] E. N. Cinicioglu, P. P. Shenoy, and C. Kocabasoglu, “Use ofradio frequency identification for targeted advertising: A collaborativefiltering approach using bayesian networks,” in Proceedings of the9th European Conference on Symbolic and Quantitative Approachesto Reasoning with Uncertainty, ser. ECSQARU ’07. Berlin,Heidelberg: Springer-Verlag, 2007, pp. 889–900. [Online]. Available:http://dx.doi.org/10.1007/978-3-540-75256-1 77

[13] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl, “Itembased collabora-tive filtering recommendation algorithms,” in WWW, 2001, pp. 285–295.

[14] POS-tagger, http://ida.hum.ku.dk/tools/index.php/?lang=en, Center ForSprogteknologi.

[15] D. Lin, “An information-theoretic definition of similarity,” in ICML.Morgan Kaufmann, 1998.

[16] T. Pedersen, “Information content measures of semantic similarityperform better without sense-tagged text,” in NAACL, ser.HLT, 2010. [Online]. Available: http://portal.acm.org/citation.cfm?id=1857999.1858046

2014 IEEE International Conference on Pervasive Computing and Communications Work in Progress

202