L-INJECTION: TOWARD EFFECTIVE COLLABORATIVFILTERING USING UNINTERESTING ITEMS

 

ABSTRACT

We develop a novel framework, named as l-injection, to address the sparsity problem of recommender systems. Bycarefully injecting low values to a selected set of unrated user-item pairs in a user-item matrix, we demonstrate that top-Nrecommendation accuracies of various collaborative filtering (CF) techniques can be significantly and consistently improved. We firstadopt the notion of pre-use preferences of users toward a vast amount of unrated items. Using this notion, we identify uninterestingitems that have not been rated yet but are likely to receive low ratings from users, and selectively impute them as low values. As ourproposed approach is method-agnostic, it can be easily applied to a variety of CF algorithms. Through comprehensive experimentswith three real-life datasets (e.g., Movielens, Ciao, and Watcha), we demonstrate that our solution consistently and universallyenhances the accuracies of existing CF algorithms (e.g., item-based CF, SVD-based CF, and SVD++) by 2.5 to 5 times on averageFurthermore, our solution improves the running time of those CF methods by 1.2 to 2.3 times when its setting produces the bestaccuracy. The datasets and codes that we used in the experiments are available at: https://goo.gl/KUrmip.

CONCLUSIONS

In this paper, we proposed a novel approach, l-injection,for uninteresting items by using a new notion of pre-usepreferences. This approach not only significantly alleviatesthe data sparsity problem but also effectively prevents thoseuninteresting items from being recommended. Because theproposed approach is method-agnostic, it can be easilyapplied to a wide variety of existing CF methods. Throughcomprehensive experiments, we successfully demonstratedthat the proposed approach is effective and practical, dramaticallyimproving the accuracies of existing CF methods(e.g., item-based CF, SVD-based CF, and SVD++) by 2.5 to5 times. Furthermore, our approach improves the runningtime of those CF methods by 1.2 to 2.3 times when its settingproduces the best accuracy.

REFERENCES

[1] W. Hwang, J. Parc, S. Kim, J. Lee, and D. Lee, “Told you i didn’tlike it: Exploiting uninteresting items for effective collaborativefiltering,” in Proc. of IEEE ICDE, 2016, pp. 349–360.

[2] G. Adomavicius and A. Tuzhilin, “Toward the next generationof recommender systems: A survey of the state-of-the-art andpossible extensions,” IEEE Transactions on Knowledge and DataEngineering, vol. 17, no. 6, pp. 734–749, 2005.

[3] Y. Koren et al., “Matrix factorization techniques for recommendersystems,” IEEE Computer, vol. 42, no. 8, pp. 30–37, 2009.

[4] B. Sarwar et al., “Item-based collaboration filtering recommendationalgorithms,” in Proc. of IEEE WWW, 2001, pp. 285–295.

[5] S. Zhang et al., “Using singular value decomposition approximationfor collaborative filtering,” in Proc. of IEEE CEC, 2005, pp.257–264.

[6] P. Resnick et al., “Grouplens: an open architecture for collaborativefiltering of netnews,” in Proc. of ACM CSCW, 1994, pp. 175–186.

[7] P. Cremonesi et al., “Performance of recommender algorithms ontop-n recommendation tasks,” in Proc. of ACM RecSys, 2010, pp.39–46.

[8] R. Pan et al., “One-class collaborative filtering,” in Proc. of IEEEICDM, 2008, pp. 502–511.

[9] V. Sindhwani et al., “A family of non-negative matrix factorizationfor one-class collaborative filtering,” in Proc. of ACM RecSys, 2009.

[10] J. Ha et al., “Top-n recommendation through belief propagation,”in Proc. of ACM CIKM, 2012, pp. 2343–2346.

[11] N. Srebro and T. Jaakkola, “Weighted low-rank approximations,”in Proc. of AAAI ICML, 2003, pp. 720–727.

[12] J. Breese et al., “Empirical analysis of predictive algorithms forcollaborative filtering,” in Proc. of UAI, 1998, pp. 43–52.

[13] J. Tang, H. Gao, and H. Liu, “mtrust: discerning multi-faceted trustin a connected world,” in ”Proc. of WSDM”, pages = 93–102, year =2012.