Recommender System for News Articles using Supervised Learning

Abstract

In the last decade we have observed a mass increase of information, in particular information that is shared through smartphones. Consequently, the amount of information that is available does not allow the average user to be aware of all his options. In this context, recommender systems use a number of techniques to help a user find the desired product. Hence, nowadays recommender systems play an important role. Recommender Systems' aim to identify products that best fits user preferences. These techniques are advantageous to both users and vendors, as it enables the user to rapidly find what he needs and the vendors to promote their products and sales. As the industry became aware of the gains that could be accomplished by using these algorithms, also a very interesting problem for many researchers, recommender systems became a very active area since the mid 90's. Having in mind that this is an ongoing problem the present thesis intends to observe the value of using a recommender algorithm to find users likes by observing her domain preferences. In a balanced probabilistic method, this thesis will show how news topics can be used to recommend news articles. In this thesis, we used different machine learning methods to determine the user ratings for an article. To tackle this problem, supervised learning methods such as linear regression, Naive Bayes and logistic regression are used. All the aforementioned models have a different nature which has an impact on the solution of the given problem. Furthermore, number of experiments are presented and discussed to identify the feature set that fits best to the problem.

Full text

Recommender System for News
Articles using Supervised Learning
Akshay Kumar Chaturvedi
(MIIS Master Thesis)
Supervised by:
Dra. Filipa Peleja
(Vodafone Portugal)
Dra. Ana Freire
(Department of Information and Communications Technologies - Universitat Pompeu Fabra)

Abstract
In the last decade we have observed a mass increase of information, in particular information
that is shared through smartphones. Consequently, the amount of information that is available
does not allow the average user to be aware of all his options. In this context, recommender
systems use a number of techniques to help a user find the desired product. Hence, nowadays
recommender systems play an important role.
Recommender Systems' aim to identify products that best fits user preferences. These
techniques are advantageous to both users and vendors, as it enables the user to rapidly find
what he needs and the vendors to promote their products and sales. As the industry became
aware of the gains that could be accomplished by using these algorithms, also a very
interesting problem for many researchers, recommender systems became a very active area
since the mid 90's.
Having in mind that this is an ongoing problem the present thesis intends to observe the value
of using a recommender algorithm to find users likes by observing her domain preferences. In
a balanced probabilistic method, this thesis will show how news topics can be used to
recommend news articles.
In this thesis, we used different machine learning methods to determine the user ratings for an
article. To tackle this problem, supervised learning methods such as linear regression, Naive
Bayes and logistic regression are used. All the aforementioned models have a different nature
which has an impact on the solution of the given problem.
Furthermore, number of experiments are presented and discussed to identify the feature set
that fits best to the problem.

List of Figures
1.1. Content Based Filtering.................................................................................7
1.2. Collaborative Filtering...................................................................................8
1.3. Collaborative Vs Content Based Filtering.....................................................8
2.1. The New York Times' recommendation section..........................................10
4.1. Linear Regression........................................................................................14
4.2. Logistic Function.........................................................................................15
4.3. Bayes Rule...................................................................................................16
6.1. Linear Regression on Pre-Processed Dataset...............................................27
6.2. Logistic Regression on Pre-Processed Dataset............................................27
6.3. Naive Bayes on Pre-Processed Dataset.......................................................28
6.4. Correlation of user ratings with different features......................................28
6.5. Linear Regression results for individual users.............................................29
6.6. Logistic Regression results for individual users..........................................30
6.7. Naive Bayes results for individual users......................................................30
6.8. Performance of Extracted Features.............................................................32
6.9. Impact of Number of classes......................................................................32

List of Tables
1.1. Methods of Hybridization.............................................................................9
6.1. Description of explicit feedbacks................................................................22
6.2. Description of implicit feedbacks................................................................23
6.3. Statistics of explicit feedbacks.....................................................................23
6.4. Statistics of implicit feedbacks....................................................................24
6.5. Feature names and their respective aliases..................................................26

Table of Contents
1. Introduction ....................................................................................................... 6
2. Problem definition and thesis objective .......................................................... 10
3. Problem Formalization .................................................................................... 12
4. Background ..................................................................................................... 13
4.1. Linear Regression.....................................................................................13
4.2. Logistic Regression...................................................................................15
4.3. Naive Bayes .............................................................................................16
5. Related Work..................................................................................................17
6. Experiments....................................................................................................21
6.1. Overview...................................................................................................21
6.2. Dataset......................................................................................................22
6.3. Data Pre-processing..................................................................................24
6.4. Feature Extraction.....................................................................................25
6.5. Results.......................................................................................................26
7. Conclusions and Future work.........................................................................33

1. Introduction
This is the age of information and consumption of information has gained pace, particularly
in recent years. People in today's world are overwhelmed with data and this often creates the
problem of choosing something from a large set of options. Recommender systems identify
which products should be presented to the user, in which the user will have time to analyse
and select the desired product [Ricci et al., 2011].
Now with the advent of e-commerce websites like Amazon, it became more obvious the
important role that recommender systems play. The importance of recommender systems also
came to light in 2006. In 2006, Netflix, a global provider of streaming movies and TV series,
announced an open competition to predict user ratings for films, based on previous ratings
without any information about the users or the films. Prizes were based on improving
Netflix's own algorithm. The team that finally won the competition was able to achieve
around 10% improvement on Netflix's algorithm [Ricci et al., 2011], they were awarded 1
million US$ .
A recommender system is traditionally composed of:
 Users : People in the system who have preferences for items and people who can be
source of data as well are called Users [Ricci et al., 2011]. Each user may have a set
of user attributes, if we are using user demographic (age, gender etc.) then
demographic is a user attribute. For each user, a model can be inferred. For example,
their genre of movie or the type of books they (the users) like to read.
 Items : Products the system is choosing to recommend are known as Items [Ricci et
al., 2011]. For each item we may have a set of item attributes or properties. For
example, an actor in a movie, author of a news article or colour of an appliance.
 Preferences : These represent users' likes and dislikes [Ricci et al., 2011]. User meets
item in the preferences space, for example, a user can rate a movie 5 on a scale of 5
stars.
The components described above are used in different algorithms in various ways. Broadly
these algorithms can be classified in the following types:
 Non-personalized systems : These involve summary statistics and in some cases
product associations [Poriya et al., 2014] by using external data from the community,
like, a product that is the best seller or most popular or trending hot. It may also
provide summary of community ratings, for example, how much a population likes a
restaurant or summary of community ratings which turns into a list like, which is the
best hotel in town.

 Content based filtering : Here, users rate items and from that a model of user
preferences against the item attributes is built [Ricci et al., 2011]. An example could
be in the domain of movies. Suppose someone likes science fiction, fantasy and action
movies, and doesn't like romantic movies. Overtime the algorithm can accumulate this
and Figure out that the user has positive scores on genres like science fiction, fantasy
and action, and lower scores for romance. The algorithm might also find out that there
were some actors that user likes or dislikes. For example, the user can be a fan of
movies with the actor Bruce Willis, and not a fan of Ben Stiller. Content based
filtering uses this information to map user ratings against the attributes of the products
- in this example movies. Our approach fits in this type of algorithm as we are trying
to build a model on user feedback of news articles. Figure 1.1 provides a good
overview of Content Based Filtering.
Figure 1.1 : Content Based Filtering
 Collaborative filtering : In collaborative filtering, user ratings of other people are
used rather than attribute data to predict and recommend [Ricci et al., 2011].
Collaborative filtering builds on the idea of a user model that is a set of ratings and an
item model that is a set of ratings. Combining the two models, we get a sparse matrix
of ratings, some of its cells are filled and most are not. So, here there are two main
tasks, one is to fill the empty cells or predict a rating and second is to choose a filled
cell or recommend an item. A good overview of Collaborative Filtering is presented
in Figure 1.2.

Figure 1.2 : Collaborative filtering
The main difference between Content Based Filtering and Collaborative Filtering is
that Collaborative Filtering works on preferences of other users (users with similar
preferences for some items) to recommend new items whereas Content Based
Filtering is not at all concerned with preferences of the other users. This point is
appropriately illustrated in Figure 1.3.
Figure 1.3 : Collaborative Vs Content Based Filtering

 Hybrid methods : In these methods, a combination of two or more recommendation
algorithms are used to take or maximize advantage of some techniques and avoid or
minimize the drawbacks of another [Burke, 2002]. Various ways of combining
different algorithms are shown in Table 1.1 (for more details please see [Burke,
2002]).
Table 1.1 : Methods of Hybridization
Hybridization
method
Description
Weighted
The scores (or votes) of several recommendation
techniques are combined together to produce a single
recommendation.
Switching
The system switches between recommendation techniques
depending on the current situation.
Mixed
Recommendations from several different recommenders
are presented at the same time
Feature combination
Features from different recommendation data sources are
thrown together into a single recommendation algorithm.
Cascade
One recommender refines the recommendations given by
another.
Feature
augmentation
Output from one technique is used as an input feature to
another.
Meta-level
The model learned by one recommender is used as input to
another.

2. Problem definition and thesis objective
People tend to have preferred sources for printed news, like a newspaper or a magazine
(Figure 2.1 shows the news recommendation section of The New York Times). However, in
online world, this reality changes drastically, a user is flooded with information from
different sources in which it is not uncommon for a user to change between news portals or
read news from portals that merge news articles from different sources (Yahoo! News).
Having such a huge amount of information, it becomes difficult to select the news article that
a user will like. Consequently, users stop the consumption of news or lowers it down.
Figure 2.1 : The New York Times' recommendation section.
To overcome the problem of information overloading (difficulty in making a decision caused
by the presence of too much information), recommender systems can be used. These types of
systems use various types of feedbacks from the users to narrow down the users' options to a
small number. Rating an article is the most common feedback that is used, but ratings alone
are not always sufficient to predict the articles a user might like. So, other types of feedbacks
are used to improve the recommendation. One can try various implicit feedbacks as user is
not asked to provide them. But when it comes to explicit feedback, there are a few things one
has to take into account, firstly, the total number of explicit feedbacks used by the
recommender system should be minimum, as users usually don't have much time or don't like
to provide feedbacks. Secondly, feedback asked to the user should be easy for an average
user to understand so that he can provide the feedback easily and correctly.

The goal of the present thesis is to investigate the impact of features extracted (explained in
section 6.1) from feedback 'classes' (category a news article belongs to, for example, sports,
politics, etc.) on recommendations and also on other features. To achieve this goal, we used
the extracted features in machine learning models with different feedbacks in the dataset. We
have further created new datasets by selecting the articles belonging to the most frequent
classes and removing the articles belonging to the infrequent ones' and then performed
experiments on them to see the impact of number of classes on the performance of the
system. At last, we have done analysis of the results obtained from using feature
combinations in different models to figure out the right model and features required to make
reasonable recommendations.

3. Problem Formalization
In this section, we will formalize our problem and also give a brief overview how we
achieved our objective. Our problem of recommending news articles can be formalized as
follows: given feature(s) F, learn a model which can estimate user ratings R, such that mean
squared error (MSE) between original (R) and estimated value (r) is minimal, where
 F represents the features, it can be visualized to be a matrix of size M by N, where N
is the number of different types of feedbacks taken from user and M is the number of
articles. It contains every type of feedback taken from user except user ratings (R),
which is to be estimated. We can use all the features or select a subset.
 R is the user ratings given by user for each article. Domain of user ratings can be
binary or from a range of integers 1 to 5 or 1 to 10, in our dataset it is 1 to 5.
 r is the estimated user rating (we find 'r' using supervised learning algorithms in the
present thesis) for each article. It takes the same values as R.
 MSE or mean squared error is defined as 1
𝑀 (𝑅 − 𝑟)2, M is the number of articles.
To achieve the above objective we experimented with three Supervised Learning algorithms,
namely Linear Regression, Logistic Regression and Naive Bayes. The three algorithms are
described in sections 4.1, 4.2 and 4.3. Experiments were conducted not only with the features
available in the data set, but also with two extracted features which are described in section
6.1. After obtaining results (MSE) from our three algorithms mentioned before, these results
are compared with the result of a baseline model which is also described in section 6.1.

4. Background
Machine Learning can be described as inferring or learning models from data itself about the
process that generated the data. A more formal definition was given by Tom M. Mitchell in
[Mitchell, 1997] : "A computer program is said to learn from experience E with respect to
some class of tasks T and performance measure P if its performance at tasks in T, as
measured by P, improves with experience E". Machine Learning can be broadly classified
into three types:
Supervised Learning : When the given data contains the output explicitly and the pairs of
input and output data points are used to predict the unknown output of inputs, it is called
supervised learning [James et al., 2014] . An example could be a movie data set with different
features of movie like director, lead actor or genre (as inputs) and also user rating (as output).
Unsupervised Learning : Instead of having input with correct output, data only contains
input. Here, the objective is to cluster or group the similar inputs together [James et al., 2014]
. An example could be clustering documents in different genres given the most frequent
words in each document.
Reinforcement Learning : Here, situation is a bit similar to unsupervised learning in the
sense that it does not use input with correct output, instead it uses a reward for every action
[Mitchell, 1997] . The rewards can be both positive and negative depending on the action
taken given the input. So, here the objective is to maximize the total reward which is the sum
of all rewards which the system got after every action. An example could be playing a game
where for every right move a positive reward is given and for every wrong move a negative
reward is given, thereby the objective is to figure out the set of moves or actions which
maximizes the total reward.
Most of the machine learning algorithms fall into one of the three categories described above.
We will describe below some of the algorithms but will concentrate only on supervised
learning methods as it is the relevant type in this thesis.
4.1. Linear Regression
Linear Regression has two main objectives : the first one is to establish if there is a
relationship between two variables. This relationship can be positive i.e. both variables
increase and decrease together, or negative i.e. when one variable increases, the other
decreases and vice-versa. The second objective is to forecast output values for new inputs.
The output of this algorithm is a linear equation,
y = β0 + β1*x + έ (1)

where y is the dependent variable or the output the model will predict and x is the
independent variable or the input. β0 is the constant or intercept of the equation, it doesn't
have an intuitive interpretation. β1 is the slope and is the relationship between variables, it is
the amount by which y changes when x changes. In most cases, data is not completely linear
i.e. some data points may fit the line exactly and some may not, in that case, we get error
between actual output and predicted output, έ is that error (This error is shown in Figure 4.1).
Gradient descent algorithm is used to find optimal values of β0 and β1 which minimizes error
έ, thereby finding the best line to fit the data [James et al., 2014] . In case, data is not at all
linear then linear regression won't work.
Figure 4.1 : Linear Regression
Linear Regression can be of two types, simple linear regression and multiple linear
regression. In simple linear regression, there is only one independent variable (x). In multiple
linear regression, there are more than one or multiple independent variables (x1,x2,x3....). In
multiple linear regression, a subset of variables can be selected rather than using all of them.
Two strategies are used for this selection, one is to check the relationship of different
independent variables with dependent variable (y) and select only those independent
variables which are strongly related to y. Second strategy is based on overall model
performance, here the contribution or impact of an independent variable is checked i.e.
whether it is able to help in predicting outputs or not.

4.2. Logistic Regression
Logistic regression is also a linear model. It differs from linear regression in the context that
in linear regression, the dependent variable is continuous whereas in logistic regression, the
dependent variable is discrete. Most of the times the dependent variable is binary discrete,
having value 0 or 1 i.e. falling into one category or another. The output of logistic regression
is probability of a input belonging to one category or another [James et al., 2014] . Linear
equation used in linear regression is used here also, but here linear equation is passed into a
logistic function (logistic function is shown in figure 4.2).
Figure 4.2 : Logistic Function
Thereby, giving the following equation :
𝑃 (𝑦 = 1 | 𝑥) = 1
1+𝑒−(𝛽0+𝛽1∗𝑥) (2)
where y is the output, x is the input and β0 + β1*x is the linear equation used in linear
regression. A threshold is selected, if the probability is above the threshold, input belongs to
one category, if the probability is below the threshold, it belongs to another. Threshold is
adjusted according to number of points classified correctly, most common threshold is 0.5. A
question could be, why use logistic function? The answer is that here the output is probability
is probability and probability is always between 0 and 1, to restrict the output between 0 and
1, logistic function is used.

4.3. Naive Bayes
Naive Bayes like logistic regression predicts if an input belongs to a particular category or
not. It's output is also probability but instead of logistic function, it uses Bayes rule. Figure
4.3 describes Bayes rule appropriately :
Figure 4.3 : Bayes Rule
So the equation is :
𝑃 (𝑦 = 1 | 𝑥) = 𝑃(𝑥 | 𝑦=1)∗𝑃(𝑦=1)
𝑃(𝑥) (3)
where y is output and x is input. P(x | y) and P(y) are calculated from the given data and used
for predicting outputs of new inputs. For different values of y, P(y | x) is calculated and x is
assigned to the category (or y is assigned the value) for which P(y | x) is maximum [Rish,
2001] . An assumption is made that variables x1,x2.... are independent of each other [Rish,
2001] and therefore, P(x1,x2.... | y) can be written as P(x1 | y) * P(x2 | y)....., which is not true
is most cases and that's why this algorithm is called Naive Bayes.

5. Related work
Recommender systems have been of great interest among the scientific community in the last
few years. Several successful approaches have been proposed. These approaches can be
classified by the domains they have been used in or applied to. Approaches are further
classified by the type of data used for recommendation (implicit or explicit) and lastly by the
type of algorithm used for recommendation (collaborative, content based filtering or hybrid
approach). In this section we will go through some papers which have proposed
recommender systems in domains like news, music etc. using both explicit and implicit
feedback or either of them in different algorithms.
Amongst the many papers which both inspired and motivated me to work on this topic, one
was [Li et al., 2010]. Here, news recommendation problem was modelled as a contextual
bandit problem. Authors devised a new algorithm named as LinUCB to solve the problem.
LinUCB is a general contextual bandit algorithm and can be applied to domains other than
news recommendation. Using LinUCB, number of clicks increased by 12.5% as compared to
a standard context-free bandit algorithm.
While talking about news recommendation, one cannot miss [Das et al., 2007]. Here, both
model and memory based algorithms have been used. Memory based algorithms use
weighted average of past ratings from other users where weight is proportional to the
'similarity' between users. Pearson correlation coefficient and cosine similarity are the typical
measures used for 'similarity'. Model based ratings model the user preferences based on past
information. From model based, they used clustering techniques PLSI and MinHash, and
from memory based, they used item co-visitation. All the three algorithms assign a score to a
story. Finally, two combinations of three techniques (PLSI, MinHash and co-visitation) were
used for evaluation. In one combination, co-visitation was given a higher weight than the
other techniques (2.0 instead of 1.0). This method was named CVBiased and in another
combination PLSI and MinHash were given higher weight (2.0 instead of 1.0), which is
known as CSBiased. The baseline used was recommendation based on recent popularity
called Popular. It was observed, on average both CVBiased and CSBiased performed 38%
better than the baseline Popular on live traffic.
Another influential paper is [Liu et al., 2010] , the work done in this paper was built on [Das
et al., 2007] . Here, information filtering was used. Information filtering removes unwanted
information from an information stream. First, they did an analysis of a users' interest over a
14 month period for each category of article (categories are pre-defined). They used click
distribution of a user to study the interest. Then they used Bayes rule to predict a users'
interest for a particular period of time. Then, predictions made for particular period were
combined to predict a users interest in a long period. The predictions made till now were
users' genuine interest. But a user also gets influenced by current news trends, so Bayes rule
is used to make predictions using click distribution of a short recent period (example last
hour). Finally, users' genuine interest is combined with current news trends to get IF (article).

The information filtering score for each article IF (article) is multiplied by collaborative
filtering score for the article CF (article), and the final score after multiplication of IF (article)
and CF (article) is used to rank the articles. The collaborative filtering score CF (article) is
obtained from [Das et al., 2007] . Using this method, click through rates (CTR) were
improved by 30.9% upon existing method i.e. [Das et al., 2007] .
While papers discussed previously were concerned with news domain in particular, [Hu et al.,
2008] is not concerned with any particular domain, instead here, focus was on
recommendations using implicit feedback dataset. Authors used latent factor models to
compute the number of times a user 'u' watched a program 'i' (rui). For finding the parameters
to compute rui , alternating least squares optimization technique was used with regularization
to avoid overfitting. This approach was to recommend television shows. The interesting thing
about this approach was there was only positive feedback and no negative feedback and this
made the task of recommendation much more difficult. Evaluation was done by the users
themselves as they were provided a list of television shows sorted from most preferred to
least preferred shows. Expected percentile rank was used for comparison where latent factor
model outperformed both popularity model and neighbourhood model.
In [Baltrunas and Amatriain, 2009], micro-profiling of users was used for music
recommendation. Several sub profiles of each users were created. Instead of a single user
model several models were created to make recommendations with different context. In this
paper, context was time. So several sub-profiles were created for morning, evening, summer,
winter, etc. The idea behind these sub-profiles was that music preference of a user varies with
time but it also remains same for a particular time. To make it more clear, an example could
be, one might like inspirational music in the morning and romantic music in the evening so
there is a difference between morning and evening music preference but music preference for
morning remains same, it doesn't change from inspirational to any other type. Only implicit
feedback was used here, so as in [Hu et al., 2008], dataset only had positive opinions and not
negative. For training, popularity of an artist was considered, i.e. number of times a user has
listened to a particular artist at a particular time domain (morning, evening, etc.). In this
paper, artists were recommended rather than a music track. Offline evaluation was done,
where mean absolute error was computed for each time domain. And then, a weighted
average of mean absolute error was taken where weight was the number of ratings that can be
predicted in a particular time domain. The best system would be the one which could
minimize the weighted average of mean absolute error. In conclusion, it was observed that
contextual (different time domains) micro-profiling gave better results than a context free
user model.
In [Parra and Amatriain, 2011], focus was again on mapping implicit feedback to explicit
ratings. The domain was music recommendation. The contributions of this paper are, a study
of the relation between implicit feedback and explicit ratings, analysis of impact of other
variables and a linear model to predict ratings. A user study was conducted to collect data,
users were asked to rate an album on a 1 to 5 scale. Three variables were selected for
analysis, implicit feedback (play count for a user on a given item), global popularity
(aggregated play count of all users on a given item) and recentness (time elapsed since user

played a given item). Values of all the three variables were divided into three bins - low,
medium and high. It was observed that quantized implicit feedback and distribution of ratings
were related, quantized recentness and distribution of ratings were also related but not as
strongly as implicit feedback and no significant relation was found between quantized global
popularity and distribution of ratings. Analysis of relationship between other variables
showed some relation between recentness and implicit feedback only and not any other pair.
Linear regression was used for predicting ratings using three variables and RMSE (root mean
squared error) was used for evaluation. Four models were used for prediction, first one was
only with implicit feedback, in second one recentness was also added with implicit feedback,
in third model global popularity was added to the second model and in the fourth and last
model a variable which was multiplication of implicit feedback and recentness was added to
the third model. the baseline used was average of ratings. Although performance of all
models were relatively same, but they all had average improvement of 6.5% over the
baseline.
[Parra et al., 2011] was built upon [Parra and Amatriain, 2011]. In [Parra et al., 2011], instead
of linear regression, logistic regression was used for prediction. Here, a random effect of user
was also added in the model. Finally, rating is predicted by computing expected value of
rating across all ratings and their probabilities for a user and a item. A second dataset was
collected which contained only implicit feedback i.e. play count of each album per user.
Experiments were performed on this dataset and also the dataset collected in [Parra and
Amatriain, 2011] . While conducting experiments on first dataset [Parra and Amatriain, 2011]
, a variable, concerts per year and random effect of user was used with the three variables
used in [Parra and Amatriain, 2011] . In case of second dataset, as it did not have concerts per
year variable in it all the other variables used in first dataset were used. In the end, results of
four approaches were compared. First approach was the one used in [Hu et al., 2008], in
second approach log transformation of play counts used in [Hu et al., 2008] were used, third
approach is the one where logistic regression was used with implicit feedback, recentness and
global popularity, in fourth approach concerts per year is also used with the variables used in
third approach, fourth approach was only used on first dataset as concerts per year was
available for that dataset only. For evaluation, Mean average precision (for relevant albums)
and normalized Discounted cumulative gain (for ranking of albums) metrics were used as
final goal in this paper was recommending albums rather than just predicting ratings. Most
popular album was used as baseline and it outperformed all the other approaches in the first
dataset as the dataset was very sparse. In second dataset, results were different, baseline
performed worst, second approach improved upon first one and although third approach
performed better than first one it was more or less equal to second approach.
Another paper which is built upon [Hu et al., 2008] is [Johnson, 2014]. Here also like [Hu et
al., 2008], latent factor models are used, though here the approach is probabilistic which was
not the case in [Hu et al., 2008]. Here, the algorithm used was called logistic matrix
factorization. The probability of interaction between a user and a item is given by a logistic
function parameterized by the sum of the inner product of user and item latent factor vectors
and user and item biases. Similar to [Hu et al., 2008], a confidence function is chosen, this

function can be replaced any other function, in the paper it is given by αrui where α is the
tuning parameter and rui is rating of user u for item i. Then, likelihood of the observation
matrix is computed, regularization is also performed using zero mean spherical Gaussian
priors to avoid overfitting. Local maximum of likelihood function is found using alternate
gradient descent method. As in [Hu et al., 2008], here also play counts are used for
prediction, a play count is only recorded if a user has listened to a song for more than 30
seconds continuously thereby removing the bias that a user started a song but didn't like it.
Mean Percentage Ranking (MPR) metric was used for evaluation as was the case in [Hu et
al., 2008]. Logistic Matrix Factorization was able to outperform both the model used in [Hu
et al., 2008] and also baseline popularity model.
For recommending news, [Li et al., 2010], [Das et al., 2007] and [Liu et al., 2010] used
history of logged-in users, [Garcin et al., 2012] considers only current visit data where users
don't log in. Authors decided to build a candidate set of most fresh news articles. Potential
candidates for recommendation were the articles in candidate set which were accessed at least
once. Most popular news article was taken as baseline. Three approaches were considered,
collaborative filtering at the level of news items, content based recommendation and a hybrid
approach where collaborative filtering is applied at the level of topics. In the first approach,
news reading was modelled as a Markov process given the sequential nature of news reading,
news articles were considered states. Two parameters were used at this stage namely past and
future, past refers to the number of news articles to be considered for recommendation
whereas future refers to the number of steps needed to reach a given state. Transition
probability is computed for pairs, using these transition probabilities a score is generated
which is reading probability of a news item n in f steps given the sequence s, here the order of
sequence matters. In second approach, Latent Dirichlet Allocation (LDA) was used with
cosine similarity. In third and final approach, news articles were clustered based on topic
distributions, then a transition matrix is build for clusters. K-means is used for clustering.
Once the clusters are determined, a probabilistic model is build similar to first approach by
replacing a sequence with a cluster. Two metrics were used for evaluation, Mean average
precision (MAP) and S@5, it is equal to 1 if the immediate successor of the current items is
recommended among the first 5 recommended news stories, 0 otherwise. The training set
contains all news items accessed before time t and the testing set has items accessed after
time t. First approach i.e. collaborative filtering at the level of news items outperformed
others in both metrics.
There are similarities and differences between the works described above and the work we
have done. The similarity and difference is in the evaluation metric, machine learning model
and features used in the model. There are papers which have used Naive Bayes [Liu et al.,
2010] , linear regression [Parra and Amatriain, 2011] and logistic regression [Parra et al.,
2011] , but none of them have shown a comparison of these three models. In evaluation,
RMSE has been used [Parra and Amatriain, 2011] but most of them have preferred MAP and
other metrics, some have even tested their systems on live traffic. In features, none of them
have used joint and mean conditional probabilities of labels givens by users for a news
article, however, we have experimented with feature novelty which is similar to feature
recentness used in [Parra and Amatriain, 2011] and [Parra et al., 2011] .

6. Experiments
6.1. Overview
In this section, we will give an overview of our experiments and also explain the extracted
features and baseline used in our experiments. The first step in performing experiments was
to collect data, instead of collecting data ourselves (as process of data collection can be both
expensive and time consuming), we downloaded data used in a CMU (Carnegie Mellon
University) project which seemed suitable for our experiments. The details of the data set are
given in section 6.2. Next step was data pre-processing where we removed ambiguities and
handled missing values, detail are in section 6.3. After data pre-processing, process of feature
extraction was performed, details of how feature extraction was done is in section 6.4. After
feature extraction, different features were used in different algorithms (as described in
sections 4.1, 4.2 and 4.3) to predict user ratings. In section 6.5, we have presented the results
of our experiments.
Although the features available in the data set can be used for news recommendation, we
decided to extract features and use them also for news recommendation. Two features were
extracted namely Joint Conditional Probability (JCP) and Mean Conditional Probability. The
difference in these two features is at the last step of computation. We calculate the
conditional probability of individual labels or tokens (given by an user to an article) given (or
conditioned upon) user rating of the article. If an article has been given only one token or
label, then there will be no difference between JCP and MCP. If an article has been given
multiple tokens or labels, then there can be difference between JCP and MCP. The difference
is as follows :
Joint Conditional Probability (JCP) : Conditional probability of individual tokens or labels
are multiplied together and the result of the multiplication of the conditional probabilities of
individual tokens or labels is JCP.
Mean Conditional Probability (MCP) : For MCP, mean of conditional probabilities of
tokens or labels is computed and the mean is known as MCP.
The motivation behind extracting and using these particular features came from [Liu et al.,
2010] . Using similar features in Naive Bayes, authors of [Liu et al., 2010] were able to
develop a very efficient news recommender system. Although, [Liu et al., 2010] was built
upon [Das et al., 2007] , what caught our eye was the simplicity of features and methods used
in [Liu et al., 2010] . While doing literature review, we couldn't find any paper where these
features were used, what we found was people using linear and logistic regression with
features like recentness, etc. So we decided to use JCP and MCP in linear regression, logistic
regression and Naive Bayes for performing news recommendation.

Baseline : For baseline, we used popularity model. In popularity model, the most popular or
most frequent value is always predicted. For example, if user rating '5' is most frequent when
compared to other values of user rating in the given data set, then each predicted user rating
will be 5.
6.2. Dataset
The dataset we used for our experiments was collected in Carnegie Mellon University for the
project PIIR (Proactive Personalized Information Integration and Retrieval) [Özgöbek et al.,
2014]. Original unclean dataset contained information on 10010 user ratings on news articles.
Dataset contains both explicit and implicit feedback. Table 6.1 shows the description of
explicit feedbacks in the dataset.
Table 6.1 : Description of explicit feedbacks.
Feedback
Description
user_like
An integer between 1 and 5 representing how much a user likes an article.
Relevant
An integer between 1 and 5 representing how much a user thinks an article is
relevant.
Novelty
An integer between 1 and 5 representing how new the article is according to the
user.
Readability
Representing whether the article is readable or not is according to the user, is
either 0 or 1.
Authority
Representing whether the news article is authoritative or not according to the
user, is either 0 or 1.
Classes
A string representing the classes news article belongs to (for example, |news),
an article can belong to multiple classes (|news |movies |Broadway |music), a
user can create any class, there are no predefined classes. '|' serves as delimiter.
There were no classes for some articles which means the user didn't give a
feedback.
The variables described in Table 6.1, except the variable "classes", contain value -1 for few
articles. This means user didn't provide any feedback for that variable in a specific article.
Table 6.2 shows the description of implicit feedbacks in the dataset. Table 6.3 and 6.4 gives
statistics of both explicit and implicit feedbacks in the data set.

Table 6.2 : Description of Implicit feedbacks
Feedback
Description
TimeOnMouse
Number of milliseconds the user spent on moving the mouse.
TimeOnPage
Number of milliseconds the user spent on a page or news article.
EventOnScroll
Number of clicks on the scroll bars.
ClickOnWindow
Number of clicks inside browser window but not on scroll bars.
TimeOnHScroll
Number of milliseconds the user spent on using horizontal scroll.
TimeOnVScroll
Number of milliseconds the user spent on using vertical scroll.
NumOfPageUp
Number of pages the user scrolled up.
NumOfPageDown
Number of pages the user scrolled down.
MSecForPageUp
Number of milliseconds the user spent on scrolling a page up.
MSecForPageDown
Number of milliseconds the user spent scrolling a page down.
NumOfUpArrow
Number of clicks on up arrow key.
NumOfDownArrow
Number of clicks on down arrow key.
MSecForUpArrow
Number of milliseconds the user spent on up arrow key.
MSecForDownArrow
Number of milliseconds the user spent on down arrow key.
To use variables as features in the models (sections 4.1, 4.2, 4.3), we selected all the variables
in the explicit feedback with the exception of 'user_like' because it is the target variable. The
purpose of selecting all the variables was to check their impact on predicting user ratings. In
implicit feedback however, we selected only 'TimeOnMouse' and 'TimeOnPage', rejecting all
others as they were sparse vectors (a vector whose most of the elements are zero).
Table 6.3 : Statistics on explicit feedback
Feedback
min
Max
Mean
Variance
user_like
1
5
3.35
1.68
Relevant
1
5
3.41
1.86
Novelty
1
5
3.43
2
Readability
0
1
0.87
0.11
Authority
0
1
0.82
0.15

Table 6.4 : Statistics on implicit feedback
Feedback
min
max
Mean
Variance
TimeOnMouse
0
320611
1976.4
29049887.3
TimeOnPage
0
5138378
71430.3
18683642234
EventOnScroll
0
160
0.92
13.7
ClickOnWindow
0
81
0.87
5.5
TimeOnHScroll
0
0
0
0
TimeOnVScroll
0
0
0
0
NumOfPageUp
0
19
0.1
0.67
NumOfPageDown
0
19
0.12
0.86
MSecForPageUp
0
5978
17.3
32784.86
MSecForPageDown
0
7391
22.46
50489.13
NumOfUpArrow
0
23
0.08
0.52
NumOfDownArrow
0
91
0.96
18.24
MSecForUpArrow
0
7378
23.26
46440.9
MSecForDownArrow
0
17743
175.1
641665.3
6.3. Data Pre-processing
Data pre-processing is usually the first step of data mining and knowledge discovery. Data
pre-processing can have an impact on generalization performance of Machine Learning
algorithm [Kotsiantis et al., 2006]. Most of the real world datasets suffer with problems of
missing values and ambiguities, similar was the case with our dataset. To deal with these
issues in the feedback 'classes', data pre-processing was done. There were a few steps
involved in data pre-processing:
 First, we removed all the rows or information on articles which didn't belong to any
class. There were 2094 user feedbacks in which article didn't belong to any class at
first, after removing all of them we were left with around 8000 user feedbacks and
540 classes.

 Next, we removed ambiguities in the feature 'classes', ambiguities like there were
several classes 'USA' , 'usa' and 'us' referring to the same country or entity, we
replaced all of them with a single class.
 After the second step, we removed all the classes which occurred only once as
machine learning requires occurrence of each item at least twice so that one can be
used in training and other for testing.
After the third step, we checked again if there was any article still left which didn't belong to
at least one class. We stopped when we had removed sufficient amount of ambiguities and
each article belonged to at least one class.
As for the other features, we replaced blanks with '0' in variables 'TimeOnMouse' and
'TimeOnPage', for variables like 'readability', 'novelty', 'relevant' and 'authority', blanks and -1
were replaced with lowest value they could have i.e. 1 for 'relevant' and 'novelty', 0 for
'readability' and 'authority'.
After performing all the steps above, we were left with 6625 user feedbacks and 335 different
classes.
6.4. Feature Extraction
For the reasons explained in section 6.1, we wanted to use conditional probabilities of
'classes' in models like Linear and Logistic Regression, so, we extracted two features from
the feature 'classes' namely 'joint conditional probability' and 'mean conditional probability'.
The steps involved in the process were:
 First, separate classes into tokens, for example, if an article belongs to classes 'sports'
and 'basketball', the two tokens will be 'sports' and 'basketball' (in case if an article
belongs to only one class, then there is only one token). This is done for all articles.
 Compute the joint frequency of each token with each 'user_like' i.e. {1,2,3,4,5}, joint
frequency is the total number of times a token appears with a particular 'user_like'.
 Compute the frequency for all values of 'user_like'.
 Now, we compute the conditional probability of token given 'user_like' or
𝑃 (𝑡𝑜𝑘𝑒𝑛|𝑢𝑠𝑒𝑟_𝑙𝑖𝑘𝑒). We do this by dividing joint frequency of token and 'user_like'

by the frequency of the particular 'user_like' which has appeared with token. For
example,
𝑃 (′𝑠𝑝𝑜𝑟𝑡𝑠′|𝑢𝑠𝑒𝑟_𝑙𝑖𝑘𝑒 = 5) = 𝑗𝑜𝑖𝑛𝑡 𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦 𝑜𝑓 ′𝑠𝑝𝑜𝑟𝑡𝑠 ′ 𝑎𝑛𝑑 ′𝑢𝑠𝑒𝑟 _𝑙𝑖𝑘𝑒 ′ = 5
𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦 𝑜𝑓 ′𝑢𝑠𝑒𝑟 _𝑙𝑖𝑘𝑒 ′ = 5 (4)
 To calculate joint conditional probability, we multiply the conditional probabilities of
all the tokens of an article (only if there are multiple tokens, otherwise conditional
probability becomes joint conditional probability).
 To calculate mean conditional probability, we take the mean of the conditional
probability of all the tokens of an article (only if there are multiple tokens, otherwise
conditional probability becomes mean conditional probability).
We also calculated 'TimeOnMouse' and 'TimeOnPage' in seconds to use as features. The new
features were named 'TimeOnMouse (sec)' and 'TimeOnPage (sec)'.
6.5. Results
To present our results in a concise way, we have assigned our features short aliases. Table 6.5
provides description of each alias.
Table 6.5 : Feature names and their respective aliases
Feature name
Alias
Relevant
RL
Novelty
N
Readability
RD
Authority
A
TimeOnMouse
TM
TimeOnPage
TP
joint conditional probability
JCP
mean conditional probability
MCP

Results that we obtained after performing experiments (Linear Regression, Logistic
Regression and Naive Bayes) on our pre-processed dataset are shown in the Figures 6.1, 6.2
and 6.3 (these results are for all users, results for individual users are given in Figures 6.5, 6.6
and 6.7).
Figure 6.1 : Linear Regression on Pre-Processed Dataset
Figure 6.2 : Logistic Regression on Pre-Processed Dataset
0
0.2
0.4
0.6
0.8
1
1.2
1.4
MSE
Features
Linear Regression(Pre-Processed)
with JCP
with MCP
Alone
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
RL,N,RD,A,TM,TP
RL,N,RD,A,TM(s),TP…
N,RD,A,TM(s),TP(s)
N, RD, A, TM, TP
RD, A, TM(s),TP(s)
RD, A, TM, TP
A,TM(s),TP(s)
A,TM,TP
TM(s),TP(s)
TM,TP
TM
TP
TP(s)
TM(s)
JCP
MCP
RL, N, RD, A
RL
N
RD
A
RL, A
RL, RD
RL, N
MSE
Features
Logistic(Pre-processed)
JCP
MCP
Alone

Figure 6.3 : Naive Bayes on Pre-processed Dataset.
After we observed that features 'relevant (RL)' and 'novelty (N)' were performing
exceptionally well alone, we decided to compute each feature's correlation with 'user_like',
our results are presented in Figure 6.4.
Figure 6.4 : Correlation of user ratings with different features
From Figure 6.4, we can clearly see that features 'relevant (RL)' and 'novelty (N)' are strongly
correlated with 'user_like', 'readability (RD)' and 'authority (A)' are somewhat correlated with
'user_like' whereas other features are not at all correlated with 'user_like'. This gives a good
0
1
2
3
4
5
MSE
Features
Naive Bayes(Pre-processed)
with JCP
with MCP
Alone
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Corr with
user_like
Features
Correlation
correlation

indication that features 'relevant (RL)','novelty (N)', 'readability (RD)' and 'authority (A)' will
be able to approximate 'user_like' better than other features.
We also have to take into account that all these strongly correlated (with 'user_like') features
are explicit feedback, in a real world scenario a user doesn't really give that much explicit
feedback and systems have to rely on implicit feedback. As we know that usually a class an
article belongs to is given by the system and not the user (in our dataset classes are given by
user), so in further experiments we have reduced the number of classes and performed our
experiments to see the changes in the performance of features in different models.
Figure 6.5 : Linear Regression results for individual users
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
M
S
E
User
Linear Regression results for individual users
N,MCP
N,JCP
N
MCP
JCP

Figure 6.6 : Logistic Regression results for individual users
Figure 6.7 : Naive Bayes results for individual users
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
M
S
E
Users
Logistic Regression results for individual users
N,MCP
N,JCP
N
MCP
JCP
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
M
S
E
Users
Naive Bayes results for individual users
MCP, N
JCP,N
N
MCP
JCP

In Figures 6.1, 6.2 and 6.3 are the results of different models on all users, whereas in Figures
6.5, 6.6 and 6.7 are the results of different models on individual users.
With the different features available to us, models could be tested with many combinations of
features, so we decided to go with few combinations in case of all users. But as we know that
most of these features are explicit feedbacks and won't be available in a real world data set,
so we dropped all of the features and kept only novelty (N) as it is used in recommender
systems as evident by [Parra and Amatriain, 2011] and [Parra et al., 2011] . For individual
users, we have only shown results using feature novelty (N) and our extracted features i.e.
JCP and MCP. For baseline, we used popularity model i.e. recommending the most popular
article, this is also a very common practice as evident by [Hu et al., 2008], [Parra and
Amatriain, 2011] , [Parra et al., 2011] and [Johnson, 2014] .
In case of all users, the result was that linear regression model with our extracted features was
able to outperform the popularity model or baseline but it was unable to outperform the
explicit feedbacks as they were strongly correlated with user_like and our extracted features
were not as clear in Figure 6.4.
In case of individual users, linear regression model with our extracted features was able to
outperform the popularity model or baseline for most of the users but not all.
We performed experiments by reducing the number of classes too and found out that number
of classes was having an impact on performance of our models which contradicts the
founding in [Garcin et al., 2012].
Below in Figure 6.8 are the mean square errors we got using only extracted features with
different models which shows that linear regression outperforms all other models with our
extracted features . And Figure 6.9 shows impact of number of classes (these results are on all
users rather than individuals), as the number of classes decrease, performance of models
increases (or mean square error decreases).
Figure 6.8 : Performance of Extracted Features
1.1
1.15
1.2
1.25
1.3
1.35
1.4
1.45
Baseline
Linear
Logistic
Naive
M
S
E
Models
Performance of Extracted Features
JCP
MCP

Figure 6.9 : Impact of Number of classes
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
335
100
80
60
40
20
M
S
E
Number of classes
Impact of Number of classes
Linear
Logistic
Naive

7. Conclusions and Future work
In this thesis, we have shown how the class of an article can be used for news
recommendation. We have seen step by step how our system developed. First, we defined
and formalized the problem. Then, we pre-processed the data thereby handling ambiguities
and missing values. After that, we extracted two features. And lastly, we used the extracted
features and features available in the data set to predict how much a user likes a news article.
The most important conclusion that we can draw from our results is that the class of the
article can be used to improve a recommendation algorithm. News articles are strongly
related to its class or category and this has an impact on users preferences. Using a supervised
algorithm we were able to observe the impact of this feature (i.e. articles’ class). Out of the
three learning models i.e. linear regression, logistic regression and Naive Bayes, linear
regression turned out to be the best one.
Extracted features were able to outperform popularity model but they are still not able to
outperform explicit feedback like novelty (N), at least not with the three models we have
used.
Reducing the number of classes has an impact on performance of every model. As the
number of classes decrease, performance of every model increases i.e. number of classes and
performance of a model are inversely proportional to each other. This finding is contradictory
to finding of [Garcin et al., 2012] where they found no impact of number of classes on
performance. The difference could be because in our dataset, 'classes' was an explicit
feedback, it is possible that in their dataset the feedback which is similar to 'classes' is
implicit feedback.
Extracted features i.e. Joint Conditional Probability (JCP) and Mean Conditional Probability
(MCP) though performed reasonably well alone, average of mean square error on all datasets
was around 0.98 for linear regression and around 1.14 for logistic regression and Naive
Bayes, however they were unable to have any influence on other features.
Naive Bayes turned out to be the weakest among three models as it performed worse with
TimeOnMouse (TM) and TimeOnPage (TP), with average of mean square error being above
4. The bad performance of Naive Bayes was not surprising because Naive Bayes is not a
robust classifier, and the problem in hand requires a model that is able to detect correlations
that are not easily detected with small datasets
In future work, we would like to test our approach on a dataset where classes are implicit
feedback rather than explicit feedback thereby performance would be more real world and we
would also be able to observe any difference in performance of implicit versus explicit
feedback.
There are many other options of algorithms that perform supervised learning. In the present
thesis, we have worked with only three of them, however it would be interesting to evaluate

others. Another interesting approach would be to solve the problem of news recommendation
as unsupervised learning problem. For example, models that are able to detect underlying
variables that are enclosed in the news articles (i.e. LDA).
However, our approach is one of the many approaches and in the future we would like to
propose different and more complex approaches with ultimate goal of designing a news
recommendation system which are able to improve the state-of-the-art.

References
[Li et al., 2010] L. Li, W. Chu, J. Langford, and R.E. Schapire. A contextual-bandit approach
to personalized news article recommendation. In Proceedings of the 19th International
Conference on World Wide Web (WWW), 2010.
[Das et al., 2007] A. Das, M. Datar, A. Garg, S. Rajaram. Google news personalization:
scalable online collaborative filtering. In Proceedings of the 16th International Conference on
World Wide Web (WWW), 2007.
[Liu et al., 2010] J. Liu, P. Dolan, E. Pederson. Personalized news recommendation based on
click behavior. In Proceedings of the 14th International Conference on Intelligent User
Interface (IUI), 2010.
[Hu et al., 2008] Y. Hu, Y. Koren, and C. Volinsky. Collaborative Filtering for implicit
feedback datasets. In Proceedings of 8th International Conference on Data Mining, 2008.
[Baltrunas and Amatriain, 2009] L. Baltrunas and X. Amatriain. Towards time-dependant
recommendation based on implicit feedback. In Workshop on Context-Aware Recommender
Systems in 3rd ACM Conference on Recommender Systems, 2009.
[Parra and Amatriain, 2011] D. Parra and X. Amatriain. Walk the talk: Analyzing the relation
between implicit and explicit feedback for preference elicitation. In Proceedings of the 19th
Conference on User Modeling, Adaptation, and Personalization, 2011.
[Parra et al., 2011] D. Parra, A. Karatzoglou, X. Amatriain, and I. Yavuz. Implicit feedback
recommendation via implicit-to-explicit ordinal logistic regression mapping. In Workshop on
Context-Aware Recommender Systems in 5th ACM Conference on Recommender Systems,
2011.
[Johnson, 2014] C. C. Johnson. Logistic matrix factorization for implicit feedback data. In
Workshop on Distributed Machine Learning and Matrix Computations in Advances in Neural
Information Processing Systems 27, 2014.
[Garcin et al., 2012] F. Garcin, K. Zhou, B. Faltings, V. Schickel. Personalized News
Recommendation Based on Collaborative Filtering. In Proceedings of the International Joint
Conference on Web Intelligence and Intelligent Agent Technology, 2012.
[Mitchell, 1997] Machine Learning, Tom Mitchell, McGraw Hill, 1997.
[Rish, 2001] I. Rish. An empirical study of the Naive Bayes classifier. IJCAI Workshop on
Empirical Methods in Artificial Intelligence, 2001.
[Poriya et al., 2014] A. Poriya, N. Patel, T. Bhagat and R.Sharma. Non-Personalized
Recommender Systems and User-based Collaborative Recommender Systems. International
Journal of Applied Information Systems 6(9), 22-27 (2014).

[Ricci et al., 2011] F. Ricci, L. Rokach, B. Shapira and P.B. Kantor, Recommender Systems
Handbook, Springer, 2011.
[Burke, 2002] R. Burke. Hybrid Recommender Systems: Survey and Experiments. User
Modeling and User-Adapted Interaction 12(4), 331–370 (2002).
[James et al., 2014] G. James, D. Witten, T. Hastie & R. Tibshirani. An Introduction to
Statistical Learning with Applications in R, Springer, 2014.
[Özgöbek et al., 2014] Ö. Özgöbek, N. Shabib, J. A. Gulla. Data Sets and News
Recommendation, 2nd International Workshop on News Recommendation and Analytics,
2014.
[Kotsiantis et al., 2006] S.B. Kotsiantis, D. Kanellopoulos, P.E. Pintelas. Data preprocessing
for supervised leaning. International Journal of Computer Science 1(2), 111-117 (2006).