An Energy-Efficient Mobile Recommender Systems Bingchun Zhu Dung Phan Hien Le February 22, 2011

Recommender Systems Agenda Introduction –Recommender System(RS) –Motivation –Problem formulation Algorithm Experiment Results Conclusion

Recommender Systems RS: What are they and Why are they RS: identify user interests and provide personalized suggestions. Enhance user experience –Assist users in finding information –Reduce search and navigation time Increase productivity Increase credibility Mutually beneficial proposition

Recommender Systems Types of RS Three broad types: 1.Content based RS 2.Collaborative RS 3.Hybrid RS

Recommender Systems Types of RS – Content based RS Content based RS highlights –Recommend items similar to those users preferred in the past –User profiling is the key –Items/content usually denoted by keywords –Matching “user preferences” with “item characteristics” … works for textual information

Recommender Systems Types of RS – Collaborative RS Collaborative RS highlights –Use other users recommendations (ratings) to judge item’s utility –Key is to find users/user groups whose interests match with the current user –More users, more ratings: better results –Can account for items dissimilar to the ones seen in the past too

Recommender Systems Types of RS: Hybrid Hybrid model The combination of two above models

Recommender Systems MOBILE Recommender System IS widely studied before BUT - Mostly based on user ratings - and is only exploratory in nature SO Unique features distinguishing mobile RS remains open The combination of two above models

Recommender Systems Motivation Taxi services are very popular and: –Energy consumption counted; –Successful story of drivers are different –Data related to individuals and objects are rich –Mobile RS provides users access to personalized recommendation anytime, anywhere

Recommender Systems To Provide more useful “local navigation” options High density of customer looking for the Cab THEN: Potential Travel Distance(PTD) LCP Skyroute algorithm

Recommender Systems Problem Definition Mobile sequential recommendation problem, which recommends sequential pickup points for Taxi driver to maximize his business success. Recommend a travel route for a Cab driver in a way such that the potential travel distance before having customer is minimized

Recommender Systems Problem Formulation

Recommender Systems Problem Formulation Assume a available set of N potential pick-up points: C = {C1, C2…Cn} And P={P(C1), P(C2)…, P(Cn)} is the probability set, where P(Ci) is estimated probability at each pick-up point.

Recommender Systems Problem Formulation is the set of directed sequences is the number of all possible driving routes where is the length of route is the probabilities of all pick-up points containing in

Recommender Systems Mobile Sequential Recommendation(MRS) Problem is PTD function Driver current position

Recommender Systems Problem Formulation

Recommender Systems Computational Complexity

Recommender Systems Sequential Recommendation Algorithm Potential Travel Distance Function (PTD) –an objective function which is used to evaluate condidate routes –property of PTD LCP algorithm –an algorithm which is used for pruning the search space offline SkyRoute algorithm –an algorithm for seeking optimal recommendation routes

Recommender Systems Recommended Driving Route c 1 : a pick-up event happens with probability p(c 1 ) c 2 : a pick-up event may happen with probabolity (1-p(c 1 ))p(c 2 ) –only when no pick-up event happens at c 1, this event happens.... c 4 : a pick-up event happnes with probability (1-p(c 1 ))(1-p(c 2 ))(1- p(c 3 ))p(c 4 ) An exapmel of Recommended Driving Route with the length of suggensted driving route L = 4

Recommender Systems Potential Travel Distance Function PTD is defined as the expected distance for a cab before picking up a customer in the route R L :

Recommender Systems PTD Function Property Lemma. The Monotone Property of the PTD Function –the PTD function is strictly monotonically increaseing with each attribute of vector DP. Vector DP is a vector combined by vector D and vector P With this property, it’s possible to determine a candidate route is better than the other without computing PTDs.

Recommender Systems PTD Function Property A recommended driving route R 1 with a length L, associated with the vector DP 1, dominates another route R 2 with a length L, associated with vector DP 2, iff the following two conditions hold: every element in DP 1 is not worse than it peer in DP 2 at lease element in DP 1 is better than its peer in DP 2 By this definition, if a candidate route A is dominated by a candidate route B, A cannot be an optimal route. element 2 B dominates A B dominates C element 1 A B C

Recommender Systems Constrained Sub-route Dominance

Recommender Systems LCP Pruning Algorithm LCP Pruning algorithm –For two sub-routes A and B with a length L, which includes only pick-up points, if sub-route A is dominated by sub-route B under Definition 2, the candidate routes with a length L which contain sub-route A will be dominated and can be pruned in advance.

Recommender Systems LCP algorithm prunes the search space offline –LCP algorithm will enumerate all the L- length sub-routes; –then prune the dominated sub-routes by difinition 2 offline. this pruning process can be done offline before the position of a taxi driver is known

Recommender Systems SkyRoute and its Property With lemma 4, if we can find skyline routes first, and then search the optimal driving routes from the set of skyline routes. This way can eliminate lots of candidates without computing the PTD function.

Recommender Systems Backward Pruning C5C5 C6C6 C7C7 PoCab R1R1 R2R2

Recommender Systems SkyRoute Algorithm Input C: set of pick-up points P: probability set for all pick-up points Dist: pairwise drive distance matrix of pick-up points L: the length of suggested drive route PoCab: current position Offline Processing (LCP) –Enumerate all sub-routes with length of L from C –Prune and maintain dominated Constrained Sub-routes with length L using sub-route dominance. –Maintain the remained non-dmonated sub- routes with length L, denoted as

Recommender Systems Online Processing –Enumate all candiate routes by connecting PoCab with each sub-route of –for i = 2: L-1 decide dominated sub-routes with i-th intermediate pick-up points and prune the corresponding candidates by using Backward pruning. update the candidate set by filtering the pruned candidates in above step. –end for –Select the remained candidate routes with length of L from the loop above –Final typical skyline query to get optimal skyline routes

Recommender Systems Keywords PTD function: - a function to compute the Potential Travel Distance before having a customer LCP algorithm: - a route pruning algorithm. - can be done offline before the position of a cab is known SkyRoute algorithm: - a route pruning algorithm - SkyRoute includes: + LCP offline pruning + Online pruning when the position of a cab is known

Recommender Systems Recommendation Process Obtaining the Optimal Driving Route: - Using LCP and SkyRoute for pruning candidates - Compute PTD function for all remaining candidates - Get the route with minimal PTD value Other challenge: How to make the recommendation for many cabs in the same area?

Recommender Systems Recommendation Process(cont.) Circulating mechanism - search k optimal drive routes - NO.1 route to the 1st coming empty cab - NO.2 route to the 2 nd coming empty cab - … - More than k empty cabs? Repeat from NO.1

Recommender Systems Experimental Data Real world data: - GPS location traces of approximately 500 taxis collected around 30 days in San Francisco Bay area - Number of pick-up points: 10 - Travelling distances between pick-up points are measured with Google Map API

Recommender Systems Experimental Data(cont.) Synthetic data: - Randomly generate pick-up points within a specific area - Generate pick up probability by a standard uniform distribution - Using Euclidean distance instead of driving distance - 3 sets: 10, 15, 20 pick-up points respectively

Recommender Systems Optimal Routes with Real World Data L=3: → C1 → C3 → C2 L=4: → C1 → C3 → C2 → C7 L=5: → C4 → C1 → C2 → C3 → C7

Recommender Systems Evaluated Algorithms BFS(Brute Force Search): - Compute the PTD value for all candidate routes - Find the minimum value as the optimal route LCPS (LCP Search) - Use LCP algorithm for offline pruning - Compute PTD for remained candidate routes - Get the minimum value as the optimal route SR(BNL)S: Sky Route + BNL (Block Nested Loop) - Using SkyRoute algorithm for pruning - Applying BNL for the remained candidates to get skyline routes SR(D&C)S: SkyRoute + D&C (Divide and Conquer ) - SkyRoute algorithm for pruning - D&C algorithm to get skyline routes

Recommender Systems Experiment Results A Comparison of Search Time - LCPS overperforms BFS and SR(D&C)S

Recommender Systems Experiment Results(cont.) Comparison of Search Time(L=3) on Synthetic Data Set

Recommender Systems Experiment Results(cont.) The pruning effect

Recommender Systems Experiment Results(cont.) Comparison of Skyline Computing

Recommender Systems Multi Evaluation Functions Skyline computing is time consuming Given a cab and fixed potential pick- up points: - Skylines are needed to compute only one time - Search space is pruned drastically => Skyline computing will have advantage with multi evaluation criteria

Recommender Systems Multi Evaluation Functions(cont.) Using 5 different evaluations (including PTD) Select 5 corresponding optimal drive routes

Recommender Systems Conclusion This paper developes an energy-efficient mobile recommender system for Taxi drivers. This system is able to recommend a sequence of potential pick- up points for a driver such that the potential travel distance before having customer is minimized. This paper provides a Potential Travel Distance(PTD) function for evaluating candidate sequences and two recommendation algorithms LCP and SkyRoute. LCP algorithm outperforms BFS and SkyRoute when searching for one optimal route. However, SkyRoute has better performance than BFS and LCP when there is an online demand for different optimal driving routes.

Recommender Systems THANK YOU !!! Questions??