collaborative nowcasting for contextual recommendation · 2016-05-02 · motivation and problem...

Motivation and Problem DefinitionCollaborative Nowcasting Model

ExperimentsConclusion and Future Work

Collaborative Nowcasting for Contextual

Recommendation

Yu Sun †‡1, Nicholas Jing Yuan ‡2, Xing Xie ‡3,

Kieran McDonald ∗4, Rui Zhang †5

† University of Melbourne{

1sun.y,

5rui.zhang}@unimelb.edu.au

‡ Microsoft Research ∗ Microsoft Corporation{

2nicholas.yuan,

3xing.xie,

4kieran.mcdonald}@microsoft.com

April 14th 2016

Y. Sun, N. J. Yuan, X. Xie, K. McDonald and R. Zhang Collaborative Nowcasting for Contextual Recommendation



Outline

1 Motivation and Problem Definition

2 Collaborative Nowcasting Model

3 Experiments

4 Conclusion and Future Work



ExperimentsConclusion and Future Work Motivation Problem Definition

Outline


Motivation

Problem Definition


Nowcasting

Model Formulation

Parameter Estimation

3 Experiments

Set-up

Results





Proactive Experiences on Mobile Phones

Digital assistants: Cortana, Google Now, Siri.

Proactive experiences

recommend “the right information at just the right

time” a and

help you “get things done” b

even “before you ask” c.

ahttp://www.google.com/landing/now/bhttp://dev.windows.com/en-us/cortanachttp://www.apple.com/ios/whats-new/





Information types: videos, news, traffic, weather,apps, places, (calendar, stock prices, sports) etc.

Each type in such a layout is called a card





Limited display size =⇒ showone or two cards

Which card a user needs ⇐=

intent

to have dinner =⇒restaurant cardto drive home =⇒ trafficcard

To recommend “the right

information at the right time”,need monitor intent




Intent and Context

Intent! Contextexternal context: physical environment (e.g., time,

location)

internal context: users’ states (e.g., activity, usage of

apps)

Example (From Context to Intent)

context: 6:00 p.m., in the office intent: to drivehome

context: just left a shopping center, using Yelp intent:to find a restaurant




Intent and Context

Relationship between context and intent is difficult to model

intent and context change swiftly

exhibit strong sequential correlation

Context itself is heterogeneous and complicated

all contemporaneous information related to the intent

Challenge to model

structure of context

relationship between context and intent




Existing Work

Traditional recommendation models cannot tackle the challenge

only deal with a given intent (e.g., to find movies, books)

recommend new items fulfilling the given intent

Time-aware recommendation models also cannot apply

do not consider other context besides time

not suitable for swiftly changing context/intent

Context-aware recommendation models do not work either

do not take sequential correlation into account

consider only external context (e.g., time, location)




New Recommendation Paradigm

Personal Digital Assistant-Style Recommendation

user-centered rather than product/item-centered

recommendation based on multiple types of intent

First step: monitoring real-time intent by context

Wide Applications

Proactive experiences

Online advertisement




Outline


Motivation

Problem Definition


Nowcasting

Model Formulation


3 Experiments

Set-up

Results





Intent Monitoring Problem

The studied problem is defined as follows:

Definition (Intent Monitoring)

Given a starting time t0, a monitoring granularity ∆, a type of

intent γ and the context X ut

of user u, the intent monitoring

problem is to predict whether user u has intent γ with context

X ut

for each time step t of length ∆ starting from t0.

Example

Time step 10 a.m. 11 a.m 12 p.m. 1 p.m. Now

News intent 0 0 1 1 ?




Nowcasting Model FormulationParameter Estimation

Outline


Motivation

Problem Definition


Nowcasting

Model Formulation


3 Experiments

Set-up

Results






Why Nowcasting?

UW Engineering Bldg: Waterproofing went on one day1

1Picture from: Cliff Mass, University of Washington





Why Nowcasting?

Washed off a few hours later1






Why Nowcasting?

Reapplied the next day. How much did this cost?1






Difference between Nowcast and Forecast

historical data

variable ofinterest

side data

(a) Nowcast

historical data

variable ofinterest

(b) Forecast

Nowcast: prediction of current or very near future

Difference: side data

contemporaneous withmore frequently available (e.g., industrial output → GDP)





Side-Data Used in Nowcasting

In meteorology: nowcasting weather

atmospheric conditions from aircraft

water vapor distributions from GPS receivers

social media data from Facebook, Twitter, etc.

In macroeconomics: nowcasting GDP

personal consumption, industrial production

surveys, financial variables (e.g., interest rates, CPI)

Google trend data

In data mining: nowcasting rainfall, illness rates

search engine query log (e.g., Google trend)

posts in social media (e.g., Twitter)





Existing Nowcasting Model Cannot Apply

Thunderstorm: linear regression with exponential smoothing

variable of interest quite different from intent

GDP nowcasting: dynamic factor model

granularity much larger than hours

macroeconomic variables are non-personalized

Rainfall nowcasting: Bootstrapped LASSO + regression

cannot address the personalized scenario

hard to obtain textual features for personalized intent





Outline


Motivation

Problem Definition


Nowcasting

Model Formulation


3 Experiments

Set-up

Results






The Panel

Context 7→ stochastic processes

Historical/side data 7→ time series

All series for user u 7→ panel X u

Time Step 10 a.m. 11 a.m 12 p.m. 1 p.m. Now

Facebook 306 0 915 32 257

Skype 0 1853 0 0 -McDonald’s 0 1256 652 0 0

IKEA 0 0 0 532 1247

Dist-to-Office 10.4 8.3 9.1 21.3 -Day-of-Week 6 6 6 6 6

News Intent 0 0 1 1 ?





Latent Factor Structure

Latent Factors

We assume that xi ,t ∈ X has structure xi ,t = λ′i· ft + ξi ,t , where

ft = (f1,t , .., fR,t )′, λi = (λi ,1, .., λi ,R)

′, and ξi ,t ∼ N (0, ψ2i ,t).

Written in matrix form

xt = Λft + ξt

where xt = (x1,t , .., xN,t )′, Λ = (λ1, ..,λN )

′, ξt = (ξ1,t , .., ξN,t )′

Factor Transition

To exploit sequential pattern, we assume dynamics of latent

factors have structure

ft = Aft−1 + Bωt

where A ∈ RR×R, B ∈ R

R×Q, and ωt ∼ WN(0, IQ).





Outline


Motivation

Problem Definition


Nowcasting

Model Formulation


3 Experiments

Set-up

Results






Estimation Overview

Xu

User panel

X

Λu F

Loading

Collaborativelatent factors

Λu F

Xu

For each u

Fu

Personalizedlatent factors

For each u

su

Intent

Extracting Collaborative Latent Factors Kalman Filtering Regression

Figure: Collaborative Nowcasting Model





Collaborative Latent Factors: Using CP Decomposition

XN

MT

u1 ur

v1 vr

w1 wr

≈ + · · · +

Estimation of Parameters

After CP decomposition

Xu ≈ UD

(u)V

′

where D(u) = diag(Wu,1, . . . ,Wu,r ), and U ∈ RN×R, V ∈ R

T×R,

W ∈ RM×R. We have

F = V′ and Λ

u = UuD

(u).





Using PARAFAC2 Decomposition

X u V ′Gu

Lu

≈

Estimation of Parameters

After PARAFAC2 decomposition

Xu≈ G

uHL

uV

′

where Gu ∈ RNu×R , H ∈ R

R×R is invariant to u, Lu ∈ RR×R.

We have

F = V′ and Λ

u = GuHL

u.





Personalized Latent Factors

Collaborative latent factors arenot sufficient

static common structureunsuitable forpersonalized scenario

Apply Kalman Filter on F u andX u

Time Update (Prediction)

ft = Aft−1 + Bωt

Pt = APt−1A′ + Ψt

Measurement Update

(Correction)

Kt = PtΛ′(ΛPtΛ

′+Ψt)−1

ft = ft + Kt(xt − Λft)Pt = (I − KtΛ)Pt



ExperimentsConclusion and Future Work Set-up Results

Outline


Motivation

Problem Definition


Nowcasting

Model Formulation


3 Experiments

Set-up

Results





Data Sets

Proactive log of a commercial digital assistant from

June 10th 2015 to July 9th 2015

clicks as indicators of intenteight types of intent: news, weather, etc.in total contain 20,807 anonymous users

Collect intent-related context, apps used and venuesvisited




Measurements and Compared Methods

Measurements

Macro F-measure: average performance among all users

Micro F-measure: performance per instance

Compared Methods

BoostedTree: used in existing contextual ranking models

FM: (factorization machine) for next-basket recommendation

NowcastIndi: the macroeconomic nowcasting model

CNowcastCP: CP decomp. for collaborative latent factors




Outline


Motivation

Problem Definition


Nowcasting

Model Formulation


3 Experiments

Set-up

Results





Effect of Parameters R and Q

0.9

0.95

1

1.05

1.1

2 3 4 5 6

Relative Macro F-measure

R

News

Weather

Finance

Sports

0.9

0.95

1

1.05

1.1

1 2 3

Relative Macro F-measure

Q

News

Weather

Finance

Sports

Observation

R = 4, Q = 2 is a good choice

small performance variance




Comparison across Models

Macro F-measure

Model News Events Weather Places Finance Calendar Traffic Sports

BoostedTree 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000

FM 0.877 1.102 1.459 3.465 1.263 9.179 1.332 1.395

NowcastIndi 1.746 2.643 4.403 12.70 3.788 14.92 5.800 4.221

CNowcastCP 1.766 2.513 4.329 12.16 3.412 15.33 5.483 4.195

CNowcast 1.963 2.950 4.904 14.13 4.680 16.95 7.377 5.264

Micro F-measure

Model News Events Weather Places Finance Calendar Traffic Sports

BoostedTree 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000

FM 1.040 1.280 1.497 4.951 0.932 7.231 1.114 1.276

NowcastIndi 1.365 1.733 2.223 8.073 1.526 8.019 1.997 1.625

CNowcastCP 1.422 1.686 2.301 7.893 1.427 8.447 2.048 1.636

CNowcast 1.513 1.927 2.432 9.026 1.822 8.888 2.572 2.037




Comparison across Monitoring Granularity

0

2

4

6

8

10

12

14

16

4 3 2 1

Performance Ratio

Monitoring Granularity (h)

Macro F-measure

Micro F-measure

(a) Ratio to BoostedTree

0

2

4

6

8

10

12

14

16

4 3 2 1

Performance Ratio

Monitoring Granularity (h)

Macro F-measure

Micro F-measure

(b) Ratio to FM

Observation

monitoring granularity ր ⇒ performance advantage ր

closer to “now" ⇒ suitable for nowcasting




Conclusion

Contextual Intent Monitoring

Nowcasting users’ real-time intent with context is key for

personal digital assistant-style recommendation.

Collaborative Nowcasting Model

The collaborative nowcasting model effectively models the

complicated relationship between context and intent via

nowcasting and collaborative capabilities.


collaborative nowcasting for contextual recommendation · 2016-05-02 · motivation and problem...

Documents