Discovering Causal Dependencies in Mobile Context-Aware Recommenders

ocelotgiantAI and Robotics

Nov 7, 2013 (3 years and 1 month ago)

59 views

Discovering Causal Dependencies in
Mobile Context
-
Aware Recommenders

Ghim
-
Eng Yap
1
, Ah
-
Hwee Tan

School of Computer Engineering

Nanyang Technological University


Hwee
-
Hwa Pang

School of Information Systems

Singapore Management University

1

Ghim
-
Eng Yap is sponsored by a graduate scholarship from the
Agency for Science, Technology and Research (A*STAR).

Context
-
Aware Recommender Systems


What are
Recommender Systems
?


Systems designed to augment the social recommendation
process in aid of decisions. Examples:


What do we mean by
Context
?


Any information that can characterize the interaction between
an user and the application. Examples:

Information
retrieval
systems

Recommenders for
restaurants, books,
movies, music, etc

Availability of
vegetarian food

Current
weather
conditions

Context Acquisition Challenges in
Mobile Context
-
Aware Recommenders


Resource Limitations


We need to identify the minimal set of context for a
particular user so as to save on context acquisitions.


Missing Context Values


The missing values may be recoverable given an explicit
encoding of the causal dependencies among context.

The solution: Bayesian Network


A directed acyclic graph


Encodes the complete causal dependencies among context,
and also between the various context and the target variable.


Handles missing context values effectively


This explicit encoding of user
-
specific causal dependencies
among context elements maintains high prediction accuracy.

preferred
ventilation
preferred
category
current attire
Is air-
conditioned
category
attire
requirement
of desired
category
of preferred
ventilation
attire is
appropriate
score
BN Learning Program
-

CaMML



State
-
of
-
the
-
art automated causal discovery learning method


Searches for Bayesian networks according to causal structure.


Uses the MML (Minimum Message Length) principle as metric.


Employs the Metropolis algorithm for its stochastic sampling.



Basic ideas:


Sample the space of all possible models (subject to constraints);


For each visited real model, compute a representative model
and count only on these representative models;


MML posterior of a representative = sum of members’ posteriors;


Best model is the representative with the highest MML posterior.

BUT state
-
of
-
the
-
art BN learning schemes are slow when the
number of learning variables is large!

Adopt a Tiered Context Model


A lower tier deals with a large set of context and item attributes


User context: user preferences as well as physical situation.


Item attributes: properties of each of the items to be ranked.


An upper tier deals with a small set of markers and a item score


Item markers: set of task
-
specific concerns for a typical user.


Item score: measures suitability of items based on markers.

Learned Bayesian Networks
User Contexts
Item
Markers
Item Score
Item
Attributes
Tiered Model enables Scalable Learning


No need to learn on entire large set of context and attributes


BN learning on just the upper tier to identify minimal set of markers.


These markers and their corresponding context and item attributes
are retained for the learning of causal dependency among context.


Final BN is personalized to that user
-

it encodes dependencies
among those context variables important specifically to him/her.

Learned Bayesian Networks
User Contexts
Item
Markers
Item Score
Item
Attributes

We get a compact and explanatory model of the user
-
specific
context dependencies that can predict on new (future) items.

A Restaurant Recommender


A lower tier deals with the large set of context and item attributes


26 user context: E.g. “preferred category”, “current attire”.


30 restaurant attributes: E.g. “category”, “attire requirement”.


An upper tier deals with a small set of markers and a item score


21 markers: E.g. “of desired category”, “attire appropriate”.


1 score: a value between 0.0 to 1.0 to rank each restaurant.

Learned Bayesian Networks
User Contexts
Item
Markers
Item Score
Item
Attributes
Learned Bayesian Network
User Context
Restaurant
Markers
Restaurant
Score
Restaurant
Attributes
Experimental Validation


Purpose:


To show that the automatic learning of dependencies among
context can effectively overcome the two context
-
acquisition
challenges in mobile context
-
aware recommenders, namely


Identifying the minimal set of context to minimize costs
, and


Maintaining accurate prediction with missing context values
.


Observation data:


Generated using a set of preference rules that represent different
user logics in considering markers, and a set of causal models
that state the dependencies among respective subset of context.

Datasets

Score for a restaurant is computed as

Five users are modeled, each with a different preference rule and
causal model that are unknown to the learning system. A thousand
consistent examples are generated for each user.

Learning Minimal Set of Markers


Purpose:


To validate that the minimal set of context truly important to a
certain user could be effectively discovered in the Bayesian
network that is learned automatically from data.


Procedure:


Extract just the markers and score information from the data;


Learn on this raw data using CaMML, and retain only markers
that are directly linked to the score node;


Repeat until a learned model has all its markers linked to score.

Learned Bayesian Network
User Context
Restaurant
Markers
Restaurant
Score
Restaurant
Attributes
Observations

Discovered Bayesian networks for User 1:

Performance metric:

Handling Missing Context Values


Purpose:


To verify that automatic learning of BN from data can capture the
causal dependencies among context, so that prediction on score
remains accurate despite crucial context values being missing.


Procedure:


Using the same datasets as before, we perform a 5x2
-
fold CV;


Each training set consists of just the score node and the minimal
set of restaurant markers, user context and restaurant attributes;

Learned Bayesian Network
User Context
Restaurant
Markers
Restaurant
Score
Restaurant
Attributes
Observations


Learning just on markers and score


Both suffer >20% drop in
accuracy with missing values.


We compare prediction accuracy of the
learned Bayesian networks (BN) to that
of the J4.8 decision trees (DT)
.


Learning of user
-
specific casual
dependencies among context


BN clearly outperforms DT.

Conclusions

Any comments to share with us?



Bayesian network learning from data can be applied to


find minimal set of important context

for a user and recommendation task,


exploit causal dependencies among context to
overcome missing values
.


A two
-
tiered context model is proposed for scalable learning


abstract large set of context into a smaller set of markers,


fast and scalable learning on only the important variables.


Experiments on a restaurant recommender


reliably discover the minimal set of learning variables, and


accurate prediction even when context values are missing.