Dynamic Tensor Recommender Systems
Authors: Yanqing Zhang, Xuan Bi, Niansheng Tang, Annie Qu
JMLR 2021 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The proposed method is applied to simulations, IRI marketing data and Last.fm data. Numerical studies demonstrate that the proposed method outperforms existing methods in terms of future time forecasting. In theory, we establish the convergence rate of the proposed tensor factorization and asymptotic normality of the spline coefficient estimator. |
| Researcher Affiliation | Academia | Yanqing Zhang EMAIL Department of Statistics Yunnan University Kunming, 650504, China Xuan Bi EMAIL Carlson School of Management University of Minnesota Minneapolis, MN, 55455-0438, USA Niansheng Tang EMAIL Department of Statistics Yunnan University Kunming, 650504, China Annie Qu EMAIL Department of Statistics University of California Irvine, CA, 92697-3425, USA |
| Pseudocode | Yes | Algorithm Implementation Algorithm 1: (Initialization) Input all observed yi1i2 id(t) s, the number of factors r, tuning parameter λ, initial value θ0 and a stopping criterion ε = 10 4. 2: (Individual factors update) At the lth iteration, estimate {P1, P2, , Pd, α}. (i) For each Pk, solve (4) through parallel computing and obtain b Pk . Then calculate Jb Pk through (8). (ii) For α, solve (6) and obtain bα . Then calculate Jˆα through (8). (iii) Assign b Pk l b Pk , if Jb Pk = max{Jb P1 , Jb P2 , , Jb Pd , Jˆα }. bα(l) bα , if Jˆα = max{Jb P1 , Jb P2 , , Jb Pd , Jˆα }. 3: (Subgroup factors update) At the lth iteration, estimate {q(1), q(2), , q(d), β}. (i) For every q(k), solve (5) and obtain bq(k) . Then calculate Jbq(k) through (8). (ii) For β, solve (7) and obtain bβ . Then calculate Jˆβ through (8). (iii) Assign bq(k) l bq(k) , if Jbq(k) = max{Jbq(1) , Jbq(2) , , Jbq(d) , Jˆβ }. bβ(l) bβ , if Jˆβ = max{Jbq(1) , Jbq(2) , , Jbq(d) , Jˆβ }. 4: (Stopping Criterion) Stop if max{Jb P1 , Jb P2 , , Jb Pd , Jˆα , Jbq(1) , , Jbq(d) , Jˆβ } < ε. Set the final estimator bθ = bθl. Otherwise set l l + 1 and go to step 2. |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described. It only mentions the license for the paper itself. |
| Open Datasets | Yes | In Section 6, we apply the proposed method to the IRI marketing data and Last.fm data. IRI Marketing Data (Bronnenberg et al., 2008). The Lastfm-1K dataset collected by Last.fm API (Celma, 2010) to evaluate the performance of the proposed method. The dataset is available at http: //ocelma.net/Music Recommendation Dataset/lastfm-1K.html |
| Dataset Splits | Yes | In each simulation, we consider the number of time points as T = T1 + T2, where the tensor data in the first T1 = 12 time points are set as the training data, and the tensor data in the last T2 time points are used as the testing data. For evaluating the forecasting performance at future time points, we consider T2 = 8 or 12. ... For selecting the above parameters, we set the data from the beginning of 2001 to the end of 2009 (i.e., the first 108 months) as the training set and the data from the beginning of 2010 to the end of 2010 as the validation set... We randomly sample 80% of the data from 2001 to 2010 as a training set and 20% of the data from the entire year of 2011 as the testing set. The random sampling is replicated 50 times. ... We randomly sample 80% of the data from February of 2005 to May of 2008 (i.e., the first 40 months) as a training set and 20% of the data from June of 2008 to June of 2009 (i.e., the last 13 months) as the testing set. The random sampling is replicated 50 times. |
| Hardware Specification | Yes | All experiments are implemented using Window 10 with 1.99 GHz Intel Core i7 Processor and 16 GB memory. |
| Software Dependencies | No | The paper mentions implementing experiments but does not specify any software libraries, frameworks, or their version numbers. |
| Experiment Setup | Yes | The number of latent factors is set as r = 3. ... For all methods, we select the number of latent factors r ranging from 3 to 30. For the REM method and the proposed methods, we select a tuning parameter λ from 1 to 29. For the BPTF methods, we use the default values of the remaining parameters. ... In our numerical study, we set κ = 2 and adopt truncated polynomial bases. |