UniSRec icon indicating copy to clipboard operation
UniSRec copied to clipboard

[KDD'22] Official PyTorch implementation for "Towards Universal Sequence Representation Learning for Recommender Systems".

Results 9 UniSRec issues
Sort by recently updated
recently updated
newest added

Hello, I am currently working on predicting ratings by calculating the dot product of user embeddings and item embeddings. I am considering using the Universal Sequence Representation module for user...

Interested in the pretraining process of UniSRec, I followed the instruction of README.md to pretrain in multi gpus. All the codes were downloaded from this github correctly. ![image](https://github.com/RUCAIBox/UniSRec/assets/97323096/7ccc899b-9dbf-4e03-b0f3-4c1261964b35) I run...

当我执行pretrain.py的pretrain_dataset = dataset.build()[0]这一行代码,报错new_data[k] = torch.LongTensor(value) TypeError: new(): data must be a sequence (got builtin_function_or_method)。于是我追踪调试,发现是rocbole源码的dataset.py这个文件中2213行出现错误,其中描述的是value = data[k].values,我想原作者应该是想让value是一个张量,但是我调试发现data[k]本身就是一个Tensor了,data[k].values得到的反而是一个方法,而不是一个Tensor,所以导致2136行:new_data[k] = torch.LongTensor(value),报错:new_data[k] = torch.LongTensor(value) TypeError: new(): data must be a sequence (got builtin_function_or_method)。 请问大家遇到过这个问题吗?我的做法是手动修改了value...

你好,我按照默认参数复现了一下Scientific数据集,发现了一些问题: 1. 从头开始训练要比pretrain-finetune效果更好。 2. 预训练后,finetune部分参数要比finetune所有参数好,说明finetune会过拟合。 ![image](https://github.com/RUCAIBox/UniSRec/assets/56312792/9172c81b-c9ff-4299-aa2c-4497c8599828) 又仔细研究了一下,发现pretrain之后,固定了SASrec,所以才会出现pretrain对比任务对下游任务影响较大。模型参数很小,这样做合理吗? 如果训练bert,我认为预训练是必要的,但是不需要训练bert,参数量很小,从头训练效果更好,预训练是不是多此一举? ![image](https://github.com/RUCAIBox/UniSRec/assets/56312792/648da169-9684-4345-b4cf-a40807583b3a)

When I run Pre-train from scratch, I face this error. I am looking forward to hearing your reply. Thanks

I would like to know what the purpose of this code: path=`pwd` for ds in Food Home CDs Kindle Movies do ln -s ${path}/../downstream/${ds}/${ds}.feat1CLS ../pretrain/FHCKM/ ln -s ${path}/../downstream/${ds}/${ds}.feat2CLS ../pretrain/FHCKM/ done

when i deal with the dataset preprocessin, i find that this link can't be find [http://deepyeti.ucsd.edu/jianmo/amazon/index.html]

我想请问一下: *.feat1CLS后缀名表示项目文本的bert表示, *.inter表示什么? *.index表示原始ID的重映射。 *.pth表示训练模型后缀名文件吗? > 不知道我理解的对不对,如果不对请告诉我一下,好嘛?

当自定义数据集最长交互序列小于50时,复现模型时无法补齐到50