Us ser Interf face Algo rithms
5 Us ser Interf face Algo rithms
In contra ast to most r research in h human ‐comp puter interac ction we see interaction d design and a algorithm design a s closely rela ated. Previo us work on c context ‐awa re user inter rfaces has fo ocused on ad daptation of user i interface ele ements in re sponse to co ontext chang ges. For exam mple, a mob bile device m may use a normal g graphical use er interface w when used d during a mee eting, but ma ay switch to a a simple men nu with a large fon nt and voice commands when placed d in a car. In n contrast to o such simple e context‐aw ware user
interface es, intent‐aw ware user int terfaces are by definitio on personalis sed and requ uire a more complex adaptati on model th hat relies on an explicit u user model t that reflects the user’s b behaviour an nd intent.
While th here is a lot of work on behaviour p profiling and d intent reco ognition mos st research e exhibits a sharp d istinction be etween HCI‐ ‐driven inte raction desi ign and syst tem ‐driven algorithm d esign. In contrast t, we see it as necessa ry for the d design of in ntent ‐aware user interfa aces to stro ongly link interacti ion design w with algorithm ms design. Fu urthermore, we see it as s essential th hat interactio on design lead alg gorithm desi gn and algo orithm desig gn is inform med by the requirement ts and cons strains of interacti ion design.
In order r to deal wit th the issue s identified above, this section disc cusses aspec cts of the p prediction system w which is par t of the use r interface m model. We l ike to point out that the e algorithms s operate above th he context da ata acquaint tance framew work in the G GAMBAS sys tem, which i is exploited t to gather histories s of travel b behaviour. T his is requir red to deriv e higher‐lev vel travel pa tterns data to make tempora al exploratio n of persona al and social travel infor mation poss ible. For this s purpose, ra aw travel events as sensed by the con ntext data a acquaintance e framewor rk is transla ated to hig gher ‐level knowled dge as it is re equired for b uilding inten nt ‐aware use er interfaces. . To clarify th his issue, we describe in this s ection for e each layer of f our user in nterface mo del a system matic approa ach to predi ct future travel be ehaviour an nd quality of f transport d data. As pa rt of this di scussion, we e provide a detailed specifica ation of the prediction a algorithms b based on fo rmal descrip ption of the input data which is processe ed and the fo orecasts whi ch are return ned as an ou utput of the p prediction al lgorithms.
5.1 Q Quality of T Transpor rtation: Da ata Minin ng and Pre ediction
The user r interface sy ystem exploi its prediction ns about the e quality of t ravel inform mation to incr rease the traveller rs’ awarenes s about the consequence es of their m mobility choic ces. The pred dictions are based on measure ements of th he crowd lev vels of past b bus rides, w hich allow u s to extrapo olate the sta te of the transpor rtation netw work into the future. The personal ide entity of use ers is not rele evant in this s context, and pers sonal data is not recorde ed in the cro wd histories s. The crowd measureme ents are com mposed of anonym ous statistica al informatio on about how w many peop ple have trav velled when and where.
5.1.1 I Input For the p prediction of f quality of t ransport dat ta, we proce ss static info ormation rela ated to the b bus route network k and timeta ble informat tion. This inf formation is usually mai ntained by t the bus prov vider and can be d defined as fo llows.
Definitio on (Bus Tran nsportation Network): A A bus transpo ortation netw work is form mally a tuple N=(S, L), where S S={s 1 , s 2 , …, s s n } denotes t the set of all l bus stops a and L={l 1 , l 2 , …, l n } repres sents the un nique bus lines of t the bus netw work.
Each bu s stops s i ∈ ∈ S thus rep presents a p possible entr ry and exit p point of the e bus system m, where traveller rs can board d buses. A b bus line l i ∈ L is schedu led to pick up passenge ers at these e stations Each bu s stops s i ∈ ∈ S thus rep presents a p possible entr ry and exit p point of the e bus system m, where traveller rs can board d buses. A b bus line l i ∈ L is schedu led to pick up passenge ers at these e stations
Definitio on (Time tab ble):
A time table is a tu uple T={(l k , s i i , , d i , s j , a j )}, w which lists fo or a trip wit h a given bus line l k ∈ L from bus stop s i ∈ ∈ S to the ne ext stop s j ∈ S the depart ture time d i (at stop s i ) a as well as the arriv val time a i (at t stop s j ).
The time e table thus encodes th e route of a a bus taken at specific t ime of a day y. s i and s j r represent neighbo ured stops on this rou te and thus s define a s shared route e segment ( (s i ,s j ). Also, since we distingui ish between n departure and arrival time for th he same sta ation, we ca an access s cheduled waiting t times at a bu us station. B Beyond the st tatic transpo ortation netw work data, w we are also r elying on dynamic c data about the traffic f lows in the b bus network k, which is co ollected by th he sensor in stalled in buses in real ‐time.
Definitio on (Crowd H History): A bus crowd h history is a tuple C={(l k k , s i , d i , s j , a a i , n)}, wher re n ∈ represen nts the num mber of trav ellers observ ved for a ri ide on a bu s segment ( (s i , s j ) with recorded
departur re time d i at stop s i ∈ S an nd arrival tim me a i at stop s j ∈ S.
Please n note that C contains for r every pair of consecut tive stops on n a bus rout te many cro owd level records for different t times of a d day. This is b because the crowd level history is a collection of f the raw sensing output from m the real‐wo orld. This dat ta needs fur ther post‐pr ocessing ste eps to get ins sight into
crowd le evel patterns s, which is su ubject to our prediction a algorithm.
5.1.2 O Output The targ get of our pr rediction of q quality of tra ansport info ormation is t he expected d crowdedne ess of the bus netw work. This ca an be formall ly defined as s follows.
Definitio on (Crowd P Prediction): T The predictio on of crowd levels can b be described d as a functio on f: T → , wher re f(l k , s i , d i , s s j , a j ) is the f forecast of t the expected d numbers o of travellers s for a given entry in
the bus time table, i.e., a trip w with bus line e l k from s i ∈ ∈ S departing g at time d i a and arriving g at s j ∈ S at time a a j .
As the d definition sho ows, the pre ediction is sp pecific for a particular b us line and t trip, since w we expect different t crowd leve ls for bus lin es and segm ments of the bus routes a t varying tim mes. Also, ple ease note that the result of the e prediction is in , mea aning that th he output ca an be a floati ting point nu umber to estimate e the future number of p passengers travelling a long the giv en bus segm ments.
5.1.3 P Prediction Algorithm
We will explore me ethods from the area of f time series s analysis to o tackle the prediction problem. These m methods are devised to d deal with nu umeric data showing cha aracteristic t trends over t time. For instance e, we can rea alistically ass sume that cr rowd levels significantly vary during the day and d reach a maximu m level at ru ush hour, wh hile being mu uch lower at non ‐rush ho our times. In n order to acc count for such pat tterns, mode els for time series analys sis are able to fit a mat hematical fu unction to gi iven data
observat tions. While not limited d to, we will l investigate the potent ial of the fo ollowing met thods for inferring g accurate cr owd predict ions:
Cen ntral tenden ncy algorithm ms can be u used to com pute expect ted values fr rom past sa mples of ran dom variabl es. For insta ance, a simp ple tendency y predictor i s the arithm metic mean, which is
easy y to implem ent and ofte en performs well enough h to achieve good result t. In order to o account easy y to implem ent and ofte en performs well enough h to achieve good result t. In order to o account
arit hmetic mea ns can also b be smoothed d over time.
ARM MA models are popular methods fo or regression analysis, wh hich can be employed to o fit time seri ies data to a given sto ochastic pro cess model for the sak ke of explor ratory data analysis.
Exa mples of th hese metho ods are auto oregressive functions, w which assum me that the ere exist dep pendencies b between con secutive cro owd level me easurements s. Such a pre ediction func tions can the n be used to o forecast a c crowd level a at a given tim me based on n the most re ecent measu urements rep orted by the e real‐time se ensor stream ms.
5.2 P ersonal Behaviour B r: Data Mi ining and Predictio on
A furthe er challenge is the discov very of the u user’s mobilit ty patterns i n the transp port systems s. While a transpor rtation syste em offers m any physica al routes to choose from m, a realistic c assumptio on is that there are e some part icular routes s with a high h popularity a and significa nce for a tra aveller. Such patterns in the tr ransport beh haviour arise from the re egular activit ties carried o out by peopl le over the c course of the wee k, such as co ommute trip ps to works a and leisure a activities to popular plac ces at the w week ‐end.
Motivate ed by this ob bservation, w we would lik ke to predict this informa ation to mak ke it availabl le for the design o of novel inten nt ‐aware use er interfaces. .
5.2.1 I Input Persona l user data i s managed s separately fr rom the bus network da ta in our sys stem. As enf forced by the arch itecture of t the GAMBAS S middleware e, this kind o of data is dir ectly collecte ed and store ed on the users’ pe ersonal mob bile devices t to protect th heir privacy n needs. In the e following, w we describe a formal
model o f the person al data we a ccess and an nalyse for ch aracteristic p patterns of t travel behavi iour.
Definitio on (Persona l Trip Histor ry):
A perso onal trip hist tory T={T 1 , T T 2 , …, T n } is denoted as a set of indepen dent user t rips. Each t rip T i ={r 1 , r 2 2 ,…r n } ∈ T c consists of a a set of bus s rides a us er takes. Formally y, each bus r ride is define ed as r=(b i , ( (s k , t k ), (s l, t l ) )), where b i ∈ L denotes s the bus lin e of that ride, s k i s the station n where the user got on the bus at t time t k and s s l refers to th he station w where the user got off the bus a at time t l.
The pers sonal trip his story models s information n about the user’s travel l choices wit thin the bus network. For this purpose, the e trip history y is segment ted into cohe erent user tr rips, where a a single user r trip lists all the b bus rides to g get from a s source to a f final destinat tion. This inc cludes all th e stations w where the user wa s required t o change th e bus. From m this inform mation, we c an derive im mportant beh havioural patterns s that help us s to find the most popula ar routes tak ken by the us ser.
5.2.2 O Output The trip history cont tains raw inf formation w which is not d directly mea ningful for t the purpose of travel planning g and user in nteraction. T he goal of th he predictio n is to trans late this info ormation to the most meaning gful travel pa atterns relev ant in the co ontext of a fu uture travel s situation.
Definitio on (Persona l Mobility G Graph):
ersonal A p mob bility graph G P =(V P , E P ) consists of a set of stations V P ⊆ S visite ed by the use er as well as the persona l routes E P ⊆ ⊆ (V P x V P x L) followed by y the user conducte ed with spec cific bus lines s.
The pers sonal mobilit ty graph enc odes the tra vel routines of the user i in a compact t form. Each personal route (s i i , , s j , l) ∈ E P h ighlights the e route taken n with bus lin ne l ∈ L from m station s i to o s j . Hence, t the graph not only y contains in nformation a about the m ost popular stops releva ant to the u user, but also o reveals The pers sonal mobilit ty graph enc odes the tra vel routines of the user i in a compact t form. Each personal route (s i i , , s j , l) ∈ E P h ighlights the e route taken n with bus lin ne l ∈ L from m station s i to o s j . Hence, t the graph not only y contains in nformation a about the m ost popular stops releva ant to the u user, but also o reveals
Definitio on (Persona l Travel Beh haviour Pred diction): The e prediction of persona al travel beh haviour is defined as a functio on p: (d, (t beg b , t end )) → G P , which r returns the personal mo obility graph h G P that describe es a user’s fu uture travel p patterns for a given day and time ho orizon as spe ecified by the e interval (t beg , t end d ) (e.g. 8am ‐ 12 am).
The outp put of the p rediction is thus a speci ific personal mobility gra aph for a sp pecific date a and time. This allo ws us to ide ntify not onl y the spatial , but also te mporal patte erns of user behaviour.
5.2.3 P Prediction Algorithm
In order to predict p personal trav vel behaviour r, we are pro oposing an a algorithm wh hich is runnin ng on the user’s m mobile device e. The predic tion algorith hm creates a n abstract re epresentatio on of the trav vel habits from pe rsonal trip h histories. We e refer to thi is representa ation as the personal tra avel profile o of a user. This pro ofile provide es the basis for statisti ical analysis of past tra avel behavio our upon w which the predictio ons are deriv ved.
Definitio on (Personal l Travel Prof file):
A perso onal travel pr rofile G T =(V T T , , E T ) consists s of the set o of station
V T ⊆ S vi isited by the e user and a statistical re presentation n of his trave el routines E E T ⊆ (V T xV T xL xTxDx ). Each edg ge (s i , s j , l, d d, t, n) ∈ E T T encodes th he frequency y n of trave ling among station s i an nd s j at a specific t time slot t at t day d.
Thus, bu us rides can b be encoded a as a transitio on from one station to an nother, labe lled with a f requency that can be learned from the use ers’ past trip p history. Thi is provides t he basis for a statistical model to discover r most popu lar routes, w which can b be found by reasoning o over the pro obabilities of f possible transitio ons in the pe ersonal trave l profile. The e algorithm will therefor re provide in nference me chanisms to extra ct only thos e routes wh hich are pred dicted to be relevant wi ith a high pr robability fo r a given time in t the future.
5.3 So ocial Beh aviour: D ata Minin ng and Pre ediction
In additi ion to perso onal travel be ehaviour, th e transporta ation networ rk can be au ugmented w ith social relations s of traveller rs. For this p purpose, we look not on nly on the tr avel history of an indivi dual, but also at those of hi is friends to o infer pred diction abou ut the socia l travel beh haviour as d described subsequ ently.
5.3.1 I Input Data
Basic inp put to our al lgorithm is k knowledge a bout the soc cial network of the users s. This inform mation is available e at existing social online e platforms s such as Face ebook or Goo ogle+. We us se this inform mation to identify the social re elations amon ng users.
Definitio on (Social N Network): T The social network en ncodes the users’ frien ndship relat tionships. Formally y, it is repre esented as a a graph G S =( (U S ,E S ), whic h contains t the set of a ll users U S a and their friendsh ip relations E S ⊆ U S xU S ). The friends of a specific c user u ∈ U are referred d to as fr(u)= ={v | (u,v) ∈ E S }.
Based o n the social network inf formation, w we are able to group us ers accordin ng to their fr riendship relations s. According to our defin nition, we ass sume mutua al friendship relations to comply with h existing Based o n the social network inf formation, w we are able to group us ers accordin ng to their fr riendship relations s. According to our defin nition, we ass sume mutua al friendship relations to comply with h existing
Definitio on (Social Tr rip History): A social trip p history H S = =(U S , T S ) of a set of users s U S contains s a set of persona l trip historie es T S ={T 1 , T 2 , …, T n }, wher re each T i re presents the e trip history y associated w with user u i ∈ U.
The soc ial history t hus represe ents a collec ction of trip data acros s different u users. This d data is a relevant t source of information to discover r more abst tract usage patterns, w which highlig ght social aspects i in the behav viour as discu ussed in the next section .
5.3.2 P Predictions s The pre diction of s social behav viour pattern ns includes mobility as pects of sev veral users. For this purpose , we extend the previou s definition o of a mobility y graph to ac ccount for th he behaviour r of social user gro ups.
Definitio on (Social M obility Grap ph): The socia al mobility g graph G S =(U, V S , E S ) is ass sociated with h a group of users U and consi ists of the st tations V S ⊆ S and social l routes E S ⊆ ⊆ (V S x V S x L x U) in the t transport network k.
As the d efinition sho ows, a social route (s i , s j, l, u) ∈ E S hig ghlights that a route with h bus line l ∈ ∈L from s i to s j is re elevant to us ser u ∈ U. Th e social mob bility graph t hus is an eff fective abstra act represen ntation of the socia al travel patt terns.
Definitio on (Social Tr ravel Behavi iour Predicti ion): The pre ediction of s social travel behaviour is s defined as a func ction p: (u, f fr(u), d, (t beg , t end )) →G S , which retur ns the social mobility gra aph G S given n a user u and his s social friend s fr(u), as w ell as a day d and time h horizon as s pecified by t the interval (t beg , t end ) (e.g. 8am m ‐ 12 am).
The pred diction of so cial travel in formation in ncludes aspe ects of both t temporal an d spatial soc cial travel behaviou ur. This allow ws us to high hlight changi ing usage pa atterns over time and ide entify comm on travel routines s among the users.
5.3.3 P Prediction Algorithm
For the prediction of social tra avel behavio our, our pre ediction algo orithm deal s with an i ncreased informat tion space w where the t travel choic es from a s set of users s need to b be represen ted. This informat tion is expos sed in a soci al travel pro ofile, which a aggregates t he travel be ehaviour from m several users in a statistical model.
Definitio on (Social Tr avel Profile) ):
A social tra avel profile G G T =(V T , E T ) co onsists of the e set of statio on V T ⊆S visited b by the user a nd his friend d, and a stat istical repres sentation of their travel routes E T ⊆ ⊆ (V T x V T x L xUxT T x D x ). Ea ach edge (s i , s j , l,u, d, t, n ) ∈ E T encod es the frequ ency n of us ser u travelin ng among station s s i and s j at a s specific time slot t at day y d.
Our pred diction algor rithm builds the social tra avel profile f for a social u user group, c consisting of f the user himself and all of h is friends w willing to sha are informat ion about th heir travel h history. Base d on the social tr ravel profile, , the probab bility of a fri iend travelli ng on a spe ecific route i in the futur e can be
compute ed. This allow ws us to det termine the relevance o of a route in its social co ontext, so th hat those compute ed. This allow ws us to det termine the relevance o of a route in its social co ontext, so th hat those