Nex ac i i y p edic ion om indi idual daily mobili y pa e ns
Xianchen Wang, Nilu e Sa i Aslam, Chen Zhong
The Ba le Cen e o Ad anced Spa ial Analysis is pa o The Ba le , UCL
Con ac : [email protected]
Aim and Objec i es
➢Model he dynamics o indi idual's daily mo emen s
➢P edic human ac i i ies in he nex ime in e al
Figu e 1: Illus a ion o he nex ac i i y p edic ion
➢Indi idual-le el seman ic
ajec o y da a ex ac ed om
mobile phone applica ions
➢The da ase consis s o eco ds
in ou dimensions: use ID,
ime, loca ion and ac i i y ype
➢The e a e 1475 use s wi h
104,069 ajec o y eco ds
wi hin 30 days (No embe 2021)
➢Human’s daily ac i i ies:
➢home (H)
➢wo k (W)
➢ea ing and d inking (EaD)
➢educa ion (Ed)
➢en e ainmen (En)
➢ equen shopping (S1)
➢non- equen shopping (S2)
➢o he s (O)
Figu e 2: Visualisa ion o all
indi idual ajec o ies
Figu e 3: Quan i a i e dis ibu ion o
di e en ypes o ac i i ies
Me hodology
➢Use embedding laye
o lea n seman ic
ela ionship
➢Cons uc aining and
es da ase s using
sliding windows
Figu e 4: Nex ac i i y p edic ion amewo k
Figu e 5: S uc u e o p oposed MTL-LSTMs
model
➢Mul i-ca ego y
classi ica ion ask
➢The p ima y challenge
lies in he imbalance
among ac i i y
ca ego ies
➢T ans o m he ask
in o mul iple bina y
classi ica ion asks,
le e aging concep s
om ensemble
lea ning and mul i- ask
lea ning
Figu e 6: E alua ing he model pe o mance using
a ious pe o mance me ics
Figu e 7: P ecision- ecall cu e o MTL-LSTM
➢The in e se ela ionship be ween ac i i y en opy and p edic ion
accu acy, i.e., lowe en opy wi h highe p edic ion accu acy
Figu e 8: Rela ionship be ween ac i i y en opy and p edic ion accu acy
Conclusion
➢Among hese me ics, he highes sco e in MTL-LSTM is cap u ed
in ecall
Resul s and Findings
Da a Desc ip ion
➢The p oposed app oach e ec i ely imp o es ecall, pa icula ly o
ou ine ac i i ies, while highligh ing challenges in p edic ing less
s uc u ed beha iou s
➢Fu he e inemen o mul i- ask lea ning s a egies and ensemble
echniques can enhance p edic ion accu acy o i egula ac i i ies
➢The p edic ion esul s
illus a e highe
accu acy o
ou ine ( egula )
ac i i ies, such as home,
wo k, and educa ion
➢Less s uc u ed
ac i i ies, such as
ea ing, en e ainmen ,
and shopping, ob ain
lowe accu acy om
indi iduals' daily ac i i y
chains
Calcula ion o mulas
o di e en me ics
