Examination:
Programming for Data Science (ID2214)
Course code: ID2214
Course name: Programming for data science
Literature and tools:
The following rules apply to both parts of the examination:
Literature, other documents, including lecture slides, notes, etc. are not allowed.
Computers, tablets, phones, etc. are not allowed for searching for answers or
communicating with anyone except for the examiner. A text editor/word pro-
cessor (on a computer, tablet or phone) may be used for writing answers to the
questions, and on part II, any tool, such as an editor, Jupyter notebook or IDE
may be used for developing the programs.
Date and time: April 14, 2020, 08:00-12:00
Examiner: Henrik Bostro¨m
Requirements to pass: 5 points on part I and 10 points on part II.
On part I, keep the text short and to the point.
On part II, only the Python standard library, the NumPy and pandas libraries
may be assumed, in addition to functions explicitly stated in the tasks.
The answers (including blank ones) should be numbered and ordered.
Unreadable answers will be ignored.
Good luck!
Part I (Theory, 10 points)
1a. Methodology, 2 points
Assume that we want to develop a model and estimate its performance on
independent data. We have therefore have decided to randomly split an available
dataset into a training and test set, using the former to train the model and
the latter to estimate its performance. However, since the learning algorithm
that we would like to use cannot directly deal with missing values, we have
decided to employ some imputation technique prior to applying the algorithm.
The question is now whether there may be any potential risk in applying the
imputation technique using the whole dataset, before randomly splitting the
data into training and testing. Explain your answer.
1
1b. Data preparation, 2 points
Assume that we want to discretize numerical features prior to applying the
decision-tree learning algorithm. Will it have any effect on the resulting tree if
we employ equal-width or equal-sized binning? Explain your reasoning.
1c. Performance metrics, 2 points
Assume that we have a binary classification model that, given a test instance,
outputs an estimate of the probability for the positive class. By choosing some
other threshold than 0.5 for whether to assign a positive or negative label to the
test instances, the predictive performance may be affected. Should we increase
or decrease the threshold to increase a) precision and b) recall, of the positive
class? Explain your answer.
1d. Combining models, 2 points
The predictive performance of an ensemble of classifiers, for which the predic-
tions are formed by averaging predictions of the individual members, is depen-
dent on the diversity of the members. Describe how an ensemble of na¨ıve Bayes
classifiers would be trained, if similar techniques that are used to form random
forests would be employed.
1e. Association rules, 2 points
Assume that we have generated a set of association rules with a specified support
and confidence, from a dataset with a set of binary features and binary class
labels, encoded as itemsets. Assume that we have selected a subset of the rules,
for which the heads (consequents) contain only a class label. If we want to use
this subset of rules to classify a novel test instance, i.e., to assign one of the two
class labels, what are the potential problems we may encounter? Explain your
reasoning.
Part II (Programming, 20 points)
2a. Data preparation, 10 points
Your task is to define the following Python function that aggregates multiple
rows with the same identifier into a single row:
aggregate(df)
which given a pandas dataframe df, where the columns correspond to numerical
features, except for a column named CLASS, which contains (categorical) class
values, and a column named ID, which contains identifiers for the instances. In
case one or more instances share the same identifier, these should in the new
dataframe be represented by a single row, where the identifier should appear in
the ID column, and the value for each numerical feature of the instance should
be the mean of the values for instances sharing the identifier, and the value for
the class label of the instance should be the mode of the values for instances
sharing the identifier (in case the mode is not unique, one of the mode values
2
may be chosen arbitrarily).
For example, given a dataframe df:
ID V1 V2 CLASS
0 1 1 1 A
1 1 2 0 A
2 2 3 1 B
3 2 4 0 C
4 3 5 1 C
then aggregate(df) should return the following dataframe:
ID V1 V2 CLASS
0 1 1.5 0.5 A
1 2 3.5 0.5 B
2 3 5.0 1.0 C
Hint: You may obtain unique values from a pandas data series values by
values.unique(), the mean by values.mean() and the first mode value by
values.mode()[0].
2b. Combining models, 10 points
Your task is to define the following Python function:
stacking(level0_df,level1_df,base_learners,learner)
which given two pandas dataframes level0_df and level1_df, with the same
set of columns, which correspond to features, except for a column named CLASS,
which contains the class values, and where the rows correspond to instances, a
list of learning algorithms base_learners (see below), and a learning algorithm
learner (see below), will return a list of base models and a stacking model,
where each base model is trained on level0_df using the corresponding base
learning algorithm, and the stacking model is trained using the class labels of
level1_df and the predictions of the base models on level1_df as features.
Each learning algorithm alg, i.e., as specified by learner and the elements
of base_learners above, can be used to generate a model from a dataframe by:
model = alg.fit(df)
Each resulting model can be used to make predictions (produce a list of class
labels) for a dataframe (in which the CLASS column is ignored) by:
predictions = model.predict(df)
For example, given two dataframes df0 and df1, with the same columns, and
the learning algorithms alg1,alg2,alg3,alg4, then:
stacking(df0,df1,[alg1,alg2,alg3],alg4)
should return a list of three (base) models trained using alg1,alg2,alg3
on df0, and a (stacking) model trained using alg4 on a dataframe with four
columns; one column for each base model with the predictions for df1, and one
column containing the class labels of df1.
3

欢迎咨询51作业君