Coming out with the states, actions, rewards design for an application is an art. We discuss among the available options in IR and evaluate the options' effectiveness and efficiency.
Designing States, Actions, and Rewards for Using POMDP in Session Search
1. DESIGNING STATES, ACTIONS,
AND REWARDS FOR USING
POMDP IN SESSION SEARCH
Jiyun Luo, Sicong Zhang, Xuchu Dong, Grace Hui Yang
InfoSense
Department of Computer Science
Georgetown University
{jl1749,sz303,xd47}@georgetown.edu
huiyang@cs.georgetown.edu
1
2. 2
E.g. Find what city and state Dulles airport is in, what shuttles ride-
sharing vans and taxi cabs connect the airport to other cities, what hotels
are close to the airport, what are some cheap off-airport parking, and what
are the metro stops close to the Dulles airport.
DYNAMIC IR- A NEW PERSPECTIVE TO LOOK AT
SEARCH
Information
need
User
Search
Engine
4. 4
¢ Rich interactions
— Query formulation
— Document clicks
— Document examination
— eye movement
— mouse movements
— etc.
4
CHARACTERISTICS OF DYNAMIC IR
5. 5
¢ Temporal dependency
5
CHARACTERISTICS OF DYNAMIC IR
clicked
documentsquery
D1
ranked documents
q1 C1
D2
q2 C2
……
…… Dn
qn Cn
I
informa(on
need
itera(on
1 itera(on
2 itera(on
n
6. 6
¢ Fits well in this trial-and-error setting
¢ It is to learn from repeated, varied attempts which are
continued until success.
¢ The learner (also known as agent) learns from its dynamic
interactions with the world
— rather than from a labeled dataset as in supervised learning.
¢ The stochastic model assumes that the system's current
state depend on the previous state and action in a non-
deterministic manner
REINFORCEMENT LEARNING (RL)
6
7. PARTIALLY OBSERVABLE MARKOV
DECISION PROCESS (POMDP)
7
……s0 s1
r0
a0
s2
r1
a1
s3
r2
a2
— Hidden states
— Actions
— Rewards
1R. D. Smallwood et. al., ‘73
o1 o2 o3
7
— Markov
— Long Term Optimization
— Observations, Beliefs
8. 8
8
Study designs of states, actions, reward
functions of RL algorithms in Session Search
GOAL OF THIS PAPER
9. A MARKOV CHAIN OF DECISION MAKING STATES
[Luo, Zhang, and Yang SIGIR 2014]
9
10. 10
¢ Partially Observable Markov Decision Process
¢ Two agents
— Cooperative game
— Joint Optimization
WIN-WIN SEARCH: DUAL-AGENT STOCHASTIC
GAME
— Hidden states
— Actions
— Rewards
— Markov
[Luo, Zhang, and Yang SIGIR 2014]
11. 11
¢ A tuple (S, M, A, R, γ, O, Θ, B)
— S : state space
— M: state transition function
— A: actions
— R: reward function
— γ: discount factor, 0< γ ≤1
— O: observations
a symbol emitted according to a hidden state.
— Θ: observation function
Θ(s,a,o) is the probability that o is observed when the system
transitions into state s after taking action a, i.e. P(o|s,a).
— B: belief space
Belief is a probability distribution over hidden states.
PARTIALLY OBSERVABLE MARKOV DECISION
PROCESS (POMDP)
1R. D. Smallwood et. al., ‘73
12. 12
SRT
Relevant &
Exploitation
SRR
Relevant &
Exploration
SNRT
Non-Relevant
& Exploitation
SNRR
Non-Relevant
& Exploration
— scooter price ⟶ scooter
stores
— collecting old US coins⟶
selling old US coins
— Philadelphia NYC travel ⟶
Philadelphia NYC train
— Boston tourism ⟶ NYC
tourism
q0
HIDDEN DECISION MAKING STATES
[Luo, Zhang, and Yang SIGIR 2014]
13. ACTIONS
— User Action (Au)
¢ add query terms (+Δq)
¢ remove query terms (-Δq)
¢ keep query terms (qtheme)
— Search Engine Action(Ase)
¢ Increase/ decrease/ keep term weights
¢ Switch on or off a search technique,
¢ e.g. to use or not to use query expansion
¢ adjust parameters in search techniques
¢ e.g., select the best k for the top k docs
used in PRF
— Message from the user(Σu)
¢ clicked documents
¢ SAT clicked documents
— Message from search engine(Σse)
¢ top k returned documents
Messages are essentially
documents that an agent
thinks are relevant.
[Luo, Zhang, and Yang SIGIR 2014]
13
14. ¢ Based on Markov Decision Process (MDP)
¢ States: Queries
— Observable
¢ Actions:
— User actions:
¢ Add/remove/ unchange the query terms
¢ Nicely correspond to our definition of query change
— Search Engine actions:
¢ Increase/ decrease /remain term weights
¢ Rewards:
— nDCG
14
[Guan, Zhang, and Yang SIGIR 2013]
2ND MODEL: QUERY CHANGE MODEL
15. SEARCH ENGINE AGENT’S ACTIONS
∈ Di−1 action Example
qtheme
Y increase “pocono mountain” in s6
N increase
“france world cup 98 reaction” in s28,
france world cup 98 reaction stock
market→ france world cup 98 reaction
+∆q
Y decrease
‘policy’ in s37, Merck lobbyists → Merck
lobbyists US policy
N increase
‘US’ in s37, Merck lobbyists → Merck
lobbyists US policy
−∆q
Y decrease
‘reaction’ in s28, france world cup 98
reaction
→ france world cup 98
N
No
change
‘legislation’ in s32, bollywood legislation
→bollywood law
15[Guan, Zhang, and Yang SIGIR 2013]
16. QUERY CHANGE RETRIEVAL MODEL (QCM)
¢ Bellman Equation gives the optimal value for an MDP:
¢ The reward function is used as the document relevance
score function and is tweaked backwards from Bellman
equation:
16
V*
(s) = max
a
R(s,a) + γ P(s' | s,a)
s'
∑ V*
(s')
Score(qi, d) = P (qi|d) + γ P (qi|qi-1, Di-1, a)max
Di−1
P (qi-1|Di-1)
a
∑
Document
relevant score
Query
Transition
model
Maximum
past
relevanceCurrent
reward/
relevance score
[Guan, Zhang, and Yang SIGIR 2013]
17. CALCULATING THE TRANSITION MODEL
)|(log)|(
)|(log)()|(log)|(
)|(log)]|(1[+d)|P(qlog=d),Score(q
*
1
*
1
*
1ii
*
1
*
1
dtPdtP
dtPtidfdtPdtP
dtPdtP
qt
i
dt
qt
dt
qt
i
qthemet
i
ii
∑
∑∑
∑
Δ−∈
−
∉
Δ+∈
∈
Δ+∈
−
∈
−
−
+−
−
−−
δ
εβ
α
17
• According to Query Change and Search Engine Actions
Current reward/
relevance score
Increase weights
for theme terms
Decrease
weights for
removed terms
Increase weights
for novel added
terms
Decrease weights
for old added
terms
[Guan, Zhang, and Yang SIGIR 2013]
18. RELATED WORK
18
¢ Katja Hofmann, Shimon Whiteson, and Maarten de
Rijke. Balancing exploration and exploitation in
learning to rank online. In ECIR'11.
¢ Xiaoran Jin and Marc Sloan, and Jun Wang.
Interactive exploratory search for multi page
search results. In WWW '13
¢ Xuehua Shen, Bin Tan, and Chengxiang Zhai.
Implicit user modeling for personalized search. In
CIKM '05
¢ Norbert Fuhr. A Probability Ranking Principle for
Interactive Information Retrieval. In IRJ, 11, 3,
2008
18
19. STATE DESIGN OPTIONS
¢ (S1) Fixed number of states
— use two binary relevance states
¢ “Relevant” or “Irrelevant”
— use four states
¢ whether the previously retrieved documents are relevant
¢ whether the user desires to explore
¢ (S2) Varying number of states
— model queries as states, n queries è n states
— infinity states
¢ document relevance score distribution as states.
¢ one document corresponds to one state
19
20. ACTION DESIGN OPTIONS
¢ (A1) Technology Selection
— a meta-level modeling of actions
¢ implement multiple search methods, and select the best
methods for each query
¢ Select the best parameters for each method
¢ (A2) Term Weight Adjustment
— adjusted term weights
¢ (A3) Ranked List
— One possible ranking of a list of documents is one
single action
¢ If the corpus size is N and the retrieved document number
is n, then the size of the action space is:
20
PN
n
= N(N −1)...(N − n +1) =
N!
(N − n)!
21. REWARD FUNCTION DESIGN OPTIONS
¢ (R1) Explicit Feedback
— Rewards generated from user’s relevance assessments.
¢ nDCG, MAP, etc
¢ (R2) Implicit Feedback
— Use implicit feedback obtained from user behavior
¢ Clicks, SAT clicks
21
22. SYSTEMS UNDER COMPARISON
¢ Luo, et al. Win-Win Search: Dual-Agent
Stochastic Game in Session Search.
SIGIR’14
¢ Zhang, et al. A POMDP Model for Content-
Free Document Re-ranking. SIGIR’14
¢ Guan, et al. Utilizing Query Change for
Session Search. SIGIR’13
¢ Shen, et al. Implicit user modeling for
personalized search. CIKM '05
¢ Jin, et al. Interactive exploratory search
for multi page search results. WWW '13
22
S1A1R1(win-win)
S1A3R2
S2A2R1(QCM)
S2A1R1(UCAIR)
S2A3R1(IES)
S1A1R2
S1A2R1
S2A1R1
23. EXPERIMENTS
¢ Evaluate on TREC 2012 and 2013 Session Tracks
— The session logs contain
¢ session topic
¢ user queries
¢ previously retrieved URLs, snippets
¢ user clicks, and dwell time etc.
— Task: retrieve 2,000 documents for the last query in each session
— The evaluation is based on the whole session. Metrics include:
¢ nDCG@10, nDCG, nERR@10 and MAP
¢ Wall Clock Time, CPU cycles and the Big O notation
23
¢ Datasets
— ClueWeb09 CatB
— ClueWeb12 CatB
— spam documents are
removed
— duplicated documents
are removed
24. EFFICIENCY VS. # OF ACTIONS ON TREC 2012
24
¢ When number of actions increases, efficiency tends to
drop dramatically
¢ S1A3R2, S1A2R1,
S2A1R1(UCAIR),
S2A2R1(QCM) and
S2A1R1 are efficient
¢ S1A1R1(win-win) and
S1A1R2 are
moderately efficient
¢ S2A3R1(IES) is the
slowest system
25. ACCURACY VS. EFFICIENCY
25
TREC 2012 TREC 2013
¢ Accuracy tends to increase when efficiency decreases
¢ S2A1R1(UCAIR) strikes a good balance between accuracy
and efficiency
¢ S1A1R1(win-win) gives impressive accuracy with a fair
degree of efficiency
26. OUR RECOMMENDATION
26
¢ If focus on
accuracy
¢ If time limit is
within one hour
¢ If want the balance
of accuracy and
efficiency
v Note: number of actions heavily effect efficiency which need to be
carefully designed
27. CONCLUSIONS
¢ POMDPs are good for session search modeling
— Information seeking behaviors
¢ Design questions
— States: What changes with each time step?
— Actions: How does our system change the state?
— Rewards: How can we measure feedback or
effectiveness?
¢ It is something between an Art and Empirical
Experiments
¢ Balance between efficiency and accuracy
27
28. RESOURCES
¢ Infosense
— http://infosense.cs.georgetown.edu/
¢ Dynamic IR Website
— Tutorials : http://www.dynamic-ir-modeling.org/
¢ Live Online Search Engine – Dumpling
— http://dumplingproject.org
¢ Upcoming Book
— Dynamic Information Retrieval Modeling
¢ TREC 2015 Dynamic Domain Track
— http://trec-dd.org/
— Please participate, if you are interested in
interactive, and dynamic search
28