randolphvi/next-basket-recommendation

About Next Basket Recommendations Based on Neural Network.

🍻 Welcome stranger

• 🎓: A Ph.D graduated from University of Science and Technology of China (USTC).
• 🎯: Interested in NLP & Education area. Mainly Focusing on Hierarchical Multi-label Classification Tasks.
• 👾: I'm a die-hard fan of the Fallout ☢️ and Witcher 🐺 series, and I especially love CRPG games & indie games. Feel free to add my Steam.

🧰 My Projects

• 🛠 Multi-label Classification: a collection of deep neural network models for beginners, designed to solve Multi-label Classification problems, built in TensorFlow.
• 🛠 Semantic Textual Similarity: a collection of deep neural network models for beginners, designed to solve Semantic Textual Similarity problems, built in TensorFlow.
• 📚 Question Difficulty Prediction: a collection of deep neural network models for Question Difficulty Prediction problems, built in TensorFlow & PyTorch.
• 🍻 HARNN: an attention-based recurrent network Approach for Hierarchical Multi-label Text Classification (HMC), accepted at CIKM 2019 and built in TensorFlow. [Papers] | [Code]
• 🍻 HmcNet: a general approach that incorporates both explicit and implicit class hierarchy constraints, designed to solve Hierarchical Multi-label Text Classification (HMC), accepted at TKDE 2022 and built in PyTorch. [Papers] | [Code]
• 🍻 HMNet: a multi-modal method that captures class dependencies to solve Educational Video Concept Prediction problems, accepted at JMLC 2023. [Papers]

🎉 News

• [2023 Aug] Gods Among Us, Baldur's Gate 3!
• [2023 Jun] I've suffered in the Darkest Dungeon, and no one is spared!
• [2023 May] I've been engrossed in the The Legend of Zelda: Tears of the Kingdom, Nintendo rules the fxxking world!
• [2023 Apr] I've cleared Octopath Traveler II, スクエニさん、さすがかよ!

All Works

2023

• Hong Yuting, Shiwei Tong, Wei Huang, Yan Zhuang, Qi Liu, Enhong Chen, et al. Search-Efficient Computerized Adaptive Testing, CIKM'2023, Accepted.
• Wei Huang, Tong Xiao, Qi Liu, Zhenya Huang, et al. HMNet: A Hierarchical Multi-modal Network for Educational Video Concept Prediction, JMLC'2023, 2023: 1-12.

2022

• Wei Huang, Enhong Chen, Qi Liu, Hui Xiong, Zhenya Huang, Shiwei Tong, et al. HmcNet: A General Approach for Hierarchical Multi-label Classification, TKDE'2022, Accepted.
• Shuanghong Shen, Qi Liu, Enhong Chen, Zhenya Huang, Wei Huang, et al. Monitoring Student Progress for Learning Process-consistent Knowledge Tracing, TKDE'2022, Accepted.
• Jiatong Li, Fei Wang, Qi Liu, Mengxiao Zhu, Wei Huang, et al. HierCDF: A Bayesian Network-based Hierarchical Cognitive Diagnosis Framework, KDD'2022, 2022: 904-913.
• Shiwei Tong, Jiayu Liu, Yuting Hong, Zhenya Huang, Le Wu, Qi Liu, Wei Huang, et al. Incremental Cognitive Diagnosis for Intelligent Education, KDD'2022, 2022: 1760-1770.
• Yuren Zhang, Enhong Chen, Binbin Jin, Hao Wang, Min Hou, Wei Huang and Runlong Yu. Clustering based Behavior Sampling with Long Sequential Data for CTR Prediction, SIGIR'2022, 2022: 2195-2200.
• Zheng Gong, Shiwei Tong, Han Wu, Qi Liu, Hanqing Tao, Wei Huang, et al. Tipster: A Topic-Guided Language Model for Topic-Aware Text Segmentation, DASFAA'2022, 2022: 213-221.

2021

• Siqi Lei, Wei Huang, Shiwei Tong, Qi Liu, Zhenya Huang, Enhong Chen, et al. Consistency-aware Multi-modal Network for Hierarchical Multi-label Classification in Online Education System, Best Student Paper, ICBK'2021, 2021: 1-8.
• Ye Huang, Wei Huang, Shiwei Tong, et al. STAN: Adversarial Network for Cross-domain Question Difficulty Prediction, ICDM'2021, 2021: 220-229.
• Shuanghong Shen, Qi Liu, Enhong Chen, Zhenya Huang, Wei Huang, et al. Learning Process-consistent Knowledge Tracing, KDD'2021, 2021: 1452-1460.
• Shiwei Tong, Qi Liu, Runlong Yu, Wei Huang, Zhenya Huang, Zachary A. Pardos, Weijie Jiang, Item Response Ranking for Cognitive Diagnosis, IJCAI'2021, 2021: 1750-1756.

2020

• Wei Tong, Shiwei Tong, Wei Huang, et al. Exploiting Knowledge Hierarchy for Finding Similar Exercises in Online Education Systems, ICDM'2020, 2020: 1298-1303.
• Shiwei Tong, Qi Liu, Wei Huang, et al. Structure-based Knowledge Tracing: An Influence Propagation View, ICDM'2020, 2020: 541-550.
• Xin Wang, Wei Huang, Qi Liu, et al. Fine-Grained Similarity Measurement between Educational Videos and Exercises, ACM MM'2020, 2020: 331-339.
• Yang Liu, Zhi Li, Wei Huang, Tong Xu, Enhong Chen. Exploiting Structural and Temporal Influence for Dynamic Social-Aware Recommendation, JCST'2020, 2020, 35(2), 281–294.

2019

• Wei Huang, Qi Liu, Enhong Chen, et al. Hierarchical Multi-label Text Classification: An Attention-based Recurrent Network Approach, CIKM’2019, 2019: 1051-1060.

next-basket-recommendation's Issues

Can you upload the file "neg_sample.pickle"?

Hi, dear RandolphVI,

I trying to run your code, the Anaconda with Python 3.6 (3.6.9) was already downloaded and installed, as well as, all needed frameworks also were installed... But after I run the train.py in Spyder the code ended with errors, because there is no "neg_sample.pickle" in the "data" folder. So, can you upload this file? Or can you give the example of its structure, from which one can understand what should be inside this file?

P.S.: I used your files "test_sample.json", "train_sample.json" and "validation_sample.json" as corresponding files mentioned in "config.py" by "self.TESTSET_DIR", "self.TRAININGSET_DIR" and "self.VALIDATIONSET_DIR". The only file that is absent is "neg_sample.pickle".

Sincerely Yours,
Sergey.

prediction after training

Did any figure out how to perform prediction after training?

Some questions about the operation of the algorithm

@RandolphVI, can you recommend, please, what paramemers in config.py should be changed
to reduce RAM memory usage? The initial preprocessing of the data (time after start of the train.py and before the start of 1st epoch's calculation) consumes 7 min (200 epoches consumes 1-1,5 hours) and my pc is almost unavailable due to memory overusage. So, it's better to reduce this processing time and RAM usage. Can you give advice how? There are parameters in Config.py:

self.cuda = False
self.clip = 10
self.epochs = 1# = 200
self.batch_size = 256
self.seq_len = 12
self.learning_rate = 0.01  # Initial Learning Rate
self.log_interval = 1  # num of batches between two logging
self.basket_pool_type = 'avg'  # ['avg', 'max']
self.rnn_type = 'LSTM'  # ['RNN_TANH', 'RNN_RELU', 'LSTM', 'GRU']
self.rnn_layer_num = 2
self.dropout = 0.5
self.num_product = 26991+1  # 商品数目，用于定义 Embedding Layer - Number of products, used to define Embedding Layer
self.embedding_dim = 32  # 商品表征维数， 用于定义 Embedding Layer - Product characterization dimension to define Embedding Layer
self.neg_num = 500  # 负采样个数 - Number of negative samples
self.top_k = 10  # Top K 取值 - Top K Value

May be, "top_k" value or "embedding_dim" should be reduced? Or something like this?
Parameters can be linked with each other and, possibly, should be changed together,
and I doubt to change them, because I don't know the code such precisely now.

P.S.: I observed problems with the code (in Win OS) and made these changes to fix them:

logging.info("!!! DREAM Model Training...") #"✔︎ DREAM Model Training..."
# logger = dh.logger_fn("torch-log", "logs/training-{0}.log".format(time.asctime()))
logger = dh.logger_fn("torch-log", "logs/Training {0}.log".format(time.strftime("%a %d-%b-%Y %H.%M.%S")))

And after I replaced all symbols ✔, ☛ and ✘ the code run like clockwork.
May be, these changes should be added to the next update in the repo?

UPD. I already figured out the question.

runing your code on gpu

hi i read your code and need too run on gpu but your model in train just working on cpu can you change
torch.mm(du_latest, item_embedding[negtives].t()).data.numpy()[0]

Prediction

How to make the prediction after training the model?

Where can I find the full ‘’ratingT12.csv‘’ dataset

Sorry to bother, is there a way to get access to the original dataset？

caculation for hitratio and ndcg

The hitratio and ndcg will be influenced by neg sample numbers. I think the calculation method for these two metrics might be wrong in some way.
The example is when neg sample is 1, the ndcg will larger than 1 and hitratio will always be 1, and when neg sample is rather small, the hitratio will be much smaller.