�ݺ�ߣ

Extending recommendation systems
with semantics and context-awareness

CCIA 2011

Victor Codina & Luigi Ceccaroni
vcodina@lsi.upc.edu lceccaroni@BDigital.org

Departament de Llenguatges i Sistemes Informàtics Health Informatics
Knowledge Engineering and Machine Learning Group Personalized Computational Medicine

Outline

 Traditional vs. Contextual recommendation

 State-of-the-art & Current limitations

 Research question

 Semantics acquisition & exploitation

 Proposed model

 Experimental evaluation

 Conclusions & Future work

Extending Recommendation Systems with Semantics and Context-Awareness 2

Traditional recommendation problem

 Regression problem:
o Given a pair (u ∈ U, i ∈ I), predict item’s degree of utility ( )
 Estimation based only on user and item information

preferences (u)

Preference Collaborative filtering (CF)
Content-based (CB)
Hybrid
Recommendation model attributes (i)
Matrix recommender


Context-aware recommendation problem

 Context as additional dimension for estimation
o Given a tuple (u, i, c), predict item’s degree of utility in context c
o Context = “situated action”
Training data
Representational view:
c Pre-filtering

c Recommendation model
Multi-Dimensional (MD)

Example:
c = (winter, cold) c Post-filtering
c1 = Season c2 = Temperature


State-of-the-art & limitations

 Adaptations of latent-factor models (MD paradigm)
 Examples:
o N-dimensional Tensor Factorization
o Bias-based Matrix Factorization with temporal dynamics
 Best prediction accuracy results on recent competitions
o E.g.: Netflix challenge (2009), Yahoo! Labs KDD Cup (2011)
 Main limitations of latent-factor models:
o Lack of transparency in explaining recommendations
o Low cold-start performance (users and items with few ratings)
o Lack of novelty and diversity of recommendations


Research questions & main assumptions

 Research questions
o Q1. Can we overcome the limitations and improve global
recommendation quality (not only prediction accuracy) by
exploiting domain and context knowledge?

o Q2. Under which conditions is this improvement maximized?

 Main assumptions
o There exists semantic relationships among entities of the
recommendation space (users, items, contexts)

o The adequate exploitation of these semantic relationships is
useful to overcome current limitations


Knowledge acquisition and representation

Domain/Context Concept Concept
concepts x y
S(x,y)?

Explicit similarity Implicit similarity

Ontology-based Statistics-based
Similarity measure - Edge-based (LCA) - Probabilistic measures (PMI)
- Node-based (MICA) - Dimensionality reduction (LSA)
- Logic-based - Graph-based (SimRank)

uses uses

Ontologies Data collections
Knowledge source - Folksonomies
- Taxonomies (ODP)
- Thesauri (Wordnet) - Item descriptions


User/Item representation

 Concept-based modeling (weighted overlay approach)

Domain knowledge
(concepts = item attributes)

User ‘u’ d2 d4 Item ‘i’
Pu Pi
d1 d3

(Degree of interest in d1) (Relevance of d3)

Interest inferring method Attribute weighting method
- Explicit feedback (Rating avg) - Structured content (IDF)
- Implicit feedback (Seen frequency) - Unstructured (TFIDF, tagshare)


Knowledge exploitation

Knowledge
Can be used for… Possible benefits
type
Measuring the semantic matching among Less rigid contextual
Contextual different context states filtering than using
exact matching
Applying semantic inference methods over Enrich item/user
user/item concept-based profiles: profiles with new
- Spreading activation concepts
- Reasoning based on DLs semantically related
Domain-
based Measuring the matching between two More precise
user/item using various semantic matching similarity
strategies: measurements that
- Pairwise (Best-pairs or All-pairs) using traditional
- Groupwise (set-, vector- or graph-based) measures


Case of study: a MD semantically-enhanced CB

 Contextual prediction model (bias-based):

Overall Contextual Contextual
rating avg User bias Item bias All-pairs Item-User semantic matching

where:
Session bias of (u,d) contextual bias of (u,d)

 Stochastic gradient descent for model training:


MovieLens Dataset

 Contextual concepts without semantics
o 3 contextual factors (season, time of the day, weekend?)
 Domain concepts with implicit semantics
o Set of pre-selected tags + set of genres
o Semantic relationships among tags acquired from folksonomy
 Original dataset pruned by selecting only items with a
certain amount of pre-selected tags


Offline experiment

 Last ratings (according to timestamp) testing
o In this way we simulate future predictions for each user
 5-fold cross validation
 Two recommendation tasks evaluated
o Rating prediction (RMSE) and Top-10 recommendation (Recall)
 Threshold-based cold-start performance evaluation
o User profile size < 25 ratings: 10% of users
 Performance comparison of the proposed model with:
o 3 model variants
o 5 baseline models
o 1 model based on matrix factorization


Results

 Paired t test significance among 4 model variants:
o Model 1 “Static-CB” (static bias + traditional Item-User matching)
o Model 2 “Static-SemCB” (static bias + “All-pairs” matching)
o Model 3 “Contextual-CB” (contextual bias + traditional matching)
o Model 4 “Contextual-SemCB” (contextual bias + “All-pairs” matching)
Global RMSE Cold-Start RMSE
0,851 0,919

0,844 0,918

0,837 0,917
0,05 0,001 0,17 0,05 0,62 0,01

0,83 0,916
Model 1 Model 2 Model 3 Model 4 Model 1 Model 2 Model 3 Model 4

(P-values in red)
E.g. P-value = 0,05 means that there is a 95% chance of being a real difference


Conclusions

 Context-awareness improves prediction accuracy for
users with a certain number of ratings (non cold-start)
o 25+ rating: 90% of users
 Semantics slightly improves cold-start performance
 The knowledge acquisition method for the MovieLens
folksonomy may be not adequate: limited domain
knowledge
 MovieLens users rate several movies at once and not just
after seeing the movie
o Rating-session--specific effects have a major influence in the
user ratings: distorted contextual information


Future work

 Extending evaluation of the proposed CB model:
o Using datasets from other domains (e.g. music, tourism, health)
o Experimenting with other sources of knowledge (e.g. Amazon
movie taxonomy)
o Experimenting with other methods for semantics exploitation
o Evaluating other properties (e.g. diversity, novelty, coverage)

 Extending CF models with the proposed semantic
approach:
o Neighborhood-based
o Matrix Factorization


Backup slides


Prediction models of all variants

 Model 1 (Static-CB):

 Model 2 (Static-SemCB):

 Model 3 (Contextual-CB):

 Model 4 (Contextual-SemCB):


�ݺ�ߣ

Extending Recommendation Systems With Semantics And Context Awareness

More Related Content

Extending Recommendation Systems With Semantics And Context Awareness