Normalize one log's knowledge_code field to a list of int ids.

Apply non-negative clipping if needed. In this implementation, discrimination is already positive because of sigmoid, so this

:param stu_id: LongTensor, student ids :param exer_id: LongTensor, exercise ids :return: FloatTensor, probability of answerin

Return student ability theta_u.

Read a two-line config file: header, then student_n,exercise_n,knowledge_n.

RCD threshold: min-max normalize, average, then power four.

Build raw RCD concept-map edges from logs. Returns ``knowledgeGraph``-style edges before reciprocal-edge splitting.

Count RCD-style concept transitions. For two consecutive logs, a transition from concept i to concept j is counted only when both exercises a

1. delete students who have fewer than min_log response logs 2. divide dataset into train_set, val_set and test_set (0.7:0.1:0.2) :return

Yield each student's ordered practice sequence. Supported inputs: - Agent4Edu: [{"user_id": ..., "logs": [...]}, ...] - RCD ASSIST: [{"lo

Load edge pairs from JSON or RCD-style TSV/TXT files.

Split RCD raw graph into prerequisite and similarity edges. Reciprocal pairs are written once as undirected/similarity edges. One-way pairs

测试模型性能，计算准确率、F1、AUC和ROUGE-2。

Split each student's sequence into training, validation, and test sets. - Train: first 80% of data (last 10% of this for validation).

Split each student's sequence into training, validation, and test sets. - Train: first 80% of data (last 10% of this for validation).