Riiid Answer Correctness Prediction - lgb baseline 学习

关注我的公众号YueTan进行交流探讨
欢迎关注数据比赛方案仓库 https://github.com/hongyingyue/Competition-solutions

链接

特征

features = ['user_questions', 'user_mean', 'content_questions', 'content_mean', 'prior_question_elapsed_time']

user_df = train[train.answered_correctly != -1].groupby('user_id').agg({'answered_correctly': ['count', 'mean']}).reset_index()
user_df.columns = ['user_id', 'user_questions', 'user_mean']

content_df = train[train.answered_correctly != -1].groupby('content_id').agg({'answered_correctly': ['count', 'mean']}).reset_index()
content_df.columns = ['content_id', 'content_questions', 'content_mean']

train['prior_question_elapsed_time'].fillna(mean_prior, inplace = True)

得分： 0.75

使用的五个特征含义分别是

• 用户回答所有问题的个数
• 用户回答问题的正确率
• 问题被回答的个数总和
• 问题被回答的正确率
• 用户在上一个question bundle中平均消耗时间

特征2

链接

FEATS = ['answered_correctly_avg_u', 'answered_correctly_sum_u', 'count_u', 'answered_correctly_avg_c', 'part', 'prior_question_had_explanation', 'prior_question_elapsed_time']


def add_user_feats(df, answered_correctly_sum_u_dict, count_u_dict):
    acsu = np.zeros(len(df), dtype=np.int32)
    cu = np.zeros(len(df), dtype=np.int32)
    for cnt,row in enumerate(tqdm(df[['user_id','answered_correctly']].values)):
        acsu[cnt] = answered_correctly_sum_u_dict[row[0]]
        cu[cnt] = count_u_dict[row[0]]
        answered_correctly_sum_u_dict[row[0]] += row[1]
        count_u_dict[row[0]] += 1
    user_feats_df = pd.DataFrame({'answered_correctly_sum_u':acsu, 'count_u':cu})
    user_feats_df['answered_correctly_avg_u'] = user_feats_df['answered_correctly_sum_u'] / user_feats_df['count_u']
    df = pd.concat([df, user_feats_df], axis=1)
    return df

# answered correctly average for each content
content_df = train[['content_id','answered_correctly']].groupby(['content_id']).agg(['mean']).reset_index()
content_df.columns = ['content_id', 'answered_correctly_avg_c']

# changing dtype to avoid lightgbm error
train['prior_question_had_explanation'] = train.prior_question_had_explanation.fillna(False).astype('int8')

得分 0.76

使用的7个特征是

• 使用的特征是增量特征，截止到目前该用户的回答正确率
• 截止到目前该用户回答正确的题目个数
• 截止到目前该用户回答问题所有个数
• content在所有回答中的正确率
• part
• 是否有解释
• 平均消耗时间
  

特征3

链接

FEATURES = ['prior_question_elapsed_time', 'prior_question_had_explanation', 'part',
	 'answered_correctly_u_avg', 'elapsed_time_u_avg', 'explanation_u_avg',
                'answered_correctly_q_avg', 'elapsed_time_q_avg', 'explanation_q_avg',
                 'answered_correctly_uq_count', 'timestamp_u_recency_1', 'timestamp_u_recency_2', 'timestamp_u_recency_3', 
                'timestamp_u_incorrect_recency']

使用的特征14个

特征4

https://www.kaggle.com/code/a763337092/lgb1215

features_dict = {
    #'user_id',
    'timestamp':'float16',#
    'user_interaction_count':'int16',
    'user_interaction_timestamp_mean':'float32',
    'lagtime':'float32',#
    'lagtime2':'float32',
    'lagtime3':'float32',
    #'lagtime_mean':'int32',
    'content_id':'int16',
    'task_container_id':'int16',
    'user_lecture_sum':'int16',#
    'user_lecture_lv':'float16',##
    'prior_question_elapsed_time':'float32',#
    'delta_prior_question_elapsed_time':'int32',#
    'user_correctness':'float16',#
    'user_uncorrect_count':'int16',#
    'user_correct_count':'int16',#
    #'content_correctness':'float16',
    'content_correctness_std':'float16',
    'content_correct_count':'int32',
    'content_uncorrect_count':'int32',#
    'content_elapsed_time_mean':'float16',
    'content_had_explanation_mean':'float16',
    'content_explation_false_mean':'float16',
    'content_explation_true_mean':'float16',
    'task_container_correctness':'float16',
    'task_container_std':'float16',
    'task_container_cor_count':'int32',#
    'task_container_uncor_count':'int32',#
    'attempt_no':'int8',#
    'part':'int8',
    'part_correctness_mean':'float16',
    'part_correctness_std':'float16',
    'part_uncor_count':'int32',
    'part_cor_count':'int32',
    'tags0': 'int8',
    'tags1': 'int8',
    'tags2': 'int8',
    'tags3': 'int8',
    'tags4': 'int8',
    'tags5': 'int8',
   # 'tags6': 'int8',
   # 'tags7': 'int8',
#     'tags0_correctness_mean':'float16',
#     'tags1_correctness_mean':'float16',
#     'tags2_correctness_mean':'float16',
#     'tags4_correctness_mean':'float16',
#     'bundle_id':'int16',
#     'bundle_correctness_mean':'float16',
#     'bundle_uncor_count':'int32',
#     'bundle_cor_count':'int32',
    'part_bundle_id':'int32',
    'content_sub_bundle':'int8',
    'prior_question_had_explanation':'int8',
    'explanation_mean':'float16', #
    #'explanation_var',#
    'explanation_false_count':'int16',#
    'explanation_true_count':'int16',#
   # 'community':'int8',
#     'part_1',
#     'part_2',
#     'part_3',
#     'part_4',
#     'part_5',
#     'part_6',
#     'part_7',
#     'type_of_concept',
#     'type_of_intention',
#     'type_of_solving_question',
#     'type_of_starter'
}
categorical_columns= [
    #'user_id',
    'content_id',
    'task_container_id',
    'part',
   # 'community',
    'tags0',
    'tags1',
    'tags2',
    'tags3',
    'tags4',
    'tags5',
    #'tags6',
    #'tags7',
    #'bundle_id',
    'part_bundle_id',
    'content_sub_bundle',
    'prior_question_had_explanation', 
#     'part_1',
#     'part_2',
#     'part_3',
#     'part_4',
#     'part_5',
#     'part_6',
#     'part_7',
#     'type_of_concept',
#     'type_of_intention',
#     'type_of_solving_question',
#     'type_of_starter'
]

features=list(features_dict.keys())

top solution

https://www.kaggle.com/competitions/riiid-test-answer-prediction/discussion/209597

• Firstly, the data are sorted by [‘user_id’, ‘timestamp’, ‘content_id’]
• created features in different array via self-designed rolling function or self-designed cumlative function
• use catboost

['content_id',
 'prior_question_elapsed_time',
 'prior_question_had_explanation',
 'correct_answer',
 'user_count',
 'user_sum',
 'user_mean',
 'item_count',
 'item_sum',
 'item_mean',
 'answer_ratio_0',
 'answer_ratio_1',
 'answer_ratio_2',
 'bundle_id',
 'part',
 'le_tag',
 'question_correct_user_ablility_mean',
 'question_correct_user_ablility_median',
 'question_wrong_user_ablility_mean',
 'question_wrong_user_ablility_median',
 'word2vec_0',
 'word2vec_1',
 'word2vec_2',
 'word2vec_3',
 'word2vec_4',
 'svd_0',
 'svd_1',
 'svd_2',
 'svd_3',
 'svd_4',
 'tags_w2v_correct_mean_0',
 'tags_w2v_wrong_mean_0',
 'tags_w2v_correct_mean_1',
 'tags_w2v_wrong_mean_1',
 'tags_w2v_correct_mean_2',
 'tags_w2v_wrong_mean_2',
 'tags_w2v_correct_mean_3',
 'tags_w2v_wrong_mean_3',
 'tags_w2v_correct_mean_4',
 'tags_w2v_wrong_mean_4',
 'real_time_wrong_mean',
 'real_time_wrong_median',
 'real_time_correct_mean',
 'real_time_correct_median',
 'task_set_distance_wrong_mean',
 'task_set_distance_wrong_median',
 'task_set_distance_correct_mean',
 'task_set_distance_correct_median',
 'mean_0_ratio',
 'mean_1_ratio',
 'mean_3_ratio',
 'mean_4_ratio',
 'mean_5_ratio',
 'mean_6_ratio',
 'mean_7_ratio',
 'mean_8_ratio',
 'mean_9_ratio',
 'mean_10_ratio',
 'user_d1',
 'user_d2',
 'task_set_distance',
 'user_diff_mean',
 'user_diff_std',
 'user_diff_min',
 'user_diff_max',
 'task_set_item_mean',
 'task_set_item_min',
 'task_set_item_max',
 'task_set_distance2',
 'task_distance_shift',
 'task_set_distance_diff',
 'task_distance_diff_shift',
 'container_mean_1',
 'container_mean_5',
 'container_std_5',
 'container_mean_10',
 'container_std_10',
 'container_mean_20',
 'container_std_20',
 'container_mean_30',
 'container_std_30',
 'container_mean_40',
 'container_std_40',
 'prior_question_elapsed_time_mean_1',
 'prior_question_elapsed_time_mean_5',
 'prior_question_elapsed_time_mean_10',
 'prior_question_elapsed_time_mean_20',
 'prior_question_elapsed_time_mean_30',
 'prior_question_elapsed_time_mean_40',
 'item_mean_mean_30',
 'item_mean_mean_40',
 'task_set_distance_mean_1',
 'task_set_distance_mean_5',
 'task_set_distance_mean_10',
 'task_set_distance_mean_20',
 'task_set_distance_mean_30',
 'begin_time_diff',
 'end_time_diff',
 'part_time_diff_mean',
 'part_session_mean',
 'part_session_sum',
 'part_session_count',
 'full_group0_item_mean_mean',
 'full_group0_item_mean_median',
 'full_group0_task_set_distance_median',
 'full_group0_timestamp_mean',
 'full_group0_timestamp_median',
 'full_group1_item_mean_mean',
 'full_group1_item_mean_median',
 'full_group1_task_set_distance_median',
 'full_group1_timestamp_median',
 'part_sum',
 'part_count',
 'part_mean',
 'part_sum_global_ratio',
 'part_sum_1',
 'part_sum_5',
 'part_mean_5',
 'part_sum_10',
 'part_mean_10',
 'cum_answer0_mean_item_mean',
 'cum_answer0_median_item_mean',
 'cum_answer0_median_task_set_distance',
 'cum_answer1_mean_item_mean',
 'cum_answer1_median_item_mean',
 'cum_answer1_mean_task_set_distance',
 'cum_answer1_median_task_set_distance',
 'cum_answer0_time_diff',
 'cum_answer1_time_diff',
 'global_task_set_shift1',
 'global_task_set_shift2',
 'global_task_set_shift4',
 'global_task_set_shift5',
 'cum_answer0_mean_wrong_time_diff',
 'cum_answer0_median_wrong_time_diff',
 'cum_answer1_mean_right_time_diff',
 'content_correct_mean',
 'content_correct_sum',
 'content_correct_count',
 'hard_answer0_time',
 'hard_answer1_time',
 'full_bundle_item_mean_mean',
 'full_bundle_item_mean_median',
 'full_bundle_task_set_distance_mean',
 'full_bundle_task_set_distance_median',
 'full_bundle_timestamp_mean',
 'full_bundle_timestamp_median',
 'bundle_sum',
 'bundle_mean',
 'bundle_count',
 'user_trend_mean',
 'user_trend_median',
 'user_trend_roll_user_ans_sum',
 'user_trend_roll_user_ans_mean',
 'user_trend_roll_user_ans_count',
 'user_trend_roll_item_ans_mean',
 'user_trend_roll_item_ans_count',
 'div_ratio1',
 'div_ratio2',
 'div_ratio3',
 'new_Feat0',
 'new_Feat1',
 'new_Feat2',
 'new_Feat3',
 'part_time_wrong_div',
 'part_time_right_div',
 'diff_lag_median_div',
 'diff_item_median_div',
 'diff_time_median_div',
 'diff_item_mean_div',
 'diff_task_set_mean_div',
 'diff_timestamp_mean_div',
 'last_20_frequent_answer',
 'last_20_frequent_answer_count',
 'last_20_frequent_answer_mean',
 'last_20_frequent_answer_sum',
 'last_user_same_answer_tf',
 'last_item_same_answer_tf',
 'last_right_time_diff',
 'last_wrong_time_diff',
 'last_5_part_time_div',
 'last_10_part_time_div',
 'last_20_part_time_div']

https://www.kaggle.com/code/tomooinubushi/62nd-solution-lightgbm-single-model-lb-0-801