Smart Geo-fencing With Location Sensitive Product Affinity
Smart Geo-fencing with Location Sensitive Product AffinityAnkur GargSunav ChoudharyPayal Bajaj Adobe Researchankugarg@adobe.comAdobe Researchschoudha@adobe.comStanford Universitypabajaj@stanford.eduSweta AgrawalAbhishek KediaShubham omAdobeshubhagr@adobe.comABSTRACTGeo-fencing is a location based service that allows sending of messages to users who enter/exit a specified geographical area, knownas a geo-fence. Today, it has become one of the popular locationbased mobile marketing strategies. However, the process of designing geo-fences is presently manual, i.e. a retailer must specify thelocation and the radius of area around it to setup the geo-fences.Moreover, this process does not consider the user’s preferencetowards the targeted product/service and thus, can compromisehis/her experience of the app that sends these communications. Weattempt to solve this problem by presenting a novel end-to-endsystem for automated design of affinity based smart geo-fences.Affinity towards a product/service refers to the user’s interest in aproduct/service. Our unique formulation to estimate affinity, usinghistorical app usage data, is sensitive to a user’s location and thus,the affinity is termed as location sensitive product affinity (LSPA).The geo-fence logic tries to capture contiguous groups of locationswhere the affinity high. Experiments on real world e-commercedataset reveals that geo-fences designed by our approach performssignificantly better at accurately targeting the users who are interested in a product. We thus show that, using historical app usagedata, geo-fences can be designed in an automated manner and canhelp enterprises target interested users with better accuracy ascompared to the present industry practices.CCS CONCEPTS Information systems Location based services; Geographicinformation systems; Mobile information processing systems; Clustering; Human-centered computing Ubiquitous and mobilecomputing;KEYWORDSGeo-fencing, Spatial Data Mining Workdone when the author was a part of Adobe ResearchPermission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page. Copyrights for components of this work owned by others than ACMmust be honored. Abstracting with credit is permitted. To copy otherwise, or republish,to post on servers or to redistribute to lists, requires prior specific permission and/or afee. Request permissions from permissions@acm.org.SIGSPATIAL’17, Los Angeles Area, CA, USA 2017 ACM. 978-1-4503-5490-5/17/11. . . 15.00DOI: 10.1145/3139958.3140059ACM Reference format:Ankur Garg, Sunav Choudhary, Payal Bajaj, Sweta Agrawal, Abhishek Kedia,and Shubham Agrawal. 2017. Smart Geo-fencing with Location SensitiveProduct Affinity. In Proceedings of SIGSPATIAL’17, Los Angeles Area, CA,USA, November 7–10, 2017, 10 pages.DOI: 10.1145/3139958.31400591INTRODUCTIONModern day mobile devices can provide a lot of contextual information of a user like current location, physical state, temperature,humidity, etc. that can aid a multitude of use cases. In the industry,location based marketing has become increasingly important due tothe expanding mobile user base. In this form of marketing, a brandtargets it’s mobile app users with an offer for a product/service,based on their geographical location. Geo-fencing is a locationbased marketing technique that allows brand marketers to pushoffers through in-app messages, location based coupons, real-timeupdates, etc. in specific geographical areas called geo-fences. Geofencing consists of two broad stages. The first stage is geo-fencedesign which comprises of a selection of key locations within anarea of interest and definition of virtual boundaries (known asgeo-fences) enclosing these locations. The second stage is real-timedetection which is about geo-fence deployment and testing for thepresence of mobile devices inside the deployed set of geo-fences inreal-time. The real-time detection problem has seen active interestfrom the research community [16, 37], but the geo-fence designproblem has not been addressed in depth. We only focus on thegeo-fence design problem in this paper.For location based marketing, the geo-fence design stage requiresa lot of manual effort on the marketer’s end, since he/she has to understand the usage patterns and user preferences in the area of interestby analyzing aggregated data. Thus, the marketer runs the risk ofdetecting only global patterns while missing out on individual andsegment level preference patterns. Geo-fences so designed wouldlack personalization, lead to unnecessary targeting and degrade theuser experience offered by the brand’s app. Frequent unnecessarytargeting and non-personalized experiences could encourage theuser to opt out or uninstall the app altogether, both of which arehighly undesirable from the brand’s perspective. The paper focuseson addressing these challenges of geo-fence design by presentingan automated approach to design smart geo-fences based on locationdependent user affinity towards a product/service. The geo-fencesare termed smart since they are sensitive to user affinity.In [3, 8, 29, 34], the authors have shown that a user’s productbrowsing behavior coupled with his present location gives a strongindication of his interest in that product at that location. In the
SIGSPATIAL’17, November 7–10, 2017, Los Angeles Area, CA, USA# Users3000200010000510152025303540# Distinct Products BrowsedFigure 1: Distinct Products viewed by UsersTable 1: Number of Distinct Stay points visited by users# Distinct Stay Points# Users (out of 4000) 1010-25 253966295sequel, we refer to both products and services as products. We termthis user interest in a product at a particular location, as locationsensitive product affinity (LSPA). It is important to note here thatthe affinity captures a long-term interest of the user towards aproduct and does not consider the temporal nature of such interestswith respect to the time of day. Estimating the affinity of a userat different locations towards a product is inherently complex dueto sparsity in the location based data collected [3]. By sparsitywe mean that, users generally browse a small number of products(Figure 1) and visit only a small number of locations (Table 1).Thus, we propose a formulation to estimate this affinity for a largenumber of products and at various locations through a sequentialutilization of pair-wise product-product, user-user, and semanticlocation-location similarities to remove most of the sparsity beforethe affinity based geo-fence design step. Location semantics refers tothe distribution of types of places in a small area around a particularlocation. We then identify segments of users having similar affinitiesover the area in which the geo-fences are to be designed and proposea systematic method to this design problem. The method tries tocapture locations with high user affinity towards the product andthus, the resulting geo-fences are personalized to each user segment.The main contributions of the paper can be summarized as follows:(1) We design a novel end-to-end system for automated design ofaffinity based smart geo-fences with several desirable properties. Firstly, the system generates geo-fences that are both product specific and personalized to user segments for improvedtargeting. Secondly, dependence on and collection of data is limited to locations where users have browsed for products withstrict adherence to location privacy policies [9, 22]. Lastly, thenumber and size of geo-fences designed are alterable to achievegood scalability of the location based targeting system usingthese geo-fences. We demonstrate the superior performanceof our system, compared to the current industry practice ofmanual geo-fence design by domain experts, on a real-worldaggregated e-commerce dataset.Garg et. al.(2) We model the latent location sensitivity in a user’s affinitytowards browsing for various products. We experimentallydemonstrate that ignoring location sensitivity in user affinitymodeling degrades the quality of the designed geo-fences. Thislocation semantics based latent sensitivity model is a majorcontributor to the superior quality of geo-fences designed byour system.(3) We mitigate the extreme sparsity problem that is typical in location based datasets (see Figure 1 and Table 1). Our three-stepuser affinity modeling approach sequentially utilizes pair-wiseproduct-product, user-user, and semantic location-location similarities to remove most of the sparsity before the affinity basedgeo-fence design step. We demonstrate that without such mitigation of the extreme sparsity in the affinity estimation process,the geo-fence design method performs poorly. This is true ofboth our approach and other affinity estimation approacheslike collaborative filtering.The rest of the paper is organized as follows. Section 2 brieflypresents the nature of the dataset and describes the problem addressed in the paper. Sections 3 and 4 develop the solution approachin detail, respectively describing the affinity model and the usersegment based geo-fence design. Section 5 describes the experimental setup and results. Section 6 describes prior art and Section 7concludes the paper.2PROBLEM DESCRIPTIONThe nature of the dataset is an important aspect towards understanding both the problem at hand and the subsequent solutionapproach. In the first half of this section, we explain the natureof the dataset and the preprocessing applied to it with full detailsdeferred to Section 5. In the second half of this section, we give aconcrete definition of our problem and an overview of the solutionapproach.2.1Nature of DatasetOur dataset consists of aggregated usage data from e-commerceretailers’ mobile application. This data contains logs of all interactions that the users have with these mobile applications. Out of allpossible interactions, we are only interested in the ones where theusers browse a product. Additionally, for each browsing activitydone by the user, his location is also recorded in the form of GPScoordinates ( latitude, longitude tuple). E-commerce retailers generally offer a range of related products. The set of these productsare organized in a multi-level hierarchy.2.2PreprocessingHere we give some preliminary information that will be neededto understand the rest of the paper. The first one relates to ournotion of location. As pointed by [15], the number of different GPSpoints can become very large and they seem to have little semanticmeaning. To counter these problems, they introduce the concept ofstay point. A stay point is defined as a geographical region where allthe locations, recorded for a user lie within a certain radius. Thereis also a time duration threshold, within which all the locationsneed to be recorded, for them to be a part of the same stay point.In our case, we use the radius of 100 meters and the time duration
Affinity based Geo-fencingSIGSPATIAL’17, November 7–10, 2017, Los Angeles Area, CA, USAUsageLogsSet of Points outsideArea of InterestStay PointDetectionAffinity ModelSet of Points insideArea of InterestGrid SquareFigure 2: Pictorial Representation of the set of all stay pointsL (blue circles and red triangles), grid squares in area of interest G, and stay points outside the area of interest L 0 (bluecircles).SemanticLocation ProductSimilarityBasedAffinitySet ofGeofencesof 5 minutes. This helps to capture a small region where the userhas stayed for a while and carries a semantic meaning. The staypoint is represented by the mean of GPS coordinates that lie insideit and the location semantics of the stay point models the types ofplaces in the area. In our solution, we assign a single stay pointfor all consecutive transactions, of a user, that occur within a fixedtime duration and the recorded GPS coordinates lie inside a fixedradius. All future references to location in this paper mean the staypoint assigned to each GPS coordinate in our database, using thestay point detection algorithm in [15].The main problem that we attempt to solve in this paper can thus,be summarized as: How to algorithmically design smart geo-fences,in the area of interest, to improve the response of users to offers sentfor a particular product?2.32.4Problem DefinitionWe define the set L, which is the set of all stay points that has beendetected in the dataset (blue circles and red triangles in Figure 2).In the current geo-fencing workflow, the marketer decides an areaof interest, like a part of the city, in which the offer for a particularproduct needs to be promoted; identifies key locations in that areaand marks a circular region around each one of them as geo-fences.This is done in order to attract users’ attention towards that productand evoke response to the offer in the form of browsing that productand/or purchasing it. We consider this area as a grid divided intosmaller grid squares (dotted green squares in Figure 2). The sizeof the grid squares is taken in such a way that it is very small incomparison the whole grid and we can assume the affinity to remainconstant within a grid square. Let G be the set of grid squares inthe area of interest and M(д) where д G is a set of all locationsl L that lie inside the grid square д. It is important to note that,by defining the area of interest, we do not mean that the datasetalso needs to be constrained to browsing activities inside that area.For disambiguation, we also define a set L 0 (set of blue circles inFigure 2) containing all locations not belonging to any grid squarein G.Further, let U represent the set of users and P be the set ofproducts that the e-commerce retailer provides. Then for each pair(u, p) U P, the usage logs consists of the values B(u, p, l) browse cnt where browse cnt represents the number of times theuser u browsed the product p at location l L. B(u, p, l) 0 if theu has not browsed p at that l. In our dataset, products are arrangedin a 3-level hierarchy where each product can be expressed in theform of a 3-tuple category, sub-category, vertical . As an example,the taxonomy for Orange Juice can be Food, Juices, Orange Juice Geo-fence DesignFigure 3: Solution FrameworkOverviewFigure 3 gives the overview of our solution with more detaileddescriptions in Sections 3 and 4. Our solution has two major components: Affinity Model: Precise knowledge of a user’s product and location specific affinity is useful to design high quality geo-fences.However, affinities are not explicitly observable and need to beinferred from the data at hand. The task is further complicated bythe extreme sparsity problem in the dataset which has its roots inthe tendency of users to browse only some available products andthat too at few locations (see Figure 1 and Table 1). To effectivelyestimate affinities and mitigate the extreme sparsity problem,we present our affinity modeling approach in Section 3. We relyon utilizing three types of latent similarities in the dataset in asequential manner (viz. product-product, user-user, and semanticlocation-location similarities) to estimate affinity values for userstowards different products at all grid squares inside the area ofinterest. The affinity obtained after employing this three stepapproach is what we term as Location Sensitive Product Affinity(LSPA). The experiments in Section 5 demonstrate that all ofthese latent similarities improve the quality of geo-fencing. Geo-fence Design: The second component is responsible for designing geo-fences for a specific product that the marketer wantsto target to his users. In existing systems, a single set of geofences are designed for a specific product, i.e. same geo-fencesfor all users. We introduce the aspect of personalization in thegeo-fence design by first identifying a small number of user segments. Users belonging to the same segment have similar affinityvalues towards that product at all the grid squares inside thearea of interest. Afterwards, we design geo-fences for each of
SIGSPATIAL’17, November 7–10, 2017, Los Angeles Area, CA, USAGarg et. al.the identified user segments such that it captures the areas ofhigh affinity for users in that segment. Thus, for a given product,we get a set of user segments and for each user segment, we geta set of geo-fences as the output from our approach.according to the indexing notation employed by MATLABr . Wenow define sim P (p, p 0 ) between any two products p, p 0 P as aweighted average of cosine similarities sim P p, p 0 , w C · simcos B C Cat(p), : , B C Cat p 0 , : w SC · simcos B SC sCat(p), : , B SC sCat p 0 , : w V · simcos B V Vert(p), : , B V Vert p 0 , : ,(2)3 AFFINITY MODEL3.1 Product-Product Similarity based AffinityIt is reasonable to expect that two similar products would commandsimilar affinities from a typical user at most locations. Indeed, inthe context of recommender systems, a high preference towardsa product has been used as a proxy to estimate interest in otherrelated products [32]. We capture this effect by a custom metricsim P (p, p 0 ) that measures similarity between products p, p 0 Pwith multi-level taxonomy information. Thereafter, we use thiscustom similarity measure to define a function AI (u, p, l) [0, 1]over all triplets (u, p, l) U P L, and designate it as theintrinsic affinity of user u towards product p at location l. Intuitively,AI (u, p, l) helps in estimating affinity values at 3-tuples (u, p, l)where B(u, p, l) 0 by utilizing information from 3-tuples (u, p 0, l)where B(u, p 0, l) , 0.The construction of sim P (·, ·) that follows is inspired from theproduct taxonomy information in the dataset and item-item similarity models [28]. Each product admits a 3-level taxonomy consisting of category, sub-category, and vertical. Let us respectivelydenote by C, S and V, the sets of categories, sub-categories andverticals in the dataset. Further, for any product p P, let Cat(p),sCat(p) and Vert(p) respectively denote the category, the subcat P U egory, and the vertical for p. We define matrices B P Z ,B C [0, 1] C U , B SC [0, 1] S U and B V [0, 1] V U asÕB P (p, u) ,B(u, p, l), (p, u) P U, (1a)l LÕB P (p, u)p:Cat(p) cB C (c, u) , Õ,Õ (c, u) C U, (1b)B P (p, u)c 0 C p:Cat(p) c 0ÕB P (p, u)p:sCat(p) sB SC (s, u) , Õ,Õ (s, u) S U, (1c)B P (p, u)s 0 S p:sCat(p) s 0ÕB P (p, u)p:Vert(p) vB V (v, u) , Õ,ÕB P (p, u) (v, u) V U, (1d)v 0 V p:Vert(p) v 0where B(u, p, l) denotes the browse count as described in Section 2.3.We note that, by definition, the summation operator equivalenceÕÕÕÕÕÕÕ p Pc 0 C p:Cat(p) c 0s 0 S p:sCat(p) s 0v 0 V p:Vert(p) v 0is true. Intuitively, B C (:, u) represents a vector of browsing intensity for the user u that is aggregated across all locations in L andnormalized w.r.t. categories in C. Analogous intuitions are true forboth B SC (:, u) and B V (:, u). Note that we have used the ‘:’ symbolwhere simcos (·, ·) denotes the cosine similarity function betweentwo equal length vectors and weights w C , w SC and w V are nonnegative and chosen to satisfy w C w SC w V 1.Finally, we define the intrinsic affinity function as the weightedaverage!!Õsim P (p, p 0 )B(u, p 0, l)ÕÕAI (u, p, l) , . (3) ·sim P p, p 00B u, p 00, lp0 Pp 00 P3.2p 00 PUser-User Similarity based AffinityPrior work in friend recommendation [36] and many other recommendation systems [26] suggests that large sets of users exhibit similarities in their behavioral preferences and this phenomenon couldbe utilized by various recommendation systems to improve prediction accuracy with limited training data. Analogous to Section 3.1,we will capture the similarity between users u, u 0 U with a custom metric sim U (u, u 0 ) and use this similarity measure to define afunction AU (u, p, l) [0, 1] over all triplets (u, p, l) U P Lthat we term as the user similarity based affinity of user u towardsproduct p at location l. Intuitively, AU (u, p, l) helps in estimatingaffinity values at 3-tuples (u, p, l) where B(u, p, l) 0 by utilizinginformation from 3-tuples (u 0, p, l) where B(u 0, p, l) , 0.Similarity of user preferences is further connected to overlapsimilarity in frequented locations [15, 33]. Prior to constructingsim U (·, ·), we model this location based aspect of user similaritywith semantic representations (see [33]) for each location as comM denote theputed by the Google Places API [12]. Let Sem(l) R semantic vector for location l L, where M types of places havebeen identified. Intuitively, Sem(l) represents a distribution overthe M types of places, constructed from places that are geographically close to l. We illustrate this with an example. Suppose thatwe have M 5 types of places in our dataset, viz. restaurant, hotel,hospital, bank, and place of worship. Further, assume that thereare 10 places with identifiable types that are close to a location lof which 7 are restaurants, 2 are hotels and 1 is a bank. Then wewould have Sem(l) (0.7, 0.2, 0.0, 0.1, 0.0). The actual length ofSem(l) obtained from Google Places API [12] is M 100. We nowdefine the semantic similarity function sim L (l, l 0 ) over all l, l 0 Las sim L l, l 0 , simcos Sem(l), Sem l 0 .(4)Next, we define a hierarchical structure on elements of L toinform the construction of sim U (·, ·). We draw on the approachtaken in [15] where locations were hierarchically clustered forsimilarity matching between location sequences. We represent eachM andl L in its semantic vector representation Sem(l) R employ the DIANA algorithm [18] with the Euclidean distance
Affinity based Geo-fencingSIGSPATIAL’17, November 7–10, 2017, Los Angeles Area, CA, USA1.0C10.8Products browsed by u at L1: {p1, p2}0.6Products browsed by u’ at L’1 , L’2: {p3}L1L’1L’2C20.40.2C3Stay points ofuser u’Stay points ofuser uStay points ofother users0.0Figure 5: Heatmap depicting the difference between Location Sensitive Product Affinity distribution (over area of interest) of the same user for two different productsFigure 4: View of a clustered layermetric on these semantic vectors to derive a hierarchically clusteredstructure for all l L. The cut point of the hierarchy is decided bythe use of the Gap Statistic value [31]. Let H denote the set of alllayers in this hierarchically clustered structure between the optimalcut point and the root. For any layer h H , let C(h) denote the setof clusters formed at that layer. An example of a clustered layer isshown in Figure 4.Finally, we describe the construction of sim U (·, ·). Let us denote by sim U,C (·, ·, ·) and sim U,L (·, ·, ·), two intermediate similarity measures, that are defined as follows. Let C 0 C(h) be a clusterat layer h and let Pu, C0 P denote the subset of products browsedby user u U within locations l C 0 . If Pu, C0 and Pu 0, C0 are bothnonempty sets for some u, u 0 U, then we define the similaritymetric sim U,C (u, u 0, C 0 ) between users u and u 0 w.r.t. cluster C 0 asÕÕ sim P p, p 0p Pu, C0 p 0 Pu 0, C0 sim U,C u, u 0, C 0 ,.(5)Pu, C0 · Pu 0, C0 Ñ We let C 00 (h) , Γ C(h) Pu, Γ , Γ C(h) Pu 0, Γ , .If C 00 (h) is nonempty, we further define the layer level similaritymetric sim U,L (u, u 0, h) between users u and u 0 w.r.t. layer h H asthe average of all cluster level similarity metrics within that layerÕ sim U,C u, u 0, C 0 C 0 C 00 (h)sim U,L u, u 0, h ,,(6) C 00 (h) and in case C 00 (h) , we set sim U,L (u, u 0, h) 0. A weighted sumof the layer level similarities is used to define the user similaritybased affinity sim U u, u 0 H Õ3.3Semantic Location-Location Similaritybased AffinityAs discussed in Section 3.2, semantic similarity between locationssim L (·, ·) can be used to model similarity of preferences acrossdifferent users sim U (·, ·). Additionally, semantic location similarity should be exploitable to infer similarity of preferences for thesame user across locations. We do so by extending the definitionof the Sem(·) to operate on grid squares д G (shown in Figure 2) and defining a function AS (u, p, д) [0, 1] over all triplets(u, p, д) U P G that we term as the location semantics basedaffinity of user u towards product p at grid square д. Intuitively,AS (u, p, д) helps in estimating affinity values at 3-tuples (u, p, l)where B(u, p, l) 0 by utilizing information from 3-tuples (u, p, l 0 )where B(u, p, l 0 ) , 0.Since locations in L are actually stay points (see Section 2.2),the definitions of the location semantic function Sem(·) and thesemantic location similarity function sim L (·, ·) can be extendedunchanged to operate on areas that are larger than a stay point,e.g. Sem(д) can be computed for grid squares д G. For any gridsquare д G, let t(д) L 0 denote the set of n locations in L 0that achieve the highest values for the partially defined semanticsimilarity function sim L (д, ·). We restrict t(д) to be a subset of L 0to smooth the affinity estimate against local effects and we restrict t(д) to not exceed n to bound the computational complexity. Wedefine the location semantics based affinity function as the weightedaverageÕÕ sim L д, l 0 · AU u, p, l 0 AU (u, p, l)AS (u, p, д) , β(h) · sim U,L u, u 0, h ,l 0 t (д)l M(д)Íl 0 t (д) sim L (д, l0) . M(д) (7)(9)where the weights β(h) are selected as in [15].The user similarity based affinity function AU (u, p, l) is nowdefined as the weighted average!Õ sim U (u, u 0 )ÕAU (u, p, l) ,· AI u 0, p, l . (8)00 sim U u, uu 0 UWe define the vector valued LSPA function A(u, p) [0, 1] G foreach pair (u, p) U P by collecting all values of AS (u, p, д) overд G into a vector. Figure 5 shows the heatmap of the differencevector A(u, p) A(u, p 0 ) over G for a particular user u in our datasetwho uses two different products p and p 0 and illustrates that theLSPA function can vary significantly.h 1u 00 U3.4Location Sensitive Product Affinity (LSPA)
SIGSPATIAL’17, November 7–10, 2017, Los Angeles Area, CA, USAAlgorithm 1 CreateUserSegments(G, U, A, p)Output: U serSeд(p), Aavд (p)Steps:max ch val 0.0. Holds max. CH-value obtainedk optimal 0. Holds k that gives max. CH-valueU serSeд(p) {}. user segments at k optimalfor k 1 to U dochk , user seд K-Means(A, U, p, k)if chk max ch val thenmax ch val chkk optimal kU serSeд(p) user seдend ifend for. Calculate average affinity distribution for user segmentsfor user seд U serSeд(p) dofor д G doÍAavд (p, user seд, д) end forend forreturn U serSeд(p), Aavд (p)A(u,p,д) user seд u us er s eдGarg et. al.Algorithm 2 DesignGeofence G, U serSeд(p), Aavд (p), p, thresh Output: Geo f ence(p)Steps:for user seд U serSeд(p) do. Find of grids having average affinity greater than threshдrid set {}for д G doif Aavд (p, user seд, д) thresh thenÐдrid set дrid set {д}end ifend forдrid clusters DBSCAN(дrid set) . DBSCAN ClusteringGeo f ence(p, user seд) {}for clust дrid clusters do. Create boundary around each cluster of grids to formgeo-fenceдeo f ence α-hull(cluster )Geo f ence(p, user seд) ÐGeo f ence(p, user seд) дeo f enceend forend forreturn Geo f ence(p)4 GEO-FENCE DESIGN4.1 User SegmentationWith the help of the affinity distribution we can design geo-fencesfor all users. But designing separate geo-fences for each and everyuser would result in an unmanageably large set of geo-fences. Geofencing is an expensive setup as it requires deploying, maintainingand managing infrastructure to track devices at scale in real time.On the other hand, we can do better than designing a single geofence per product that potentially results in poor user experienceon account of mis-targeting, i.e. unnecessarily targeting users whomight not be interested in the product at that location. Thus, wepropose the idea of identifying small number of user segments anddesigning geo-fences for each segment separately.Algorithm 1 explains the approach that identifies the user segments for a particular product p. We perform K-Means clusteringwhere each user is represented by LSPA vector A(u, p) for a product p. We generate clusters using different values of K and choosethe one that gives the maximum CH-value [7]. This gives us the setof user segments U serSeд(p) for the product p and we design geofences for each of these segments separately. Also, we compute theaverage affinity distribution for each segment, Aavд (p, user seд),which represents the location sensitive product affinity for that segment. We get Aavд (p) by collecting all values of Aavд (p, user seд)over U serSeд(p).4.2Geo-fence DesignThe process of geo-fence design is implemented by the series ofsteps given by Algorithm 2. For each user segment, we deploy thestrategy: select all grid squares from the average affinity distributionhaving affinities above a certain threshold thresh, cluster nearbysquares and create a boundary around each of the clusters to get agroup of geo-fences. Now we discuss the clustering and boundarycreation steps in detail.4.2.1 Clustering Nearby High Affinity Locations. Clustering ofgrid squares having high affinity (greater than thresh) ensures thatthe area covered by a geo-fence is sizable. This allows the geofence to be actually useful in achieving the desired purpose i.e. helpretailer target potential customers. Also, since we don’t put anyrestriction on the size and shape of geo-fence, so the DBSCAN [11]algorithm is used for clustering which has the ability to find arbitrarily shaped clusters and is robust to outliers.4.2.2 Constructing Geo-fence boundary. The DBSCAN clustering gives clusters of grid squares as output from which the
based on their geographical location. Geo-fencing is a location based marketing technique that allows brand marketers to push o ers through in-app messages, location based coupons, real-time updates, etc. in speci c geographical areas called geo-fences. Geo-fencing consists of two broad stages. The rst stage is geo-fence
fencing, which has angle type and corrugated type; but security palisade fencing has only corrugated type. With high strength, sharp prongs and narrow pale spacing, the steel palisade fencing can achieve high level of security to prevent intruders. It's widely used for church fencing, residence fencing, factory fencing, road fencing and so on.
Figure 3.26: show the Camera Entry of Geo-Location Expenses system. 39 Figure 3.27: The Voice Entry for Geo-Location Expenses system 40 Figure 3.28: The QR-code Entry for Geo-Location Expenses system. 41 Figure 3.29: The different Record view filter by week, month, and year. 42 Figure 3.30: The full detail view of record in Geo-Location Expenses.
Chapter 11 Fencing Brian Freking Producers have numerous fencing options for the confinement and protection of livestock. Many traditional materials such as barbed and woven wire fences are suitable for fencing. However, new-er materials including high-tensile wire also should be considered prior to selecting a fencing type. The
DESIGN GUIDELINES PRINCIPLES SUMMARY 5. AROUND YOUR HOME Fencing to side and rear boundaries is mandatory. Front fencing is compulsory on allotments indicated on building envelope plan. Fencing Side & Rear - 1.8m high colourbond good neighbourhood fencing. - Fencing
Noble Contour Fencing can be expanded any time. . . . . . using either Noble Panels or more Noble Contour Fencing. 800-437-3966 Noble Contour Fence Noble Trainer Series Barn Noble CONTOUR FENCING is the ideal pasture or perimeter fence. It is HORSE SAFE, fittings that slide over each rail.File Size: 1MBPage Count: 8Horse Shelters · Horse Fencing · Horse Stalls & Barns · Round Pens & Arenas · Swing Gates
disputes with adjacent land owners. Remember that "good fences make good neighbors." Fencing Proper fencing is usually a major investment in the farm enterprise; therefore, the fencing layout should be carefully planned to save time and money. One of the goals of developing a good fencing system is to improve grazing efficiency.
Fencing Wire: Fencing Wire: Low Tensile Small Coils Products Shortname: LACW Size Range Available: Diameter (mm) Length (m) 5kg Coil Nominal Length (m) 10kg Coil Nominal 1.60 317 633 2.00 203 405 3.15 82 163 4.00 51 101 Applications: Diameter (mm) 1.60/2.00 Temporary electric fencing, tying and lacing wire 2.50 Electric fence wire
appropriate strategies to solve problems. Mathworld.com Classification: Number Theory Diophantine Equations Coin Problem 02-02. 14 AMC 8 Practice Questions Continued -Ms. Hamilton’s eighth-grade class wants to participate intheannualthree-person-teambasketballtournament. The losing team of each game is eliminated from the tournament. Ifsixteenteamscompete, howmanygames will be played to .
