EdNet数据集分析

https://edudata.readthedocs.io/en/latest/build/blitz/EdNet_KT1/EdNet_KT1.htmlhttps://edudata.readthedocs.io/en/latest/build/blitz/EdNet_KT1/EdNet_KT1.html知识追踪数据集介绍_sereasuesue的博客-CSDN博客_知识追踪数据集最近准备用这个数据集进行跑数据，但发现对这个数据集不够了解，花费了很多时间，最近偶然发现了几篇介绍文章特整理

Field	Annotation
user_id	student’s id
timestamp	the moment the question was given, represented as Unix timestamp in milliseconds
solving_id	represents each learning session of students corresponds to each bunle. It is a form of single integer, starting from 1
question_id	the ID of the question that given to student, which is a form of q{integer}
user_answer	the answer that the student submitted, recorded as a character between a and d inclusively
elapsed_time	the time that the students spends on each question in milliseconds

We randomly selected 5000 tables from all the students for analysis,which accounted for about 0.64% of the total data set, and added a column named user_id to the original table

import os
path=r'D:\EdNet-KT1\KT1'
d=[]
table_list=[]
s=pd.Series(os.listdir(path))
file_selected=s.sample(5000).to_numpy()
for file_name in file_selected:
    data_raw=pd.read_csv(path+'\\'+file_name,encoding = "ISO-8859-15")
    data_raw['user_id']=pd.Series([file_name[:-4]]*len(data_raw))
    d.append([file_name[:-4],len(data_raw)])
    data=pd.DataFrame(data_raw,columns=['user_id']+data_raw.columns.to_list()[:-1])
    table_list.append(data)
df=pd.concat(table_list)
pd.set_option('display.max_rows',10)
df=df.reset_index(drop=True)
df

	user_id	timestamp	solving_id	question_id	user_answer	elapsed_time
0	u717875	1565332027449	1	q4862	d	45000
1	u717875	1565332057492	2	q6747	d	24000
2	u717875	1565332085743	3	q326	c	25000
3	u717875	1565332116475	4	q6168	a	27000
4	u717875	1565332137148	5	q847	a	17000
...	...	...	...	...	...	...
574251	u177603	1530371808931	15	q6984	b	44250
574252	u177603	1530372197614	16	q7335	c	95750
574253	u177603	1530372198181	16	q7336	a	95750
574254	u177603	1530372198879	16	q7337	c	95750
574255	u177603	1530372199425	16	q7338	b	95750

574256 rows × 6 columns

General Feature

df.describe()

	timestamp	solving_id	elapsed_time
count	5.742560e+05	574256.000000	5.742560e+05
mean	1.546425e+12	875.902859	2.599017e+04
std	2.019656e+10	1941.978009	3.376126e+04
min	1.494451e+12	1.000000	0.000000e+00
25%	1.531720e+12	77.000000	1.600000e+04
50%	1.548410e+12	311.000000	2.100000e+04
75%	1.564817e+12	900.000000	3.000000e+04
max	1.575306e+12	18039.000000	7.650000e+06

len(df.question_id.unique())

11838

This shows there are totally 11838 questions.

Missing Value

print('Part of missing values for every column')
print(df.isnull().sum() / len(df))

Part of missing values for every column
user_id         0.000000
timestamp       0.000000
solving_id      0.000000
question_id     0.000000
user_answer     0.000556
elapsed_time    0.000000
dtype: float64

This indicates that there are no missing values in all columns except user_answer. A missing value in user_answer indicates that some students did not choose an option.

df.fillna('not choose',inplace=True)

Fill in not choose in the position of the missing value

Sort user_id

user_count_table=pd.DataFrame(d,columns=['user_id','count'])
ds=user_count_table.sort_values(by=['count'],axis=0).tail(40)
fig = px.bar(
    ds,
    x = 'count',
    y = 'user_id',
    orientation='h',
    title='Top 40 active students'
)

fig.show("svg")

We use the number of questions that students have done as an indicator of whether a student is active. This figure shows the 40 most active students.

ds=df.loc[:,['user_id','elapsed_time']].groupby('user_id').mean()
ds=ds.reset_index(drop=False)
ds.columns=['user_id','avg_elapsed_time']
ds_tail=ds.sort_values(by=['avg_elapsed_time'],axis=0).tail(40)

fig_tail = px.bar(
    ds_tail,
    x = 'avg_elapsed_time',
    y = 'user_id',
    orientation='h',
    title='Bottom 40 fast-solving students '
)
fig_tail.show("svg")
ds_head=ds.sort_values(by=['avg_elapsed_time'],axis=0).head(40)
fig_head = px.bar(
    ds_head,
    x = 'avg_elapsed_time',
    y = 'user_id',
    orientation='h',
    title='Top 40 fast-solving students'
)
fig_head.show("svg")

We take the average time it takes students to do a question as an indicator of how fast students do it. These two figures respectively show the fastest and slowest students among the 5000 students, and the average time they spent doing the problems 

 

Note that some students spend very little time doing the questions, and the time is almost zero. We can almost judge that these students did not do the questions at all, and they chose blindly. We remove these students and rearrange them

bound=5000 # If the average time of doing the topic is less than 5000, it means that the student is most likely to be bad
ds=df.loc[:,['user_id','elapsed_time']].groupby('user_id').mean()
ds=ds.reset_index(drop=False)
ds.columns=['user_id','avg_elapsed_time']
bad_user_ids=ds[ds['avg_elapsed_time']<bound]['user_id'].to_list()
df_drop=df.drop(df[df['user_id'].isin(bad_user_ids)].index)
print('bad students number is ',len(bad_user_ids))
print('length of table after dropping is ',len(df_drop))

bad students number is  61
length of table after dropping is  567778

After dropping

ds=df_drop['user_id'].value_counts().reset_index(drop=False)
ds.columns=['user_id','count']
ds_tail=ds.sort_values(by=['count'],axis=0).tail(40)
fig_tail = px.bar(
    ds_tail,
    x = 'count',
    y = 'user_id',
    orientation='h',
    title='Top 40 active students after dropping some students'
)
fig_tail.show("svg")

This figure shows the 40 most active students after dropping some bad students.

ds=df_drop.loc[:,['user_id','elapsed_time']].groupby('user_id').mean()
ds=ds.reset_index(drop=False)
ds.columns=['user_id','avg_elapsed_time']

ds_head=ds.sort_values(by=['avg_elapsed_time'],axis=0).head(40)
fig_head = px.bar(
    ds_head,
    x = 'avg_elapsed_time',
    y = 'user_id',
    orientation='h',
    title='Top 40 fast-solving students after dropping some students'
)
fig_head.show("svg")

This figure respectively show the more reasonable fastest students among the 5000 students than before, and the average time they spent doing the problems.

Sort question_id

ds=df.loc[:,['question_id','elapsed_time']].groupby('question_id').mean()
ds=ds.reset_index(drop=False)
ds_tail=ds.sort_values(by=['elapsed_time'],axis=0).tail(40)
fig_tail = px.bar(
    ds_tail,
    x = 'elapsed_time',
    y = 'question_id',
    orientation='h',
    title='Top 40 question_id by the average of elapsed_time'
)
fig_tail.show("svg")
ds_head=ds.sort_values(by=['elapsed_time'],axis=0).head(40)
fig_head = px.bar(
    ds_head,
    x = 'elapsed_time',
    y = 'question_id',
    orientation='h',
    title='Bottom 40 question_id by the average of elapsed_time'
)
fig_head.show("svg")

 

 

We can judge the difficulty of this question from the average time spent on a question.

These two figures reflect the difficulty of the questions and shows the ids of the 40 most difficult and 40 easiest questions.s

Appearence of Questions

ds=df['question_id'].value_counts().reset_index(drop=False)
ds.columns=['question_id','count']
ds_tail=ds.sort_values(by=['count'],axis=0).tail(40)
fig_tail = px.bar(
    ds_tail,
    x = 'count',
    y = 'question_id',
    orientation='h',
    title='Top 40 question_id by the number of appearance'
)
fig_tail.show("svg")
ds_head=ds.sort_values(by=['count'],axis=0).head(40)
fig_head = px.bar(
    ds_head,
    x = 'count',
    y = 'question_id',
    orientation='h',
    title='Bottom 40 question_id by the number of appearance'
)
fig_head.show("svg")

These two images reflect the 40 questions that were drawn the most frequently and the 40 questions that were drawn the least frequently

ds2=df['question_id'].value_counts().reset_index(drop=False)
ds2.columns=['question_id','count']
def convert_id2int(x):
    return pd.Series(map(lambda t:int(t[1:]),x))
ds2['question_id']=convert_id2int(ds2['question_id'])
ds2.sort_values(by=['question_id'])
fig = px.histogram(
    ds2,
    x = 'question_id',
    y = 'count',
    title='question distribution'
)
fig.show("svg")

Question’s Option Selected Most Frequently

ds=df.loc[:,['question_id','user_answer','user_id']].groupby(['question_id','user_answer']).count()

most_count_dict={}
for id in df.question_id.unique():
    most_count=ds.loc[id].apply(lambda x:x.max())[0]
    most_count_dict[id]=most_count
ds2=ds.apply(lambda x:x-most_count_dict[x.name[0]],axis=1)
ds2=ds2[ds2.user_id==0]
ds2=ds2.reset_index(drop=False).loc[:,['question_id','user_answer']]
ds2.columns=['question_id','most_answer']
ds2.index=ds2['question_id']
ds2['most_answer']

question_id
q1        b
q10       d
q100      c
q1000     c
q10000    b
         ..
q9995     d
q9996     a
q9997     d
q9998     a
q9999     b
Name: most_answer, Length: 12215, dtype: object

This shows the most selected options (including not choose) for each question.

Note that if there are multiple options for a question to be selected most frequently, the table will also contain them.

Choices Distribution

ds = df['user_answer'].value_counts().reset_index(drop=False)
ds.columns = ['user_answer', 'percent']

ds['percent']=ds['percent']/len(df)
ds = ds.sort_values(by=['percent'])

fig = px.pie(
    ds,
    names = ds['user_answer'],
    values = 'percent',
    title = 'Percent of Choice'
)

fig.show("svg")

We use a pie chart to show the distribution of the proportions of a, b, c, d and not choose among the options selected by the 5000 students.

Sort By Time Stamp

import time
import datetime
df_time=df.copy()
columns=df.columns.to_list()
columns[1]='time'
df_time.columns=columns
df_time['time'] /= 1000
df_time['time']=pd.Series(map(datetime.datetime.fromtimestamp,df_time['time']))
df_time

	user_id	time	solving_id	question_id	user_answer	elapsed_time
0	u717875	2019-08-09 14:27:07.449	1	q4862	d	45000
1	u717875	2019-08-09 14:27:37.492	2	q6747	d	24000
2	u717875	2019-08-09 14:28:05.743	3	q326	c	25000
3	u717875	2019-08-09 14:28:36.475	4	q6168	a	27000
4	u717875	2019-08-09 14:28:57.148	5	q847	a	17000
...	...	...	...	...	...	...
574251	u177603	2018-06-30 23:16:48.931	15	q6984	b	44250
574252	u177603	2018-06-30 23:23:17.614	16	q7335	c	95750
574253	u177603	2018-06-30 23:23:18.181	16	q7336	a	95750
574254	u177603	2018-06-30 23:23:18.879	16	q7337	c	95750
574255	u177603	2018-06-30 23:23:19.425	16	q7338	b	95750

574256 rows × 6 columns

This table shows the result of converting unix timestamp to datetime format

question distribution by time

ds_time_question=df_time.loc[:,['time','question_id']]
ds_time_question=ds_time_question.sort_values(by=['time'])
ds_time_question

	time	question_id
503014	2017-05-11 05:17:10.922	q129
503015	2017-05-11 05:17:34.561	q8058
503016	2017-05-11 05:17:56.806	q8120
503017	2017-05-11 05:18:22.591	q157
503018	2017-05-11 05:18:43.085	q52
...	...	...
108215	2019-12-03 00:48:27.437	q776
108216	2019-12-03 00:59:38.437	q10847
108217	2019-12-03 00:59:38.437	q10844
108218	2019-12-03 00:59:38.437	q10845
108219	2019-12-03 00:59:38.437	q10846

574256 rows × 2 columns

This table shows the given questions in chronological order.And we can see that the earliest question q127 is on May 11, 2017, and the latest question q10846 is on December 3, 2019.

ds_time_question['year']=pd.Series(map(lambda x :x.year,ds_time_question['time']))
ds_time_question['month']=pd.Series(map(lambda x :x.month,ds_time_question['time']))
ds=ds_time_question.loc[:,['year','month']].value_counts()

years=ds_time_question['year'].unique()
years.sort()
fig=make_subplots(
    rows=2,
    cols=2,
    start_cell='top-left',
    subplot_titles=tuple(map(str,years))
)
traces=[
    go.Bar(
        x=ds[year].reset_index().sort_values(by=['month'],axis=0)['month'].to_list(),
        y=ds[year].reset_index().sort_values(by=['month'],axis=0)[0].to_list(),
        name='Year: '+str(year),
        text=[ds[year][month] for month in ds[year].reset_index().sort_values(by=['month'],axis=0)['month'].to_list()],
        textposition='auto'
    ) for year in years
]
for i in range(len(traces)):
    fig.append_trace(traces[i],(i//2)+1,(i%2)+1)

fig.update_layout(title_text='Bar of the distribution of the number of question solved in {} years'.format(len(traces)))
fig.show('svg')

 

1. These three figures show the distribution of the number of problems solved in each month of 2017, 2018, and 2019.

2. And the number of questions solved is gradually increasing.

3. And the number of questions solved in March, 4, May, and June is generally small.

user distribution by time

ds_time_user=df_time.loc[:,['user_id','time']]
ds_time_user=ds_time_user.sort_values(by=['time'])
ds_time_user

	user_id	time
503014	u21056	2017-05-11 05:17:10.922
503015	u21056	2017-05-11 05:17:34.561
503016	u21056	2017-05-11 05:17:56.806
503017	u21056	2017-05-11 05:18:22.591
503018	u21056	2017-05-11 05:18:43.085
...	...	...
108215	u9476	2019-12-03 00:48:27.437
108216	u9476	2019-12-03 00:59:38.437
108217	u9476	2019-12-03 00:59:38.437
108218	u9476	2019-12-03 00:59:38.437
108219	u9476	2019-12-03 00:59:38.437

574256 rows × 2 columns

This table shows the students who did the questions in order of time.And we can see that the first student who does the problem is u21056, and the last student who does the problem is u9476.

ds_time_user=df_time.loc[:,['user_id','time']]
ds_time_user=ds_time_user.sort_values(by=['time'])
ds_time_user['year']=pd.Series(map(lambda x :x.year,ds_time_user['time']))
ds_time_user['month']=pd.Series(map(lambda x :x.month,ds_time_user['time']))
ds_time_user.drop(['time'],axis=1,inplace=True)
ds=ds_time_user.groupby(['year','month']).nunique()

years=ds_time_user['year'].unique()
years.sort()
fig=make_subplots(
    rows=2,
    cols=2,
    start_cell='top-left',
    subplot_titles=tuple(map(str,years))
)
traces=[
    go.Bar(
        x=ds.loc[year].reset_index()['month'].to_list(),
        y=ds.loc[year].reset_index()['user_id'].to_list(),
        name='Year: '+str(year),
        text=[ds.loc[year].loc[month,'user_id'] for month in ds.loc[year].reset_index()['month'].to_list()],
        textposition='auto'
    ) for year in years
]
for i in range(len(traces)):
    fig.append_trace(traces[i],(i//2)+1,(i%2)+1)

fig.update_layout(title_text='Bar of the distribution of the number of active students in {} years'.format(len(traces)))
fig.show('svg')

1. These three graphs respectively show the number of students active on the system in each month of 2017, 2018, and 2019.

2. And we can see that the number of active students in 2019 is generally more than that in 2018, and there are more in 2018 than in 2017, indicating that the number of users of the system is gradually increasing.

3. Note that the number of students is not repeated here