mtat.03.319 business data analytics - courses.cs.ut.ee · lecture 7: cross selling & upselling...
Post on 07-Jul-2018
221 Views
Preview:
TRANSCRIPT
Tips
• Youalreadyknowaboutfollowingtips• “Thecostofacquiringanewcustomerisoftenaround4timesmoreexpensivethanitistoselltoanexistingcustomer.“
• Somethingnew:• Themostsuccessfulbusinesspracticestoachievethisarebyup-selling andcross-selling.
CrossSelling:Definition
• Cross-sellinvolvesthesaleofmultipleproductsofferedbyasingleproduct/serviceprovidertoaneworexistingcustomer.
TipsForCross/Upselling?
CrossSelling• PeersAlsoBought• Incentives• OnProductCopy• DiscountedSecondBuy• BuildARelationshipAndThenAsk
UpSelling• Sellthebenefitsoftheup-sell• KeepTheUp-SellBelow25%OfTheOriginalOrder
• HighlightYourUp-sell
ReturnofCross/UpSellingStrategy?
• Amazonreportedlyattributesasmuchas35percentofitssalestocross-sellingthroughitsoptionsoneveryproductpage
• “customerswhoboughtthisitemalsobought”and• “frequentlyboughttogether”.
HowtoIncreaserevenuesfromexistingornewcustomers?
Source:https://www.slideshare.net/NYCPredictiveAnalytics/building-a-recommendation-engine-an-example-of-a-product-recommendation-engine
Up-selling
Cross-selling
Howtosolvethisproblem?
• Whatproductstorecommendtowhom ?
• Solution:Technicalbasedapproach
• Atwhatstage ofthebrowsingprocesstoshowotheroptions?
• Solution:Psychologicalbasedapproach
• Outofscopeofthiscourse.
Whatproductstorecommendtowhom ? :RecommenderSystems
GoalofaRecommenderSystem:Identifyproductsmostrelevanttotheuser(Eg.Topnoffers).
QuizWhatareusersandmatchingitemsthefollowingcases:
a.)LinkedIn
b.)Facebook
c.)Amazon
d.)Netflix
(Users: members, Items: jobs)
(Users: members, Items: members)
(Users: members, Items: products, e.g., books)
(Users: members, Items: movies, TV shows)
TypesofRecommendationSystems
• PopularityBasedSystem• Classificationbased• CollaborativeFiltering
• NearestNeighbor(rememberKNNclassificationtechnique?)• MatrixFactorization(wewillnotcover)
Solution1:PopularitybasedRecommenderSystem
Recommenditemsviewed/purchasedbymostpeopleRecommendations:Rankedlistofitemsbytheirpurchasecount
QuizWhichofthefollowingistrueofapopularitybasedrecommendersystem?
CangeneratePersonalizedRecommendations?
CanuseContext(Eg.timeofday)?
CanuseUserFeatures?
CanuseItemFeatures?
CanusePurchaseHistory?
IsitScalable?
Solution2:ClassificationModelUse features of both products as well as users in order to predict
whether a user will like a product or not.
UserFeatures(Eg.Age,Gender)
ProductFeatures(Eg.cost,quality)
PurchaseHistory
Classifier LikeorNot?
Limitation:Difficulttocollecthighqualityinformationaboutproductsandusers.
QuizWhichofthefollowingistrueofaClassificationmodelbasedrecommendersystem?
CangeneratePersonalizedRecommendations?
CanuseContext(Eg.timeofday)?
CanuseUserFeatures?
CanuseItemFeatures?
CanusePurchaseHistory?
IsitScalable?
Solution3:NearestneighborCollaborativeFiltering
Item-based
Recommenditemsthataresimilartotheitemstheuserbought.
Similarityisbaseduponco-occurenceofpurchases.
“ItemsAandBwerepurchasedbybothusersxandy,sotheyaresimilar.”
User-based
Finduserswhohaveasimilartasteofproductsasthecurrentuser.
Similarityisbaseduponsimilarityinusers’purchasingbehaviour.
“UserxissimilartouserybecausebothpurchaseditemsA,BandC.”
Fig.Source:http://www.salemmarafi.com/code/collaborative-filtering-with-python/
SimilarUsers
• Considerusersxandywithratingvectorsrx andry• WeneedsimilaritymetricSim(x,y)• CapturetheintuitionthatSim(A,B)>Sim(A,C)
HP1 HP2 HP3 TW SW1 SW2 SW3
A 4 5 1
B 5 5 4
C 2 4 5
D 3 3
Users
Movies
SimilarUsers:Jaccard Similarity
• Jaccard similarity(A,B)=!"Ç!#!"È!#
• Jaccard distance=1- !"Ç!#!"È!#
• Sim(A,B)=1/5;Sim(A,C)=2/4• Sim(A,B)<Sim(A,C):Ignorestheratingvalues
HP1 HP2 HP3 TW SW1 SW2 SW3
A 4 5 1
B 5 5 4
C 2 4 5
D 3 3
Users
Movies
SimilarUsers:CosineSimilarity
• Cosinesimilarity(A,B)=Cos(𝑟%,𝑟&)• -1:dissimilar,0:orthogonal;+1:similar• Sim(A,B)=0.38;Sim(A,C)=0.32
• Sim(A,B)>Sim(A,C):butnotmuch
HP1 HP2 HP3 TW SW1 SW2 SW3
A 4 0 0 5 1 0 0
B 5 5 4 0 0 0 0
C 0 0 0 2 4 5 0
D 0 3 0 0 0 0 3
Users
Movies
NOTE:Fillemptyvaluesby0
Problem:Treatmissingvaluesasnegative
SimilarUsers:CenteredCosine
HP1 HP2 HP3 TW SW1 SW2 SW3
A 4 5 1
B 5 5 4
C 2 4 5 0
D 3 3
Users
MoviesNormalizedratingsbysubtractingtherowmean
Avg. Rat
10/3
14/3
11/3
6/2=3
HP1 HP2 HP3 TW SW1 SW2 SW3
A 2/3 0 0 5/3 -7/3 0 0
B 1/3 1/3 -2/3 0 0 0 0
C -5/3 1/3 4/3 0
D 0 0 0 0 0 0 0
Eachrowaddition=0
Ineachrow,originalvalue– Avg.Rat
Ratingsarecenteredaround0.+:userslikedit- :usersdidnotlikedit
SimilarUsers:CenteredCosine(2)Us
ers
Movies
HP1 HP2 HP3 TW SW1 SW2 SW3
A 2/3 0 0 5/3 -7/3 0 0
B 1/3 1/3 -2/3 0 0 0 0
C -5/3 1/3 4/3 0
D 0 0 0 0 0 0 0
• Sim(A,B)=0.09;Sim(A,C)=-0.56• Sim(A,B)>>Sim(A,C):butnotmuch
• Capturesintuitionbetter• Missingratingstreatedas“average”• Handles“tougherraters”and“easyraters”
Alsoknownaspearson correlation.
RatingPredictions
• Goal:PredictionforuserX anditemi• Whatweneed:
• Let𝑟'betheratingfortheuserX.• LetNbethesetofkusersmostsimilartoX,whohaverateditemi.
• Option1: 𝑟*+ =-.∑ 𝑟0+�0∈3 (Average)
• Option2:𝑟*+ =∑ 456!67�6∈8∑ 456�6∈8
(WeightedAverage)
Foraneighboryin(∈) thesetN
s isthesimilarityoftheuserxanditsneighbory
Item– ItemCollaborativeRating
• ForitemI,findothersimilaritems.• EstimateratingforitemIbasedonratingsforsimilaritems• Canusesomesimilaritymetricsandpredictionfunctionsasinuser-usermodel.
• 𝑟*+ =∑ 47:!5:�:∈8(7:5)∑ 47:�:∈8(7:5)
𝑠+= :similarityofitemsIandj𝑟*+:ratingsofitemi bytheuserxN(i:x):setofitemssimilartoi ,ratedbyuserx.
Item– ItemCollaborativeFiltering
? : Estimatetheratingofmovie1bytheuserC
Ratingsarebetween1to5Emptyboxes:unknownrating
Sim=PearsonCoeff.1) Subtractmeanrating𝑚+ fromeachmoviei.
1) 𝑚- =(1+3+5+5+4)/2=3.62) Row1=(-2.6,0,-0.6,0,0,1.4,0,0,1.4,0,0.4,0)
2) ComputeCosinesimilaritiesbetweenrows
RememberN=2Select2moviessimilarto1andratedbyuser5.
WeightedAverage=(0.41*2+0.59*3)/(0.41*0.59)
=2.6
𝑟*+ =∑ 47:!5:�:∈8(7:5)∑ 47:�:∈8(7:5)
PerformanceMetricforRecommendationSystems
RelevantItemsthatarealsorecommended
IrrelevantItemsthatarerecommended
Relevantitemsthatarenotrecomms
AllRecommendations(madeontrainingdataset)
AllRelevantItems(Allitemsinthetestset)
Precision:MeasureofExactness
Recall:MeasureofCompleteness
NumberofrelevantproductsbeingrecommendednumberItemsbeingrecommended.
#relevantproductsbeingrecommendedtotalnumberrelevantitems.
UsertoUserVsItemtoItem
• UsertoUser• Problem:Sparse:Usershavelimitedinterests(inbuying)
• Item-ItemoutperformsUser-User• UsersaremorecomplexthanItems• ItemshavelimitedgenrethanUsers.• ItemsimilaritymakesmoresensethanUserssimilarity
top related