2. Goal
Long-term goal is to design a time-aware
recommender system, which can accurately
predict user's taste, given the current time.
● The vision is to model a single user u by
many micro profiles u1, u2, ..., un that best
represent the user in a particular time span.
Challenges
Implicit user feedback
Continuous temporal domain
Predict taste on new items rather than user behavior
L. Baltrunas & X. Amatriain 10/25/09
3. Outline
Approach & Challenges
Last.fm data set
Evaluation protocol
Empirical study
Latest and future work
L. Baltrunas & X. Amatriain 10/25/09
4. Approach: Challenges
Approach
How to combine the predictions generated for
each of the profiles and how to present the final
predictions.
Future work
How to discover meaningful time partitions (micro-
profile) based on the time cycles. Each partition
should represent a time slice where user has
similar repetitive behavior.
Investigated a simple non-personalized, non-
overlapping case of time partitioning.
L. Baltrunas & X. Amatriain 10/25/09
5. Last.fm Data
Implicit data:
Collected during a two year
period
Only Spanish users
#users 338
#tracks 322.871
#artists 16.904
#entries 1.970.029
We converted it to explicit
data: 1 to 5 stars system
[Celma'08]
L. Baltrunas & X. Amatriain 10/25/09
6. Evaluation of the System
The evaluation of a recommender system tries to
estimate the users' satisfaction for a given
recommendation.
Our goal is to predict the taste on new items rather
than user behavior.
We measure the accuracy of the system using Mean
Absolute Error (MAE).
Problem with continuous contextual variable:
The exact partitioning of the time domain defines the
ground truth that we want to predict.
L. Baltrunas & X. Amatriain 10/25/09
7. Error Measure: Our Approach
We allow only non
overlapping partitioning
We propose to compute
error E, given partitioning,
recommender and data:
L. Baltrunas & X. Amatriain 10/25/09
8. Experimental Evaluation
We used Last.fm data.
Matrix factorization as
the rating prediction
method.
We used 5 fold cross-
validation.
Finally, we do not look
into personalized
partitions but rather
evaluate global ones.
L. Baltrunas & X. Amatriain 10/25/09
9. Accuracy of the Method
We use a pre-defined time segmentation, for day,
week and year.
When using only the data of the segment the
accuracy E of the prediction improved for all our
observed segmentations.
L. Baltrunas & X. Amatriain 10/25/09
10. Towards Optimal Split of the Profiles
Day cycle is partitioned
into two segments each
spanning for 12 hours.
We used 3 different
methods to predict the
best partitioning:
True Error Cross Validation Cross Validation –
expensive, accuracy can
be increased by adding
more folds.
Explained Variance.
Information Gain.
Explained Variance Information Gain
10/25/09
11. Current work (1)
Generating artificial profiles
● In order to evaluate the goodness of the segmentation
measures we need a ground truth
● We inject artificial temporal changes in user profiles
and then compute how well the different segmentation
measures detect them
L. Baltrunas & X. Amatriain 10/25/09
12. Current work (2)
Is the approach domain or dataset specific?
● We are currently working on using the same
approach on IPTV data using viewing data
● Initial results are promising but not conclusive
L. Baltrunas & X. Amatriain 10/25/09
13. Future Work
Finding
Optimized Segments Including variable number
and per-user segmentation
Evaluation of the micro-profiling approach:
Prediction generation using (hierarchical) micro-profiles at
different temporal granularity
Recommendations at different levels, i.e., genre, artist,
album and track.
Extend the context information to include:
The current song.
The current album.
The current genre and mood of a song.
L. Baltrunas & X. Amatriain 10/25/09