This document presents an approach for recommending temporal aspects of entities based on multiple models. It introduces the problem of temporal entity aspect recommendation given an entity and time. The approach uses cascaded identification of event type and time period, and combines long-term salience features and short-term interest features using multiple ranking models. Experiments show that the ensemble model outperforms single models and baselines in recommending relevant temporal aspects.

1. Multiple Models for Recommending
Temporal Aspects of Entities
Tu Nguyen1, Nattiya Kanhabua2, Wolfgang Nejdl1
Presenter: Besnik Fetahu1
1. L3S Research Center / Leibniz Universität Hannover, Germany
2. NTENT España, Barcelona, Spain
1

2. Motivation
2ESWC 2018, Heraklion, Greece
Australia Open

3. Motivation
3ESWC 2018, Heraklion, Greece
Winners
Nominations
Movies Actors
Location
Athletes
Australia Open
Winners
Schedule
DrawResults

4. Motivation
4ESWC 2018, Heraklion, Greece
t
Jan Feb Mar
Querying
time

5. Motivation
5ESWC 2018, Heraklion, Greece
Long-term
cumulativeness
vs. Short-term
interest

6. Motivation
• Definition: Given a “search task” defined as an atomic information
need, a temporal “entity aspect” is an entity-oriented search task
with time-aware intent.
• Problem (Temporal Entity-Aspect Recommendation): Given an
event entity e and hitting time t as input, find the ranked list of
entity aspects that most relevant with regards to e and t.
6ESWC 2018, Heraklion, Greece

7. Approach Overview
7ESWC 2018, Heraklion, Greece

8. Approach Overview
8ESWC 2018, Heraklion, Greece
Sub-Task

9. Approach Overview
9ESWC 2018, Heraklion, Greece
Ranking
Task

10. Multi-criteria Learning
10ESWC 2018, Heraklion, Greece
• Multiple Ranking Models
• Idea: divide-and-conquer, each feature-set performs better for certain
entity type and at certain event time.
1. Probability the event entity e, at time t, of type C ∈ {Breaking, Anticipated}
2. Probability e is with subject to C is at event time T ∈ {Before, During, After}
1 2

11. Sub-task
11ESWC 2018, Heraklion, Greece
• Time and Type Cascaded Identification
• Semantic relation between task labels
• à joint-learning in cascaded manner
• Features
• Seasonality
• Trending
• Auto-correlation
• Prediction Errors
• SpikeM fitting parameters[1]
[1] Matsubara, Yasuko, et al. "Rise and fall patterns of informationdiffusion: model and implications." Proceedings of the18thACM
SIGKDD international conferenceon Knowledge discovery and data mining. ACM, 2012.
02060100140
observed
202530
trend
0204060
seasonal
−4002040
1990 1995 2000 2005
random
Time
Decomposition of additive time series

12. Ranking Features
12ESWC 2018, Heraklion, Greece
• Salience features
• Mainly extracted from Wikipedia or long duration query logs
• Avg. TF-IDF
• Language Model
• MLE, Entropy: reward most (cumulated) frequent aspects
• Short-term interest features
• Mainly extracted from recent query logs
• Trending velocity
• Temporal click entropy
• Cross correlation
• Temporal LM

13. Datasets
13ESWC 2018, Heraklion, Greece
• AOL query logs
• 03-2006 to 05-2006: 3 months
• Over 30 mil. Queries
• Manual construction:
• 837 entity queries
• 300 event-related queries
• Ground-truth: 70 queries (Breaking: 30, Anticipated: 30)
• Google Trend 2017
• 500 event-related queries
• Wikipedia dump 2006

14. Methods for Comparison
14ESWC 2018, Heraklion, Greece
• Random walk with restart (RWR)
• SOTA query auto-completion:
n Most popular completion
n Recent MPC
n Last N query distribution
n Predicted next N query distribution
• SVM-salience: with all salient features
• SVM-timeliness: with all short-term interest features
• SVM-all: with all features

15. Experiments-Subtask
15ESWC 2018, Heraklion, Greece
• RQ: How good is the classification method in identifying the most
relevant event type and period with regards to the hitting time?

16. Experiments
16ESWC 2018, Heraklion, Greece
• RQ: How do long-term salience and short-term interest features
perform at different time periods of different event types?

17. Experiments (2)
17ESWC 2018, Heraklion, Greece
• RQ: How do long-term salience and short-term interest features
perform at different time periods of different event types?

18. Experiments (3)
18ESWC 2018, Heraklion, Greece
• RQ: How does the ensemble ranking model perform compared to the
single model approaches?

19. Conclusion and Future Work
19ESWC 2018, Heraklion, Greece
• We studied the temporal aspect suggestion problem for entities in
knowledge bases.
• Focused on a “global” recommendation based on public attention.
• Group and Personalization recommendation (+ search context) is
interesting for future work.

20. Thank you!
20ESWC 2018, Heraklion, Greece