Temporal dynamic interaction

Download Report

Transcript Temporal dynamic interaction

A Temporal Context-Aware Model for User
Behavior Modeling in Social Media Systems
Hongzhi Yin1 , Bin Cui1 , Ling Chen2
Zhiting Hu1, Zi Huang3
Peking University
2University of Technology, Sydney
3University of Queensland
1
1
Outline
■
■
Motivation
Offline Modeling
 TCAM-Temporal Context-Aware Mixture Model
■
Online Recommendation
 Computation of Ranking Score
 Fast Top-k recommendation Algorithm
■
Experiments
 Experimental Setup
 Experimental Results
■
Conclusions
2
Background
With the rising popularity of social
media, a better understanding of users’
online behaviors is of great importance
for the design of many applications,
such as personalized recommendation,
information filtering, behavioral
targeting and computational advertising.
3
Temporal dynamic interaction
■
What is temporal dynamic interaction?
 That user 𝑢 rates or likes item 𝑣 at time 𝑡 does not means 𝑢 is willing to rate 𝑣 at other
time
 Temporal dynamic interaction requires the recommender system to recommend right
items to the right users at the right time.
■
What leads to the temporal dynamic interaction
 The social context, we call the temporal-dynamic social context is temporal context,
which reflects the attention of the public during a particular time period
 For example, the temporal context in February is “Valentine's Day”; while the temporal
context in October is “Halloween”.
4
■
To capture the temporal context is critical to improve temporal
recommendation
 Some items related to temporal context can only be recommended at the right temporal context
 E.g., when it is Halloween, the events “TAKAGISM” are good recommendations. But on
Valentine's Day, users may think that these recommendations are stupid, although they are
generally interested in the things of thriller and horror.
5
■
By investigating multiple social media systems, we observe that user
behaviors are generally influenced by two factors

Internal Factor: the intrinsic interest of the user
 External Factor: the temporal context, i.e., the attention of the pubic during a time
period, e.g., social events
■
Moreover, the two factors have different degrees of influence on user
rating behaviors on different social media platforms, even for different
users.
6
Challenge
■
The challenge is that data representing user interests and data associated
with the temporal contexts are mixed up together. E.g., some of posts in
Twitter are associated with user interests while others reflect the social
context.
■
We needs
 To precisely identify user intrinsic interests from the mixture data
 To discover the temporal context from the mixture data
 To determine the influence degrees of the two factors
7
Tag Cloud from My Twitter
My Interests:
Recommender System
User Model Data
Learning Modeling
Temporal Context:
Alibaba IPO
MH 370 Malaysia
8
Outline
■
■
Motivation
Offline Modeling
 TCAM-Temporal Context-Aware Mixture Model
■
Online Recommendation
 Computation of Ranking Score
 Fast Top-k recommendation Algorithm
■
Experiments
 Experimental Setup
 Experimental Results
■
Conclusions
9
Temporal Context - Aware Mixture Model
■
■
■
■
■
Topic/Topic Model: a multinomial distribution over a set of items (or words)
We introduce two different types of topics to model user interests and
temporal contexts, respectively.
 User-Oriented Topics 𝜙𝑧
 Time-Oriented Topics 𝜙𝑥′
A user-oriented topic is a summary of regular user interests, the popularity of
which remains stable over time
A time-oriented topic is a temporally coherent theme, and often driven by
real-life events. It has a clear temporal feature that the popularity of the topic
presents increasing or declining trends over time and reaches its peak only
during a certain period of time.
The interest distribution of a user 𝑢 is represented as a multinomial
distribution over a set of user-oriented topics 𝜃𝑢 , and each temporal context
is modeled as a multinomial distribution over a set of time-oriented topics 𝜃′𝑡 .
10
Two types of topics
An Example of Two Types of Topics Detected from Delicious
11
Model Structure of TCAM
𝜆𝑢
1 − 𝜆𝑢
𝜃𝑢
𝜙1
𝜃′𝑡
𝜙′3
12
Parameter Estimation in TCAM
■
The log-likelihood of the observed users’ temporal online behaviors stored in the Cuboid 𝑪
■
E-M algorithm to estimate model parameters
…
E-Step
Compute expectation
Q(; n )
M-Step
…
Maximize, closed form solution
E-M algorithm can be easily expressed in MapReduce, which is scalable to large-scale datasets.
13
Outline
■
■
Motivation
Offline Modeling
 TCAM-Temporal Context-Aware Mixture Model
■
Online Recommendation
 Computation of Ranking Score
 Fast Top-k recommendation Algorithm
■
Experiments
 Experimental Setup
 Experimental Results
■
Conclusions
14
Online Recommendation
■
■
■
We use E-M method to infer model parameters offline, such as user
interest 𝜃𝑢 , the temporal context 𝜃′𝑡 , user-oriented topics 𝜙𝑧 and timeoriented ϕ′𝑥 , and mixing weights 𝜆𝑢 .
When receiving a query 𝑞 = 𝑢, 𝑡 , a new multinomial distribution, 𝜗𝑞 ,
is first constructed by combining 𝜃𝑢 and 𝜃′𝑡 . Since they have different
topic space, we need to expand user interest and temporal context
spaces to be of the same dimension, as follows:
The index of user-oriented topic 𝑧 is still in the range of [1, 𝐾1 ] , but the
index of time-oriented topic x is renumbered in the range of [𝐾1 + 1,
𝐾1 + 𝐾2 ]. We use
to denote the weight of item 𝑣 on dimension ,
defined as
15
Online Recommendation
■
Computation of Ranking Score
 Each query 𝑞 is represented by a vector 𝜗𝑞 with K dimensions, and each item is also
represented by a vector 𝜑𝑣 .
 Given a query 𝑞 = 𝑢, 𝑡 , the ranking score of each item 𝑣 is computed as the
inner product of the two vectors:
■
The naïve online recommendation algorithm:
 Compute the ranking scores for all available items
 Find top-k ones with largest ranking scores
 Good for small-scale problem, not feasible for large-scale problem
16
Fast Online Recommendation Algorithm
■
To speed up the online recommendation, we transform our online
top-k recommendation task to the classic top-k query processing
problem in the filed of database.
 Each item has K attributes (topics), and its value on attribute z is defined as
the weight on topic z.
 Considering that our defined ranking function is strictly monotone given a
query, we decide to extend the classic Threshold-based algorithm (TA) to
retrieve the optimal top-k items.
■
The TA algorithm has the nice property of correctly finding the
top-k items by scanning the minimum number of items rather
than all items. (Ronald Fagin, PODS 2001)
17
Query
TA - Recommend
Top-k items
Retrieval
Online Recommendation
Offline Modeling
Sorted Lists
Column Index
User
Profile DB
List1
...
List K
TCAM
Model
User
Interest
Temporal
Context
Outline
■
■
Motivation
Offline Modeling
 TCAM-Temporal Context-Aware Mixture Model
■
Online Recommendation
 Computation of Ranking Score
 Fast Top-k recommendation Algorithm
■
Experiments
 Experimental Setup
 Experimental Results
■
Conclusions
19
Experiments
■
Datasets
 To evaluate the performance of TCAM in temporal recommendation, we conduct experiments
on MovieLens, Digg and Douban Movie Datasets.
 To evaluate the performance of TCAM in topic discovery, we conduct experiments on
Delicious Dataset.
■
Comparison Approaches




UT-User Topic Model
TT - Time Topic Model
BPRMF – Bayesian Personalized ranking framework based on matrix factorization (UAI’09)
BPTF – Bayesian probabilistic tensor factorization model (SDM’10)
20
Effectiveness of Temporal Recommendation
Temporal Recommendation Accuracy on Digg
Temporal Recommendation Accuracy on MovieLens
21
Efficiency w.r.t Online Recommendation
22
User-Oriented Topics
U1
U6
windows 0.049 resources 0.034
tools 0.048
education
U7
U8
U13
news 0.107
program 0.028
food 0.034
latest 0.102
python 0.019
recipe 0.033
current 0.099
Ruby 0.016
Cooking 0.030
0.031
Freeware 0.038
interactive
0.020
firefox 0.038
Teaching 0.020
World 0.094
javascript 0.015
Dessert 0.026
Google0.029
science 0.019
events 0.084
software 0.014
Shopping 0.021
security 0.028
tools 0.015
newspaper 0.084
tutorial 0.011
Home 0.016
23
Time-Oriented Topics
■
The USA public’s attention in 2009
T6
T8
T9
T10
7.6-7.15
7.1-7.6
10.7-10.15
6.24-6.30
july 0.012
july 0.035
free 0.012
michael 0.038
free 0.010
happy 0.020
nobel 0.012
jackson 0.036
summer 0.008
day 0.016
prize 0.011
rip 0.007
live 0.007
firework 0.009
peace 0.008
farrah 0.007
potter 0.006
independ 0.006
win 0.008
dead 0.005
harry 0.006
celebrate 0.005
obama 0.008
sad 0.005
24
Temporal Distribution of Time-Oriented Topics
25
Conclusions and Future Work
■
We proposed a temporal context-aware mixture model TCAM to model
users’ rating behaviors on social media systems, and designed a fast top-k
recommendation framework.
 Individual Interests
 Temporal Context (i.e., the attention of the public during a particular time period)
■
In the future work, we will divide user interests into long-term interests
and short-term interests:
 Individual long-term interests are generally stable
 Their short-term interests are allowed to change over time, i.e., being dynamic
■
The user-time-item cuboid is very sparse. To alleviate the problem of data
sparseness, we attempt to exploit the social correlation information (the
homophily phenomenon in the social network) as well as the temporal
correlation information.
26
Q&A
Thank You
27