Abstract:We address the problem of entry-wise low-rank matrix completion in the noisy observation model. We propose a new noise robust estimator where we characterize the bias and variance of the estimator in a finite sample setting. Utilizing this estimator, we provide a new robust local matrix completion algorithm that outperforms other classic methods in reconstructing large rectangular matrices arising in a wide range of applications such as athletic performance prediction and recommender systems. The simulation re… Show more
“…• RLMC [11]: A new robust local matrix completion algorithm that characterize the bias and variance of the estimator in a finite sample setting. • RegSVD [12]: A rating prediction algorithm based on SVD.…”
Recommendation algorithms based on collaborative filtering show products which people might like and play an important role in personalized service. Nevertheless, the most of them just adopt explicit information feedback and achieve low recommendation accuracy. In recent years, deep learning methods utilize non-linear network framework to receive feature representation of massive data, which can obtain implicit information feedback. Therefore, many algorithms are designed based on deep learning to improve recommendation effects. Even so, the results are unsatisfactory. The reason is that they never consider explicit information feedback. In this paper, we propose a Hybrid Granular Algorithm for Rating Recommendation (HGAR), which is based on granulation computing. The core idea is to explore the multi-granularity of interaction information for both explicit and implicit feedback to predict the users ratings. Thus, we used Singular Value Decomposition model to get explicit information and implicit information can be received by multilayer perception of deep learning. In addition, we fused the two part information when the two models are jointly trained. Therefore, HGAR can explore the multi-granularity of interaction information which learned explicit interaction information and mined implicit information in different information granular level. Experiment results show that HGAR significantly improved recommendation accuracy compared with different recommendation models including collaborative filtering and deep learning methods.
“…• RLMC [11]: A new robust local matrix completion algorithm that characterize the bias and variance of the estimator in a finite sample setting. • RegSVD [12]: A rating prediction algorithm based on SVD.…”
Recommendation algorithms based on collaborative filtering show products which people might like and play an important role in personalized service. Nevertheless, the most of them just adopt explicit information feedback and achieve low recommendation accuracy. In recent years, deep learning methods utilize non-linear network framework to receive feature representation of massive data, which can obtain implicit information feedback. Therefore, many algorithms are designed based on deep learning to improve recommendation effects. Even so, the results are unsatisfactory. The reason is that they never consider explicit information feedback. In this paper, we propose a Hybrid Granular Algorithm for Rating Recommendation (HGAR), which is based on granulation computing. The core idea is to explore the multi-granularity of interaction information for both explicit and implicit feedback to predict the users ratings. Thus, we used Singular Value Decomposition model to get explicit information and implicit information can be received by multilayer perception of deep learning. In addition, we fused the two part information when the two models are jointly trained. Therefore, HGAR can explore the multi-granularity of interaction information which learned explicit interaction information and mined implicit information in different information granular level. Experiment results show that HGAR significantly improved recommendation accuracy compared with different recommendation models including collaborative filtering and deep learning methods.
“…Yet another method that has gained popularity in interpolation is matrix completion (MC) technique, in which the aim is to reconstruct all the data of a given matrix using its low or high rank property by using small [14]. MC technique has been used to interpolate the missing dataset in various fields, like -seismic data processing, image processing, wireless data processing and others [15].…”
Section: Introductionmentioning
confidence: 99%
“…Recent researches are direct and indirect REM generation models are on 2D outdoor without considering the altitude [12][13] [14][24]- [26], but very few works are on indoor 3D environment [20], [27]- [31].…”
In this paper, a novel texture-patch transformed (TPT) three dimensional (3D) matrix completion (MC) method has been proposed with the support of novel 3D measuring points (MPs) locating algorithm to generate practical received signal strength (RSS) database assisted indoor 3D radio environment map (REM) of ultra-high frequency (UHF) television (TV)-band. The exploration of TV-band results in TV white and grey space (TV-WS and TV-GS), which are competent resolution to recoup excess data traffic through cognitive radio networks (CRNs) by dynamic spectrum access (DSA) by secondary user (SU). Maximum wireless data traffic generates in indoor and altitude considered exploration of REM achieves high data rate, so selecting interpolation algorithm is important for getting accurate and timely generated REM. Many MC algorithm shows better results than standard interpolation methods. Instead of using layer-by-layer MC algorithm, TPT-MC algorithm could be used through 3D↔2D conversion. Patch size has been considered through symmetric dataset profile. MC criteria based analysis shows TPT-MC algorithm takes lesser no. of MPs than layer-by-layer MC algorithm. Singular value thresholding (SVT) algorithm is used MC algorithm. TPT-SVT shows advantage over layer-by-layer SVT algorithm on RMSE, correlation, best-fit-line and simulation time on same no. of dataset. The result analysis shows that TPT-SVT algorithm is better in RMSE, closest best-fit-line and correlation coefficient than 2D IDW2, 2D K-NN, 2D kriging, TPT-IDW2, TPT-K-NN, TPT-kriging, 3D IDW2 and layer-by-layer SVT algorithm. Computation time of TPT-SVT is better than 3D IDW2 and SVT. TPT-SVT algorithm takes lesser no. of dataset than SVT algorithm for faithful MC.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.