Action Space Learning for Heterogeneous User Behavior Prediction

Authors: Dongha Lee, Chanyoung Park, Hyunjun Ju, Junyoung Hwang, Hwanjo Yu

IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that METAS outperforms the state-ofthe-art methods in predicting users heterogeneous actions, and its entity space represents the useruser and item-item similarities more clearly than the space trained by the other methods.
Researcher Affiliation Academia 1Pohang University of Science and Technology, Pohang, Republic of Korea 2University of Illinois at Urbana-Champaign, Urbana, IL, USA
Pseudocode Yes Algorithm 1: Hard triplet mining algorithm
Open Source Code No The paper states 'We implement METAS and CML using Tensorflow [Abadi et al., 2016] to run on GPU, and utilize the JAVA source code of SPTF released by [Yin et al., 2017].' This indicates they used existing open-source tools but does not provide a link or explicit statement about releasing their own implementation code for METAS.
Open Datasets Yes Tmall is the behavior dataset collected by Tmall of Alibaba, one of the biggest e-commerce platforms, and it is published by [Yin et al., 2017]. Taobao is also users-commodities behavior data on Alibaba s mobilecommerce platforms, publicly available from Ali Mobile Recommendation Algorithm Competition.
Dataset Splits Yes For each user, we leave out a single observed action per action type for testing, and use the rest for training. In our experiments, we leave out an additional observed action for a validation set. ... For each dataset, we tune the hyperparameters by using a grid search and use the optimal values that show the best H@10 on the validation set.
Hardware Specification No The paper states 'We implement METAS and CML using Tensorflow [Abadi et al., 2016] to run on GPU'. This is a general statement and does not specify any particular GPU model, CPU, or other hardware details.
Software Dependencies No The paper mentions 'Tensorflow [Abadi et al., 2016]' and 'Adam optimizer [Kingma and Ba, 2014]' but does not provide specific version numbers for these software dependencies. It also mentions 'JAVA source code of SPTF released by [Yin et al., 2017]'.
Experiment Setup Yes The MLP networks of METAS are equipped with a single layer, 50% dropout probability, and Re LU activation. For each dataset, we tune the hyperparameters by using a grid search and use the optimal values that show the best H@10 on the validation set. We set the margin size α = 2 and the mini-batch size (i.e., the number of triplets in a mini-batch) b = 200. ... We set D1 = D2 = D to reduce the model complexity, and investigate the performance changes with respect to the dimension size D with a range of {50, 100, 150, 200, 250}. ... In Algorithm 1, we set the candidate itemset size c to the size of mini-batch, and n = 3.