Introduction

Uploaded by

Actor shorts

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as DOCX, PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

20 views3 pages

Introduction

Uploaded by

Actor shorts

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as DOCX, PDF, TXT or read online on Scribd

You are on page 1/ 3

INTRODUCTION

SOCIAL media provides a public platform to easily exchange information

with each other, and nowadays people spend a lot of time every day on various
social media platforms. Since social media occupies a large part of the daily lives
of modern people, many people are interested in researching how to extract data
from social media. An example of information that could be gained from social
media is the popularity score. Specifically, this score tells how many people
viewed a post, and a larger number of views means more influence. Social media
popularity prediction (SMP) is the task of estimating the popularity score using the
available data of a given social media post.

Estimating the popularity score is hard because of the many and complex
factors that affect popularity. Quality of content and relevance to viewers are some
of the factors, and these are difficult to measure. Other factors such as real-life
events are tough to include in a prediction model. Recent SMP methods attempt to
tackle these complex factors by adding more modalities [4, 5, 7, 12, 17], such as
images [14, 39], relationship networks [25], temporal context [13], tags, and
categories.

Although increasing the number of modalities is a good approach to the

works, it also increases the complexity of the model, in terms of architecture,
memory consumption, number of modules, etc. Alternatively, the paper [7, 26, 27,
28, 29, 30] is also a multi-modal approach but in its pipeline, it represented images
as captions (i.e. texts). Different modalities could be converted to another modality
using existing technologies. Image captioning converts images to texts. There exist
speech-to-text methods already. From the social graph of a post, we could extract
different numeric values, such as the number of the neighbors for each node.

Moreover, the popularity of posts may be affected by user information.

Many studies have shown that there is a high correlation between image popularity
and users [20, 32, 33]. One of the reasons is that the users have their own
followers, different users may have different numbers of followers. Generally,
posts written by the user with more followers have a higher chance to receive more
views and likes. And the temporal and spatial information may affect the
popularity as well, the earlier post should get more people’s attention, and if the
user uploads the post in a special location, it will attract more attention too.

In this paper, we proposed a network that exploits semantic (text) and

numerical (number) modalities to estimate the popularity of a social media post
based on the self-attention mechanism. Due to the data type discrepancy, we
divided the data into semantic and numerical branches. In the semantic branch, the
image contents are transferred to caption texts and tags, all of the textual features
are converted into tokens, each token has an associated with word embedding [23],
since the attention mechanism [9] is shown effective to extract contextual
information, to better aggregate the sequence of embedding, we also develop a
feature attention mechanism for the purpose, which can deal with dispensing
recurrence, and convolutions entirely. Using only the semantic features modality is
not sufficient for some types of social media posts, so we used the numerical
features as well which can be easily converted into scalars, such as timestamps,
geo location. After preprocessing, we extracted and fused the features in both
modalities respectively, and assemble two models to calculate the popularity score.
The contributions of this work are 3 fold:
_ We designed a network that adopts an attention mechanism and exploits multiple
features in two modalities to perform model ensemble, the network can be easily
extended to include more different modalities furthermore, which is able to solve
problems with heavy categories.
_We analyzed the influence of semantic features on the model performance.
Moreover, we generated additional numerical features, the result indicates the
derived features are beneficial to improve our network performance.
_We demonstrated that our method outperforms the other state-of-the-art methods
in Social Media Popularity Dataset.