exploring statistical language models for recommender systems [recsys '15 ds poster]
TRANSCRIPT
Exploring Statistical Language Modelsfor Recommender SystemsDaniel [email protected] – http://www.irlab.orgInformation Retrieval Lab, Computer Science Department, University of A Coruña
Information Retrieval (IR)Goal Retrieve relevant documents according to the information
need of a user.Examples Search engines.Methods They can be based on:
Vector Vector Space Model.Matrix factorisation Latent Semantic Indexing.Probabilistic modelling Language Models.
Information Fitering (IF)Goal Select relevant items from an information stream for a
given user.Examples spam filters, recommender systems.Methods Some Collaborative Filtering methods are:
Vector Pairwise similarities (cosine, Pearson, etc.).Matrix factorisation SVD, NMF.Probabilistic modelling LDA.
Overview• Information Filtering (IF) and Information Retrieval (IR) are two sibling fields.• Statistical Language Models are a successful technique in IR → Explore how to apply them to recommendation.• We start by improving the current adaptation of Relevance-Based Language Models to Collaborative Filtering [1].
Relevance-Based Language Models
IR RecSys
Query Target userDocument Neighbour
Term Item
RM2 : p(i|Ru) ∝ p(i)∏j∈Iu
∑v∈Vu
p(i|v) p(v)p(i)
p(j|v)
• Iu is the set of items rated by the user u.• Vu is the set of neighbours of the user u.• p(i|u) is computed smoothing the maximum likelihood es-timate.• p(i) and p(v) are the item and user priors.
Smoothing methodsSmoothing deals with data sparsity and plays a similar role tothe IDF using a background model: p(i|C) =
∑v∈U rv,i∑
j∈I, v∈U rv,j[3].
Jelinek-Mercer(JM)
pλ(i|u) = (1− λ) ru,i∑j∈Iu ru,j
+ λ p(i|C)
Dirichlet Priors(DP)
pµ(i|u) =ru,i + µ p(i|C)µ+
∑j∈Iu ru,j
AbsoluteDiscounting
(AD)pδ(i|u) =
max(ru,i − δ, 0) + δ |Iu| p(i|C)∑j∈Iu ru,j
PriorsPriors provide a principled way of introducing knowledge intothe recommender [2].
Uniform (U) Linear (L)
Userprior
pU(u) =1
|U| pL(u) =
∑i∈Iu ru,i∑
v∈U∑
j∈Iv rv,j
Itemprior
pU(i) =1
|I| pL(i) =
∑u∈Ui
ru,i∑j∈I
∑v∈Uj
rv,j
Experiments on MovieLens 100k
Algorithm nDCG@10 Gini@10 MSI@10
SVD 0.0946 0.0109 14.6129SVD++ 0.1113 0.0126 14.9574NNCosNgbr 0.1771 0.0344 16.8222UIR-Item 0.2188 0.0124 5.2337PureSVD 0.3595 0.1364 11.8841RM2-JM 0.3175 0.0232 9.1087RM2-DP 0.3274 0.0251 9.2181RM2-AD 0.3296 0.0256 9.2409RM2-AD-L-U 0.3423 0.0264 9.2004
Research directions• Some techniques developed for solving IR problemscan be effectively applied to recommendation.• Probabilisticmodels from IR are competitive recom-mendation algorithms although there is still room forimprovements.• Language Models provide an interpretable and prin-cipled way of generate recommendations.• Using different priors [2] or clustering algorithms forthe neighbourhoods [1] can improve RM2.• We envision as future work the development ofcontext-aware and hybrid recommendations underthe Language Modelling.
Bibliography[1] J. Parapar, A. Bellogín, P. Castells, and A. Bar-
reiro. Relevance-Based Language Modelling for Recom-mender Systems. Information Processing & Management,49(4):966–980, 2013.
[2] D. Valcarce, J. Parapar, and A. Barreiro. A Study of Priorsfor Relevance-Based Language Modelling of RecommenderSystems. In RecSys ’15. ACM, 2015.
[3] D. Valcarce, J. Parapar, and A. Barreiro. A Study ofSmoothing Methods for Relevance-Based Language Mod-elling of Recommender Systems. In ECIR ’15, volume 9022,pages 346–351. Springer, 2015.
RecSys 2015, 9th ACM Conference on Recommender Systems. 16 - 20 September, 2015, Vienna, Austria.