Predicting article retweets and likes based on the title using Machine Learning

Flávio Henrique de Freitas

(Dated: September 13, 2018)

Abstract—Choosing a good title for an article is an important step in the writing pro-
cess. The more interesting the article title seems, the higher the chance a reader will in-
teract with the whole content. This project focus on predicting the number of retweets
and likes on Twitter from FreeCodeCamp’s articles based on its titles. This problem
is a classification task using Supervised Learning. With data from FreeCodeCamp on
Twitter and Medium, it was used machine learning methods including support vec-
tor machines (SVM), decision trees, gaussian naive Bayes (GaussianNB), k-nearest
neighbors, logistic regression and naive Bayes classifier for multinomial models (Multi-
nomialNB) to make the predictions. This study shows that the MultinomialNB model
performed better for retweets reaching an accuracy of 60.6% and logistic regression
reached 55.3% for likes.
Keywords— prediction, machine learning, social media, title, performance

I. DEFINITION the curiosity of the reader, proving enough information

to make the reader curious, but not enough to fully sat-
A. Project Overview isfy the curiosity. In this way, the user is forced to click
on the linked content to read the whole article.
Social networks websites have become an important Some other studies also investigate this subject using
communication tool and source of information. The deep learning on cross-domain sentiment analysis [7].
hours spent on average connected per day in the past
years is up to 6 hours [1] for adults and 9 for teenagers,
while 30% of this time is on social networks [2]. During a C. Problem Statement
normal navigation on such platforms, users are exposed
to several posts such as friends’ statuses, images, news When an author writes a text, it is expected that their
and more. With such amount of information and vari- words will influence and bring value to the readers. While
ety of content, the time for the user to decide to interact writing, the title is one of the important details that
with the content is very small. Gitte at al. [3] suggest needs to be taken into consideration, because this will
that we take around 50 milliseconds to make a good first normally be the first contact place of their work. Thus,
impression and this has proved to be very powerful in a to create a good first impression, to have more people
wide range of contexts. read the article and interact with it, choosing a good
Besides being a place for connecting with friends and title is very important.
sharing moments of the user’s life, a survey has shown Some of the most used platforms to spread ideas nowa-
that social networks are also used as a source of news days are Twitter [8] and Medium [9]. On the first one,
and information by 67% of the users [4]. Part of these articles are normally posted including external URLs and
posts are articles that can be read on an external website. the title, where users can access and demonstrate satis-
Typically such posts show the title of the article and faction with like or retweet (share) of the original post.
sometimes a small part of its content and an image. The second one shows the full text with tags to classify
Considering the offer of content and competition with the article and claps (similar to Twitter’s likes) to show
so many interesting posts, showing a proper title for the how much the users appreciate the content. A correlation
post affects the probability that a user will check the between these two networks can bring us more valuable
content. This measure has a strong impact on how many information.
readers an article will have and how much of the content The problem to be solved is a classification task us-
will be read. Furthermore, showing the user a content ing supervised learning: Predict the number of likes and
they prefer (to interact) increases the user satisfaction. retweets an article receives based on the title
It is thus important to accurately estimate the interaction
rate of articles based on its title.
D. Evaluation Metrics

B. Related Work At least one evaluation metric is necessary to quantify

the performance of the benchmarks and solution model.
In the literature is possible to find previous studies on For this project, it will be used the accuracy, which is
the area of classifying the article focused on click-baits ti- the number of correct predictions made as a ratio of all
tle detection [5] [6]. Click-bait headlines normally exploit predictions made.
 2

TABLE I. Sample of the data points

Number of correct predictions Title Retweet Like Claps
Accuracy =
Total number of predictions made ES9: JavaScript’s state of 15 48 618
the art in 2018
This metric only works well if there are a similar num- Here’s another way to think 10 30 2
ber of samples belonging to each class. For this reason, about state: How to
we will divide the range of retweets and likes count in a visually design state in
way that respects this distribution. JavaScript
How to understand 4 14 102
Gradient Descent, the most
popular ML algorithm
II. ANALYSIS

TABLE II. Complete description of the dataset fields

A. Data Exploration
Field Description
Title The content of the tweet, FreeCodeCamp
The data used to predict how titles will perform was
normally uses the title of the article from
gathered from the accounts of the non-profit organization Medium and sometimes the username of
FreeCodeCamp on Medium [10] and Twitter [11]. On the author from Twitter
both social platforms, it was possible to get public infor-
Retweet How many times that tweet was
mation about how the users interacted with the content, Count ”Retweeted” on Twitter
using as likes and retweets from Twitter, and claps from
Like Count How many times that tweet was liked on
Medium.
Twitter
Correlating the number of likes and retweets from
Medium How many times that article was marked
Twitter with a Medium article is an attempt to isolate Claps as favorite on Medium
the effect of the number of reached readers and the num-
Medium Which tags were used to classify the
ber of Medium claps. Because the more the article is
Categories article on Medium
shared in different platforms, the more readers it will
reach and the more Medium claps it will receive. Using Created at When the tweet was posted
only the Twitter statistic, it is expected that the articles URL The website of the article on Medium
reached initially almost the same number of readers (that
are the followers of the FreeCodeCamp account on Twit-
ter), and their performance and interactions are limited 1. Overall Statistic
to the characteristics of the tweet, for example, the title
of the article, that is exactly what we want to measure. We will analyze here the high-level statistics of the ar-
The FreeCodeCamp account was chosen, because the ticles. Try to understand how many times the articles
idea is to limit the scope of the subject of the articles and were on average retweeted, clapped or liked. Also, un-
predict better the response on a specif field. The same derstand the average length and number of words of the
title can perform well in one category (e.g. Technology), title.
but not necessarily in a different one (e.g. Culinary).
Also this account posts as the Tweet content the title of TABLE III. Overall Statistic
the original article and the URL on Medium. Like Retweet Claps Text Length
After getting the articles from FreeCodeCamp written
count 711.00 711.00 711.00 711.00
on Medium and shared on Twitter, there is a dataset of
711 data points. Table I shows some examples of such mean 49.29 16.44 285.26 80.62
correlation and table II explains the complete list of fields std 45.23 15.69 273.45 22.19
of the dataset. min 0.00 0.00 1.00 21.00
25% 20.00 7.00 6.00 65.00
50% 34.00 11.00 238.00 97.00
B. Exploratory Visualization 75% 63.50 20.00 471.50 97.00
max 298.00 125.00 997.00 146.00
This section will explore the data visualization of the
existing dataset and analyze the possible metrics that From this statistic is possible to understand the or-
will be used to understand the solution. We will identify der of magnitude of our dataset. Articles normally are
the relationship between each one of the features with retweeted and likes around tens of times and clapped
the overall performance of the article. hundreds of times. It is possible to check the maximum
 3

values from all the three variables, retweet and like hun-
dreds and clap thousand of times. From these numbers,
we can define what is expected from our articles and the
interaction with them. The length of the text goes from
21 to 146 characters, as expected, for a tweet content.

2. Histogram and Box plots

In this section we will check how the multiple features

are distributed.

FIG. 3. Relationship between the features

FIG. 1. Histogram
4. Title length that performed better

Here we analyze the relationship between the length of

the title with its performance. For this experiment, we
just considered the 25% top performers of each feature.

FIG. 2. Box plots

From these histograms, together with the overall

statistic and the box plots, we can notice that we have a
Gaussian distribution for the text length and the average FIG. 4. Title length performance
length is around 80 characters. Like, retweet and claps
are positive-skewed, i.e. they are concentrated on the left To avoid being biased by outliers, we removed for each
part of the graph, meaning that a small part of the arti- feature (likes, retweets and claps) analysis the data points
cles will over-perform about readers’ interaction and the that don’t fit the following formulas:
biggest part of them will generate less interaction.

Outlier < Q1 − 1.5 ∗ IQR

3. Scatter Matrix
Outlier > Q3 + 1.5 ∗ IQR
Here we try to find a relationship between the multiple
features that we gathered from Twitter and Medium. Where Q1 and Q3 are the first and third quartile and
We can notice for the image 3, we can notice a clear IQR is the Interquartile Range (IQR = Q3 - Q1 ).
relationship between the number of retweets and likes. We can notice from these graphics 4 that titles
They are directed connected, it means, the more retweets, longer than 50 and smaller than 120 characters (110 for
the more likes the article will receive and vice versa. Medium) perform better than others.
 4

After analyzing the title length and didn’t reach any

conclusion, we decided to investigate the number of words
in the title.

FIG. 7. Best Categories for like

FIG. 5. Performance of the Number of words in the title

From this second experience showed on image 5, we

reached the conclusion that the best number of words in
the title is from 9 to 17 words. To optimize the number of
retweets and likes something from 9 to 18, and for claps
from 7 to 17 words.
FIG. 8. Best Categories for claps
5. Categories that performed better

In this lexical analysis, we can notice that some words

Here we filtered the dataset and just analyzed the top
get much more attention on the FreeCodeCamp commu-
25% performers for each one of the features. We wanted
nity than others. If the intention is to make the articles
to have a clear overview of how the categories perform
reach further in numbers, talking about JavaScript, Re-
compared between them. The outliers were removed as
act or CSS will increase this change. Using the words
explained in section II B 4.
”learn” or ”guide” to describe will also make the proba-
bility higher.

C. Algorithms and Techniques

Classification is a common task of machine learning

(ML), which involves predicting a target variable taking
into consideration the previous data [12]. To reach such
classification, it is necessary to create a model with the
previous training data, and then use it to predict the
value of the test data [13]. This process is called super-
FIG. 6. Best Categories for retweet vised learning, since the data processing phase is guided
toward the class variable while building the model [14].
From this statistic, we notice that articles created Predicting the number of retweets and likes of an ar-
with the following categories can increase the number ticle can be treated as a classification problem, because
of retweets, likes and claps: ”Programming”, ”Tech”, the output will be discrete values (range of numbers). As
”Technology”, ”JavaScript” and ”Web Development”. input, the title of the articles with each word as a token
(t1, t2, t3, . . . tn), the title length and the number of
words in the title.
6. Words that performed better For this task, we evaluated the following algorithms:
Support vector machine (SVM): SVM contructs a hy-
We repeated the same strategy of limiting the 25% perplane (or a set) that can separate the points in the de-
performers for the words on the title of the article. We fined labels. The distance between the closest data points
wanted to understand if there are words that can boost and the hyperplan is named margin. An ideal separation
the interaction from the readers. The outliers were re- is defined by a hyperplan that has the largest distance
moved as explained in section II B 4. to the closest points of any class, so the challenge is to
 5

FIG. 9. Best Words for Retweet FIG. 11. Best Words for Claps

rithms can predict well with small set of data and when
there is a large number of features comparatively.
K-nearest neighbors (KNN): This algorithm takes in
consideration the k closest points (neighbors) around the
target and use them learn how to classify the desired
point.
This model was chosen, because its simple to imple-
ment, no assumption about the data is necessary and
the non-parametric nature of KNN gives an advantage in
certain settings where the data may be highly unusual.
FIG. 10. Best Words for Like Logistic regression: This model is named after the core
statistical function that it is based on, the logistic func-
tion. The Logistic regression estimates the parameters of
find the coefficients that maximize this margin. Only this function (coefficients), and as result it predicts the
these closest data points are relevant to identify (or to probability of presence of the characteristic of interest.
support the definition of) the hyperplane, and they are This model was chosen, because provides probabilities
named vectors. SVM performs linear classifications, but for outcomes and a convenient probability scores for ob-
also can efficiently perform non-linear, for this is nec- servations.
essary use a kernel trick, mapping their inputs into a
high-dimensional feature spaces. Naive Bayes classifier for multinomial models (Multi-
nomialNB): This model is similar to the Gaussian naive
This model was chosen, because it works well then big
Bayer, but the difference is that it was a multinomial
quantity of features and relatively small quantity of data
distribution of the dataset, instead of a gaussian one.
and to deal well with linear and non-linear datasets. And
due the fact we have more samples than number of fea- This model was chosen, because it works well for data
tures, it can generate a good prediction. which can easily be turned into counts, such as word
Decision trees: This model uses a decision tree to clas- counts in text. However, in practice, fractional counts
sifies the dataset into smaller subsets, and to define a con- such as TF-IDF may also work.
clusion about a target value. The tree consists of leaves, In the end, it was selected those with the best accu-
where the intermediate ones are the decision nodes and racy. To estimate it, it was used a 5-fold cross validation
the ones from the extremes are the final outcomes. that splits the dataset in 5 parts, 4 of training and 1 of
This model was chosen, because it can be easily in- testing. The implementation of this project was made
terpreted, visualized and explained. Also due the fact using Python, Numpy [15] and Scikit [16]. The full code
that this model implicitly perform variable screening or used on this project is available at: [17]
feature selection.
Gaussian naive Bayes (GaussianNB): This model is a
classification technique based on the Bayes’ Theorem. It
assumes the independence among the involved features. D. Benchmark
Nevertheless, this approach performers well even on data
that are dependent between them. This algorithm was This project run the same testing and training data
created by Bayes to prove the existence of God. It relies for multiple algorithms, the comparison between them
on the probability of an event, based on prior knowledge was used to evaluate the overall performance. The over-
of conditions that might be related to the event. all benchmark was made comparing our data with the
This model was chosen, because this family of algo- logistic regression results.
 6

III. METHODOLOGY 3. Claps: 0-50, 50-400, 400+

A. Data Preprocessing Where the range ”x-y”, means bigger than x and less
equal than y. The first range item of each feature also
1. Data cleaning
contains the zero on the range.

The first part of the data processing was to clean 3. Bag of words
the dataset. After downloading the tweets, we removed
the ones that didn’t have any URL (that points to the
To be possible to analyze the title in each data point,
Medium article) or title. Data points with values of likes,
we need to map each word into a number. This is nec-
claps or retweets that were not positive numbers or zero
essary because machine learning models normally don’t
were also excluded.
process raw text, but numerical values. To reach this, we
Words that were Twitter users were replaced by the
used a bag of words model [18]. In this model, it is taken
character ’@’ (that could be used on the statistics) and
into consideration the presence and often the frequency
words that were wrong non ASCII characters were also
of words, but the order or position is ignored.
removed.
For the calculation of the bag of words, we will use a
Some of the data points have the same URL, it means,
measure called Term Frequency, Inverse Document Fre-
that they shared more than once on the account of Twit-
quency (TF-IDF) [19]. The goal is to limit the impact of
ter. After analyzing each one of the duplicates, we no-
tokens (words) that occur very frequently.
ticed that there were two types of retweets: same URL
At this step, we processed the collection of documents
and same title; and same URL and a different title. We
and built a vocabulary with the known words. We
removed the ones of the first type. For the second type,
reached a vocabulary of 1356 words for retweets, 1399
we left, because the titles were completely rewritten and
words for likes and 1430 words for claps.
it can be considered as one different data point.
For the remaining data points, we removed the ones
that are considered outliers, as explained in II B 4. We B. Implementation
reached the numbers: retweet and likes have 711 items
(658 without outliers) each; claps has the same number
1. Training and Testing Data Split
with or without outliers, 711.

Before starting the training and the evaluation of the

2. Assigning classes to the dataset models, we split the dataset into test and training sets.
Retweets and likes have a total of 658 data points each,
For this project, we decided to classify the number of with 526 (80% approximately) as training and 132 as
retweets and likes in ranges. We wanted to make use of testing points. Claps has 711 data points, with 568 (80%
the properties of the Classification family of the Super- approximately) as training and 143 as testing points.
vised Learning algorithms.
To avoid the Class Imbalance Problem, we divided the
dataset into similar groups, as shown in image 12. 2. Training and Evaluating Models

For evaluating the model and perform the prediction,

it was chosen to start with the models: support vector
machines (SVM), decision trees, gaussian naive Bayes
(GaussianNB), k-nearest neighbors and logistic regres-
sion.
During the implementation tests, we noticed that these
models were not satisfying what we expected, so we also
trained and predicted using a naive Bayes classifier for
multinomial models (MultinomialNB).
The classification process followed the steps:

1. Load the data (title, likes, retweets and claps

FIG. 12. Range Distribution count)

The defined ranges for our features are: 2. Clean the title removing not desired words III A 1

1. Retweets: 0-10, 10-30, 30+ 3. Filter the outliers from the dataset II B 4
2. Likes: 0-25, 25-60, 60+ 4. Classify the features in ranges III A 2
 7

5. Divide the dataset in training and test Outliers: We made some tests to discover if we should
keep the outliers for the training or remove them. During
6. Create a bag of words using TF-IDF for the titles the tests, we discover if we keep the outliers, the accuracy
III A 3 was always worse.
Bag of words: To create the bag of Words, we had the
7. Train the model and calculate the accuracy
option of choosing the CountVectorizer or TfidfVector-
izer. During the simulation we got better results with
the last one, TfidfVectorizer.
3. Model Performance Metrics
Clean words: Another step during implementation was
to decide if we should keep the title in the original way or
We separated the dataset into learning and validation remove the undesired words. The tested to remove the
set. A validation set is important to reduce the risk of name of the Twitter users that appeared in some titles,
over-fitting of the chosen model. To avoid discarding some wrong characters that appeared on our dataset dur-
relevant data points, we used a cross-validation strategy. ing the crawling process. After checking the results, we
Cross-validation splits the training dataset in k folds, decided to clean up the data.
being k - 1 folders used to train the model and the last one Model’s parameters: For each model tested, we calcu-
to test it. This strategy will repeat multiple times and lated the accuracy for the default model (without any
the overall performance is the average of the computed parameter, just the default ones) and also we tried to
values. come with better parameters to evolve the accuracy. To
To estimate the model’s accuracy, we used a 5-fold test the combination of the new parameters, fine tune the
cross validation that split the dataset into 5 parts, 4 of model, we used grid search (GridSearchCV).
training and 1 of testing.

4. Challenges IV. RESULTS

During the implementation of the code available at: A. Model Evaluation and Validation
[17], the biggest challenge was to find the best accuracy,
because in the initial attempts the final accuracy was The tables IV, V and VI describe the accuracy values
close or even worse than the benchmark. To reach an we reached with the proposed model. The final accuracy
acceptable value, it was necessary to reiterate over the for each of the features are: likes is 55.3%, retweets is
same code several times and try to understand what were 60.6% and claps is 49%.
the factors that increase or decrease the this metric.
Since our input data was limited, because we are using
just the titles, we had to bring alternatives and hypothe- TABLE IV. Accuracy Likes (%)
ses that we could obtain from this value. During the ID Model Name Default Tuned
III C, we discuss several approaches that were used to 0 Benchmark 53.15 –
get a better accuracy.
1 LogisticRegression 55.30 45.45
2 GaussianNB 46.21 46.21
C. Refinement 3 DecisionTreeClassifier 43.18 45.45
4 SVC 51.51 51.51
During the models’ implementation a lot of steps were 5 KNeighborsClassifier 44.70 46.21
tested and some of them needed to be modified to reach 6 MultinomialNB 40.91 45.45
better performance. For choosing a better parameter, we 7 GradientBoostingClassifier 47.73 48.48
interacted over the options and decided on the one that
optimized the accuracy.
Ranges of likes and retweets: We tried ranges with The parameters that we obtained by the grid search of
different number of elements and also different values for the models and features are explained in the tables VII,
the ranges. Some of them were underperforming, while VIII and IX.
others reached close values. The chosen one divided the After the steps presented on the previous sections, we
ranges with a similar number of data points and also noticed that this model is robust for outliers. Even if
offers a good overview of the feature analyzed. part of the design process was to eliminate them during
New models: We increased the number of models to the training step, we can see on the image 13 the lit-
be tested in three. The classifiers previous chosen were tle variation of the final result when ignoring or not the
not reaching the desired accuracy, we decided to add new outliers.
models to try to make better predictions. The new mod- This model can be considered to reach a reasonable
els have a worse performance than the first ones. accuracy for the proposed goal. We can make further
 8

TABLE V. Accuracy Retweets (%) TABLE VII. Tuned Parameters Likes

ID Model Name Default Tuned ID Parameters

0 Benchmark 57.34 – 1 C=1, class weight=None, dual=False,

fit intercept=True, intercept scaling=1,
1 LogisticRegression 49.24 53.79
max iter=100, multi class=’ovr’, n jobs=1,
2 GaussianNB 59.09 49.24 penalty=’l2’, random state=None, solver=’liblinear’,
3 DecisionTreeClassifier 56.06 54.54 tol=0.0001, verbose=0, warm start=False
4 SVC 55.30 57.58 2 priors=None
5 KNeighborsClassifier 57.58 55.30 3 class weight=None, criterion=’gini’,
6 MultinomialNB 47.72 60.61 max depth=None, max features=None,
max leaf nodes=None, min impurity decrease=0.0,
7 GradientBoostingClassifier 56.82 55.30 min impurity split=None, min samples leaf=1,
min samples split=2, min weight fraction leaf=0.0,
presort=False, random state=None, splitter=’best’
TABLE VI. Accuracy Claps (%) 4 C=1, cache size=200, class weight=None,
ID Model Name Default Tuned coef0=0.0, decision function shape=’ovr’, degree=1,
gamma=’auto’, kernel=’linear’, max iter=-1,
0 Benchmark 42.65 – probability=False, random state=None,
1 LogisticRegression 46.85 48.95 shrinking=True, tol=0.001, verbose=False
2 GaussianNB 41.26 41.26 5 algorithm=’auto’, leaf size=50, metric=’minkowski’,
3 DecisionTreeClassifier 37.06 44.06 metric params=None, n jobs=1, n neighbors=20,
p=2, weights=’uniform’
4 SVC 49.65 49.65
6 alpha=0.5, class prior=None, fit prior=True
5 KNeighborsClassifier 39.16 41.95
7 criterion=’friedman mse’, init=None,
6 MultinomialNB 42.66 42.66
learning rate=0.1, loss=’deviance’, max depth=10,
7 GradientBoostingClassifier 44.76 37.76 max features=None, max leaf nodes=None,
min impurity decrease=0.0,
min impurity split=None, min samples leaf=2,
modifications, as explained in section V C, to reach a min samples split=0.5,
better result, but for now it is quite good. min weight fraction leaf=0.0, n estimators=100,
presort=’auto’, random state=None,
subsample=1.0, verbose=0, warm start=False
B. Justification

The benchmark used on this project was Logistic Re-

gression model, with ID 0 and named Benchmark on the
tables IV, V and VI.
In all the scenarios our models predicted better than
the benchmark, for this reason, we can assume the deci-
sions made on the step III C led us to a better model.

V. CONCLUSION

A. Free-Form Visualization
FIG. 13. Evolution Accuracy
The image 13 shows the evolution of the development
of this project. We started with the benchmark, from
there we started adding and testing features and treating 3. t2: t1 + Cleaned the words
the dataset.
During the evolution of the metrics, some variables
deprecated the accuracy value, but in further steps, it 4. t3: t2 + Added TF-IDF
made it grow. The evolution happened in the following
steps:
5. t4: t3 + Used stopwords
1. Benchmark
2. t1: Removed outliers 6. Winner: t4 + Parameters from the model tuned
 9

TABLE VIII. Tuned Parameters Retweets TABLE IX. Tuned Parameters Claps
ID Parameters ID Parameters
1 C=10, class weight=None, dual=False, 1 C=2, class weight=None, dual=False,
fit intercept=False, intercept scaling=1, fit intercept=True, intercept scaling=1,
max iter=100, multi class=’ovr’, n jobs=1, max iter=10, multi class=’ovr’, n jobs=1,
penalty=’l2’, random state=None, solver=’lbfgs’, penalty=’l2’, random state=None, solver=’liblinear’,
tol=0.0001, verbose=0, warm start=False tol=0.0001, verbose=0, warm start=False
2 priors=None 2 priors=None
3 class weight=’balanced’, criterion=’gini’, 3 class weight=None, criterion=’gini’, max depth=10,
max depth=20, max features=None, max features=None, max leaf nodes=None,
max leaf nodes=None, min impurity decrease=0.0, min impurity decrease=0.0,
min impurity split=None, min samples leaf=1, min impurity split=None, min samples leaf=1,
min samples split=2, min weight fraction leaf=0.0, min samples split=2, min weight fraction leaf=0.0,
presort=False, random state=None, splitter=’best’ presort=False, random state=None,
4 C=1, cache size=200, class weight=None, splitter=’random’
coef0=0.0, decision function shape=’ovr’, degree=1, 4 C=1, cache size=200, class weight=None,
gamma=’auto’, kernel=’linear’, max iter=-1, coef0=0.0, decision function shape=’ovr’, degree=1,
probability=False, random state=None, gamma=’auto’, kernel=’linear’, max iter=-1,
shrinking=True, tol=0.001, verbose=False probability=False, random state=None,
5 algorithm=’auto’, leaf size=10, metric=’minkowski’, shrinking=True, tol=0.001, verbose=False
metric params=None, n jobs=1, n neighbors=10, 5 algorithm=’auto’, leaf size=30, metric=’minkowski’,
p=2, weights=’uniform’ metric params=None, n jobs=1, n neighbors=20,
6 alpha=1, class prior=None, fit prior=True p=2, weights=’uniform’
7 criterion=’friedman mse’, init=None, 6 alpha=0.5, class prior=None, fit prior=True
learning rate=0.5, loss=’deviance’, max depth=10, 7 criterion=’friedman mse’, init=None,
max features=None, max leaf nodes=None, learning rate=0.5, loss=’deviance’, max depth=3,
min impurity decrease=0.0, max features=None, max leaf nodes=None,
min impurity split=None, min samples leaf=2, min impurity decrease=0.0,
min samples split=1.0, min impurity split=None, min samples leaf=1,
min weight fraction leaf=0.0, n estimators=50, min samples split=0.5,
presort=’auto’, random state=None, min weight fraction leaf=0.0, n estimators=100,
subsample=1.0, verbose=0, warm start=False presort=’auto’, random state=0, subsample=1.0,
verbose=0, warm start=False

B. Reflection
3. Clean the dataset and each title before training the
In this project, we developed classifiers to under- model (remove Twiter users and invalid characters)
stand how many times an article will receive interac- 4. Grid search to search for the best model parameters
tion like retweets and likes (both on Twitter) and claps
(on Medium). We also presented a list of words that 5. Remove outliers before processing the data
have a high change to impact positively with the read-
ers, when used on the title or the article. We classified 6. Test the dataset to discover a good relation between
and extracted information about the Categories used on train and test data points
Medium that are commonly presented on our top per-
7. Test and split the number of like, retweet and claps
formers. The number of words and length of the title
in ranges
were also discussed and presented an optimal number to
increase the success numbers. 8. Use stopwords to remove common terms of the lan-
Besides the mathematical analysis used to extract im- guage
portant characteristics of the dataset, we also developed
and trained models to predict the how an article would Following these steps listed here, this methodology and
perform. To achieve this machine learning project, some framework can be used to classify any kind of article and
features and characteristics were used: subjects that are created on Medium and shared on Twit-
ter. This solution is not limited by the context either the
1. Bag of words to tokenize the words of the title subject of the articles and can be easily reproduced to
other datasets.
2. Term Frequency, Inverse Document Frequency The hard part of the project was to reach a higher
(TF-IDF) to translate the frequency of words in accuracy than the one found with simple models, it was
the dataset necessary multiple reiterations and several modifications
 10

on the initial assumption. Reaching the 61%, 55% and [6] Lex, Elisabeth, Andreas Juffinger, and Michael Gran-
49% is not the ideal solution, but it can clearly lead to itzer. ”Objectivity Classication in Online Media.” Re-
the creation of a good title. searchGate. Proceedings of the 21st ACM Conference on
Hypertext and Hypermedia, 13 June 2010.
[7] Glorot, Xavier, Antoine Bordes, and Yoshua Bengio.
”Domain Adaptation for Large-Scale Sentiment Classi-
C. Improvement fication: A Deep Learning Approach.” WUSTL. Proceed-
ings of the 28 Th International Conference on Machine
Learning, 2011.
For future work we can think about some additional
[8] Twitter. Accessed 13 Aug. 2018. Available at:
improvements: Adding more features to the original https://www.twitter.com.
dataset making possible to relate more information to [9] Medium. Accessed 13 Aug. 2018. Available at:
the success of the article. For example, we can correlate https://www.medium.com.
the words of the title, with trendy words of the month; [10] freecodecamp. Accessed 13 Aug. 2018. Available at:
Bring more data points to train our model, would also https://medium.freecodecamp.org/ .
increase the accuracy of the solution; and try to use the [11] freecodecamp.org. Accessed 13 Aug. 2018. Available at:
position of the word on the title to classify its importance. https://twitter.com/freecodecamp.
[12] N. Abdelhamid, A. Ayesh, F. Thabtah, S. Ahmadi, W.
Hadi. MAC: A multiclass associative classification algo-
rithm J. Info. Know. Mgmt. (JIKM), 11 (2) (2012), pp.
REFERENCES 125001-1-1250011-10 WorldScinet.
[13] I.H. Witten, E. Frank, M.A. Hall. Data Mining: Prac-
[1] eMarketer Report. (2017). US Time Spent tical Machine Learning Tools and Techniques. Morgan
with Media: eMarketer’s Updated Estimates Kaufmann, Burlington, MA (2011).
for 2017. Accessed 9 Aug. 2018. Available at: [14] Thabtah, S. Hammoud, H. Abdeljaber, Parallel associa-
https://www.emarketer.com/Report/US-Time- tive classification data mining frameworks based mapre-
Spent-with-Media-eMarketers-Updated-Estimates- duce, To Appear in Journal of Parallel Processing Letter,
2017/2002142 . March 2015, World Scientific, 2015.
[2] Common Sense Media. (2015). The Common [15] NumPy. Accessed 14 Aug. 2018. Available at:
Sense Census: Media Use by Tweens and http://www.numpy.org/ .
Teens. Accessed 9 Aug. 2018. Available at: [16] scikit-learn. Accessed 14 Aug. 2018. Available at:
https://www.commonsensemedia.org/research/the- http://scikit-learn.org/stable/ .
common-sense-census-media-use-by-tweens-and-teens. [17] de Freitas, Flavio H. Jupyter notebook - Im-
[3] Lindgaard, Gitte Fernandes, Gary Dudek, Cathy M. plementation code: Predicting article retweets
Brown, Judith. (2006). Attention web designers: You and likes based on the title using Machine
have 50 milliseconds to make a good first impression! Be- Learning. Accessed 11 Sep. 2018. Available at:
haviour and Information Technology, 25(2), 115-126. Be- https://github.com/flaviohenriquecbc/machine-
haviour IT. 25. 115-126. 10.1080/01449290500330448 . learning-capstone-project/blob/master/title-success-
[4] Shearer, E., Gottfried, J. (2017). News use across so- prediction.ipynb.
cial media platforms 2017. Accessed 9 Aug. 2018. Avail- [18] Nahm, U. Y.; Mooney, R. J. Text mining with informa-
able at: http://www.journalism.org/2017/09/07/news- tion extraction. In: AAAI 2002 Spring Symposium on
use-across-social-media-platforms-2017/ . Mining Answers from Texts and Knowledge. Bases. [S.l.:
[5] Chen, Yimin, Niall J. Conroy, and Victoria L. Ru- s.n.], 2002. v. 1.
bin. ”Misleading Online Content: Recognizing Clickbait [19] Baeza-Yates, R.; Ribeiro-Neto, B. et al. Modern infor-
as ”False News”” ResearchGate. ACM WMDD, 9 Nov. mation retrieval. [S.l.]: ACM press New York, 1999. v.
2015. 463.