A Recommendation Mechanism for Contextualized mobile advertising S

A Recommendation Mechanism for Contextualized mobile advertising S. -T. Yuan et al. , Expert Systems with Applications, vol. 24, no. 4, pp. 399 -414, 2003. Jongwon Yoon 2011. 05. 04

Outline • Introduction • Proposed method – Architecture – Vector-based Representation – Recommendation Mechanism • User stereotype KB and profile • Learn user profile • Recommendation function • Experiments – Experimental measurements – Experimental user types – Results • Summary 2

Introduction • Mobile advertising – One of fields in mobile commerce – Possible to target users according to user’s contexts • It is essential that fully personalized mobile advertising infrastructure • Proposed method – A personalized contextualized mobile advertising infrastructure for advertising the commercial/noncommercial activities (MALCR) – Contributions • 1) Interactive advertising with customized recommendation • 2) Provide a representation space • 3) Recommendation mechanism using implicit user behaviors 3

Proposed method Architecture • Learn users’ profiles from implicit browsing behaviors – Difficult to obtain direct keypad inputs for every request • Two ways of service – Pull mode : the dominating mode / requests recommendations – Push mode : provide SMS if permission from users is granted 4

Proposed method Vector-based Representation Space • Features in commercial/non-commercial advertisements Attributes Category Attribute values Wholesale and retail, arts and entertainments, others Day Weekdays, weekend Time A time slot(17: 00 pm before), B time slot (17: 00 pm after) Place Outdoors, indoors and formal, indoors and informal Fee Performer Free, fee-based Top celebrities, others • Mobile Ad representation n : total number of features mi : the number of possible values for ith feature • User profile representation W Iiaj : User’s interest in the jth value of ith feature 5

Proposed method Recommendation Mechanism • Concepts – 1) Minimize users’ inputs : Use implicit behaviors – 2) Understand users’ interests – 3) Top-N scored advertisements • Browsing interface to capture implicit behaviors – Behaviors : Clicking order, clicking depth, and clicking count 6

Proposed method: Recommendation Mechanism User Stereotype KB and Profile • User Stereotype KB – Used to expedite the learning of the users’ interests – Stores a variety of typical users’ interests – Initially pre-defined (see next slide) and adjusted during usages • User profile – Use multiple user stereotype vectors Rj : the ratio of the reference of the jth stereotype vector 7

Proposed method: Recommendation Mechanism An Example of User Stereotype KB 8

Proposed method Learn User Profile: Overview • Two-level neural networks approach – One-level : Requires explicit user scoring to train (Not appropriate for mobile devices) – Two-level neural networks • User_score NN (USNN) : Calculate score using user’s implicit behaviors • Preference_weight NN (PWNN) : Calculate preference weights for the certain Ad • Flow of user profile learning – 1) Obtain user scores – 2) Use the Ads and corresponding scores as training examples of PWNN – 3) Obtain preference weights – 4) Perform sensitivity analysis and update the user profile 9

Proposed method Learn User Profile: Usage of Two-level NNs • On the request of a new stereotype – Use pre-trained user stereotype vector and NN weights – Compute customized stereotype by training PWNN • PWNN structure • Use USNN to obtain user’s score as training examples : (M_AD, Score. U) • Pre-trained USNN generates reasonable score from the value of (O, D, C) • On the use of existing stereotype – Evolve the customized user stereotype vector by training PWNN 10

Proposed method Learn User Profile: Sensitivity Analysis • Purpose – To transform PWNN outputs into the vector-based representation • Process – 1) Calculate score for each input Xi : Each input value attribute. Scorei : The output value of PWNN – 2) Compute Score. Sum – 3) Compute the preference weights Wi : Preference weight of Xi in the user streotype 11

Proposed method Recommendation Function • Recommend top-N scored advertisements – Ranks Mobile Ads relevant to a designated location • Process – 1) Compute score for each Mobile Ads – 2) Rank the scores of all Ads – 3) Recommend Top-N Ads if in the Pull mode – 4) Push Top-1 Ads to the user if in the Push mode 12

Experiments Experimental Measurements • Averaged Score. U Growth – Score computed from a user’s implicit browsing behaviors – Shows how close the Top-N match the user’s interests • Instance precision, recall, and fallout – Using learned vector representation(Top-1) and target vector representation – Instance precision = Found/(Found + False alarm) – Instance recall = Found/(Found + Missed) – Instance fallout = False alarm/(False alarm + Correctly rejected) 13

Experiments Experimental User Types • Three types : 50 users in each type – First use (Login 0) ▶ 10 Trials (Login 1 ~ Login 10) • Extremely focused(U 1) – Interests are highly concentrated – A general query is generated only at Login 0 • Extremely Scattered (U 2) – 3 general queries are generated • Middle (U 3) – Two general queries are generated in each use from Login 1 to Login 5 – Assume that recommendations conform to the user’s interests after Login 5 14

Experiments: Results Stable Interests and User Type 15

Experiments: Results Unstable Interests but Stable User Type • Interests are randomly changed at Login 3 • Four situations – Implicit change and no(L 0)/yes(L 1) weighting on the most current stereotype – Explicit change and no(L 3)/yes(L 4) weighting on the most current stereotype 16

Experiments: Results Unstable Interests and User Type • User type changing – Login 1 -3 : Extremely Focused (U 1) / Login 4 -10 : Extremely Scattered • Interests changing – Randomly changed at Login 5 17

Summary • Proposed MALCR – Mobile advertising infrastructure – Furnish a new customized recommendations – Provide a representation space • vector-based Ad and user profile representations – Devise a recommendation mechanism • Two-level NNs • Future works – Testing advertising effect measurement • L is 1 if a user exerts after receiving Top-1 • L is 0 otherwise • T is the lapse of time between the push and exertion 18