thuwangcy / hawkeskt Goto Github PK

Hi @THUwangcy, thanks for sharing your work.

While all other models use [1:] slicing to make predictions and labels,

The ASSISTment 09-10 dataset has a field order_id, which is explained on the official website as: these id's are chronological, and refer to the id of the original problem log.

So for the processing of this dataset, after grouping by user_id, should we sort by 'order_id', otherwise it will destroy the chronological order of each user's answer. Although the preprocessing part constructs the timestamp, it cannot completely guarantee the user's question order. After each user's problem is sorted by order_id, the result of the program run has changed.

您好！
您提出的HawkesKT模型让人耳目一新，关于代码仓库中对AKT模型的复现我有一些小问题。
在您复现的代码中

question_seq只被embedding为diff_parm，但是在AKT的官方仓库中，question会被embedding为 (n_question+1, embed_l)。

期待您的回复。

I noticed a weird way of invoking the TranformerLayers in SAKT model as per the below code snippet, you are only getting the output of the last layer applied on the seq_data input, you should either combine the pool or add y as input per each call!

y = seq_data for block in self.attn_blocks: y = block(mask=1, query=q_data, key=seq_data, values=seq_data)

Can you give me the slepemapy.ca data set？ please