1. Time-sensitive Personalized Query Auto-Completion
Date： 2015/01/29
Author： Fei Cai, Shangsong Liang, Maarten de Rijke
Source： CIKM’14
Advisor：Jia-ling Koh
Speaker：Sz-Han,Wang

2. Outline
Introduction
Method
Experiment
Conclusion

3. Introduction
Query auto-completion(QAC) is aimed at saving user’s time

4. Introduction A common approach:
Extract past queries from query log, and rank them by their past popularity → performance on average but far from optimal
Time and trend
MH370
movie
christmas

5. Introduction
User-specific

6. Introduction
Goal:
Propose a hybrid QAC model considers both of time-sensitivity and personalization to achieve optimal QAC effectiveness for a user.
Time-sensitivity: predict query popularity based on periodicity of a query and its recent trend
Personalization: exploit each user’s previous queries, both during the current session and from historical logs as user-specific

7. top N query completions
Flow Chart
Personalized QAC
Score user-specific query similarity
query
Time-sensitive QAC
Rank QAC candidates by predicted query popularity
top N query completions
candidates
Hybrid QAC
Rerank QAC candidates by combined time-sensitive QAC with personalized QAC

8. Outline
Introduction
Method
Experiment
Conclusion

9. Time-sensitive QAC Rank QAC candidates by predicted query popularity
Predict query popularity based on its recent trend and periodicity
Time-sensitive score = 𝑦 𝑡0+1 𝑞,λ
=λ× 𝑦 𝑡0+1 𝑞 𝑡𝑟𝑒𝑛𝑑 + 1−λ × 𝑦 𝑡0+1 𝑞 𝑝𝑒𝑟𝑖

10. Time-sensitive QAC Predict form recent trend (recent N days) harry
f: decay factor, TD(i): the interval from day i to future day t0+1
harry
harry potter
harry winston
harry style
N days = 3, f=0.95
𝑦 𝑡 ℎ𝑎𝑟𝑟𝑦 𝑝𝑜𝑡𝑡𝑒𝑟 𝑡𝑟𝑒𝑛𝑑
=0.35∗ ∗ ∗140
=131.45
𝑦 𝑡 ℎ𝑎𝑟𝑟𝑦 𝑝𝑜𝑡𝑡𝑒𝑟 𝑡𝑟𝑒𝑛𝑑
𝑡 0+1
𝑡 0
𝑡 0−1
𝑡 0−2
𝑡 0−3 85 90
110 70
𝑡 0−4
𝑇𝐷 1 =2,𝑤 1 = −1 = → 𝑛𝑜𝑟𝑚(𝑤 1 )=0.35
𝑇𝐷 2 =3,𝑤 2 = −1 = → 𝑛𝑜𝑟𝑚(𝑤 2 )=0.33
𝑇𝐷 3 =4,𝑤 3 = −1 = → 𝑛𝑜𝑟𝑚(𝑤 3 )=0.32
𝑦 𝑡 (ℎ𝑎𝑟𝑟𝑦 𝑝𝑜𝑡𝑡𝑒𝑟,1) 𝑡𝑟𝑒𝑛𝑑 =90+15=105
𝑦 𝑡 (ℎ𝑎𝑟𝑟𝑦 𝑝𝑜𝑡𝑡𝑒𝑟,2) 𝑡𝑟𝑒𝑛𝑑 =85+40=125
𝑦 𝑡 (ℎ𝑎𝑟𝑟𝑦 𝑝𝑜𝑡𝑡𝑒𝑟,3) 𝑡𝑟𝑒𝑛𝑑 =70+55=125

11. Time-sensitive QAC Predict form periodicity harry harry potter
Use autocorrelation coefficient to detect q’s periodicity
, 𝑇 𝑞 denotes q’s periodicity
𝑇 𝑞 =1 𝑦𝑒𝑎𝑟, 𝑀=3
harry
harry potter
harry winston
harry style
2005/07/16
100 95
120
2005/07/17
2004/07/17
2003/07/17
2002/07/17
𝑦 𝑡 ℎ𝑎𝑟𝑟𝑦 𝑝𝑜𝑡𝑡𝑒𝑟 𝑝𝑒𝑟𝑖 =
=105
𝑦 𝑡 ℎ𝑎𝑟𝑟𝑦 𝑝𝑜𝑡𝑡𝑒𝑟 𝑝𝑒𝑟𝑖 1 2 3
𝑟 1 = 1−2 ∗ 2−2 + 2−2 ∗ 3−2 +…+ 2−2 ∗(3−2) (1−2) 2 + (2−2) 2 + …+(2−2) 2 =−0.33
𝑟 2 = 1−2 ∗ 3−2 + 2−2 ∗ 1−2 +…+ 1−2 ∗(3−2) (1−2) 2 + (2−2) 2 + …+(1−2) 2 =−0.67
𝑟 3 = 1−2 ∗ 1−2 + 2−2 ∗ 2−2 + 3−2 ∗(3−2) (1−2) 2 + (2−2) 2 + (3−2) 2 =1

12. Personalized QAC Score user-specific query similarity
Using a combination of similarity scores 𝑆𝑐𝑜𝑟𝑒( 𝑄 𝑠 ,𝑞 𝑐 ) and 𝑆𝑐𝑜𝑟𝑒( 𝑄 𝑢 ,𝑞 𝑐 )
𝑄 𝑠 : the recent queries in the current search, 𝑄 𝑢 : the same user issued before
Personalized score =𝜔×𝑆𝑐𝑜𝑟𝑒( 𝑄 𝑠 ,𝑞 𝑐 )+ 1−𝜔 ×𝑆𝑐𝑜𝑟𝑒( 𝑄 𝑢 ,𝑞 𝑐 )

13. Personalized QAC 𝑤 𝑠 = 𝑓 𝑇𝐷 𝑠 −1 harry potter harry
, f: decay factor, TD(s): the interval between 𝑞 𝑐 and 𝑞 𝑠
𝑝 𝑞 𝑐 𝑞 𝑠 = 𝑤 𝑐𝑖 ∈ 𝑞 𝑐 𝑝 𝑤 𝑐𝑖 𝑞 𝑠 𝑁( 𝑤 𝑐𝑖 , 𝑞 𝑐 ) = 𝑤 𝑐𝑖 ∈ 𝑞 𝑐 𝑝 𝑤 𝑐𝑖 𝑊(𝑤 𝑐𝑖 ) 𝑁( 𝑤 𝑐𝑖 , 𝑞 𝑐 )
W( 𝑤 𝑐𝑖 )={𝑤:𝑤∈ 𝑞 𝑠 |𝑤 0 = 𝑤 𝑐𝑖 [0]} ,N( 𝑤 ∗ , 𝑞 ∗ ) denotes the frequency of term 𝑤 ∗ appearing in 𝑞 ∗
𝑝 𝑤 𝑐𝑖 𝑊(𝑤 𝑐𝑖 ) ≡𝑆𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦( 𝑤 𝑐𝑖 ,𝑊( 𝑤 𝑐𝑖 ))= 1 𝑊(𝑤 𝑐𝑖 | 𝑤 𝑗 ∈ 𝑊(𝑤 𝑐𝑖 ) 𝑆𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦( 𝑤 𝑐𝑖 , 𝑤 𝑗 )
= 1 𝑊(𝑤 𝑐𝑖 | 𝑤 𝑗 ∈ 𝑊(𝑤 𝑐𝑖 ) 𝑙𝑒𝑛(𝑐𝑜𝑚𝑚𝑜𝑛 𝑤 𝑐𝑖 , 𝑤 𝑗 ) 𝑚𝑖𝑛( 𝑙𝑒𝑛(𝑤 𝑐𝑖 ), 𝑙𝑒𝑛(𝑤 𝑗 ))
harry potter
harry
harry potter hogwarts
𝑝 ℎ𝑎𝑟𝑟𝑦 𝑝𝑜𝑡𝑡𝑒𝑟 ℎ𝑎𝑟𝑟𝑦 𝑝𝑜𝑡𝑡𝑒𝑟 ℎ𝑜𝑔𝑤𝑎𝑟𝑡𝑠 = × =0.84
𝑝 ℎ𝑎𝑟𝑟𝑦 𝑝𝑜𝑡𝑡𝑒𝑟 ℎ𝑎𝑟𝑟𝑦 𝑝𝑜𝑡𝑡𝑒𝑟 ℎ𝑜𝑔𝑤𝑎𝑟𝑡𝑠
𝑝 ℎ𝑎𝑟𝑟𝑦 𝑊(ℎ𝑎𝑟𝑟𝑦) = 1 2 × =0.7
𝑝 𝑝𝑜𝑡𝑡𝑒𝑟 𝑊(𝑝𝑜𝑡𝑡𝑒𝑟) = 1 1 × =1
𝑞 𝑐 term {harry, potter}
𝑞 𝑠 term {harry, potter, hogwarts}
W(harry)= {harry, hogwarts}
W(potter)= {potter}
, 𝑤 𝑢 depend on the query count

14. Hybrid QAC
Rerank QAC candidates by combined time-sensitive QAC with personalized QAC
Hybrid score =𝛾∙𝑇𝑆𝑠𝑐𝑜𝑟𝑒( 𝑞 𝑐 )+ 1−𝛾 ∙𝑃𝑠𝑐𝑜𝑟𝑒( 𝑞 𝑐 )
Standardize TSscore and Pscore
𝑇𝑆𝑠𝑐𝑜𝑟𝑒 𝑞 𝑐 ← 𝑦 𝑡0+1 𝑞,λ − 𝜇 𝑇 𝜎 𝑇
𝑃𝑠𝑐𝑜𝑟𝑒 𝑞 𝑐 ← 𝑃𝑠𝑐𝑜𝑟𝑒( 𝑞 𝑐 )− 𝜇 𝑃 𝜎 𝑃

15. Outline
Introduction
Method
Experiment
Conclusion

16. Experiment Dataset AOL sampled between 2006/03/01 and 2006/05/31
Sound and Vision sampled between 2013/01/01 and 2013/12/31

17. Experiment Evaluation metrics Measure forecast accuracy
Mean Absolute Error (MAE)
Symmetric Mean Absolute Error
Evaluate the effectiveness of QAC rankings
Mean Reciprocal Rank (MRR)
𝑦 𝑖 :𝑡ℎ𝑒 𝑡𝑟𝑢𝑒 𝑣𝑎𝑙𝑢𝑒, 𝑦 𝑖 :𝑡ℎ𝑒 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛 𝑣𝑎𝑙𝑢𝑒

18. Experiment
Query popularity prediction evaluation

19. Experiment
Impact of trade-off parameter
0.62
0.83

20. Experiment Performance of TS-QAC rankings
Performance of Hybrid QAC rankings

21. Outline
Introduction
Method
Experiment
Conclusion

22. Conclusion Adopt a combination of the two aspects of the QAC problem
Proposed to use time-series analysis for predicting its future frequency
Extend time-sensitive QAC method with personalized
The best performer λ *-H-QAC showing significant improvements over various time-sensitive QAC baselines
Use parallel processing to enhance the efficiency of the method and use other metrics to evaluate the QAC rankings