Finding the right users to be chosen as targets for advertising campaigns is not a trivial task, and it may allow important commercial advantages. A novel approach is presented here for the recommendation of new possible consumers to brands interested in distributing advertising campaigns, ranked according to the “compatibility” between users and brands. A database containing both descriptions associated with diferent brands, and textual information about users' opinions on diferent topics, is required in input. Then, sentiment analysis techniques are applied to measure to what extent the users match with the brands, based on the texts associated with their opinions. The approach has been tested on both synthetic and real datasets, and with two diferent formulations, showing promising results in all the considered experiments.
An important issue in the context of digital advertising is
how to optimize the efects of marketing communication,
trying to involve in advertising campaigns those potential
ing the distribution of advertisements to uninterested
users. Automatic systems able to suggest a set of target
users for advertising campaigns, possibly ranked
according to their potential approval rating, provide three main
benefits: (i) minimization of costs for the dissemination
of the advertising campaign through digital media, which
is often very expensive; (ii) improvement of the user
experience, since only the possibly interested customers are
contacted with advertisements which could be useful for
them; (iii) avoid the spread of useful information through
the social and other digital channels.
dation of a list of possible consumers to be suggested
as target for a specific advertisement campaign, ranked
moting the campaign. In particular, we assume that a
database containing a number of descriptions associated
with diferent brands is available, and a number of
potential customers together with their opinions on diferent
topics as well. Suitable “tag” may be generated to
summarize the brand descriptions, and the proposed approach is
based on the adoption of sentiment analysis techniques
[
CEUR
Workshop Proceedings (CEUR-WS.org) with the brands, based on the texts associated to their opinions. The proposed approach is the core of a more general framework, implementing a big data analytics platform.
The presented methodology has been validated first
Characteristic (ROC) analysis [
The authors of [
Here we propose a novel approach for the recommen- tags, have been evaluated, showing that the proposed according to their “compatibility” with the brand pro- in classifying potential new customers has been tested Let = { 1, 2, … , } be a set of users, representing possible customers of brands in the set = { 1, 2, … , }.
The main aim here is to return a set = { 1, 2, … , } of ranks, one for each brand, each containing the list of users in , ranked according to their potential “match” with the corresponding brand. This allows managers of advertising campaigns to select as targets only those users who may be potentially the most interested ones.
In order to understand to what extent each user matches with each brand, we assume that suitable tags may be associated with the brands. As an example, keywords may be extracted from textual exploration of their websites; also, such tags could be explicitly proposed by the campaign managers, according to the specific products object of the campaign. Let 1, 2, … , be the lists of tags associated to 1, … , , respectively. It is worth pointing out that diferent brands may share some common tags.
Let ∈ ( = 1, … , ) be a user and ∈ ( = 1, … , ) be a brand. The Compatibility Index between and is defined as:
In this section the approach proposed here is described in detail. In particular, it can be considered as the core of a more general platform, based on big data analytics as illustrated in Figure 1. The module Sentiment Analysis corresponds to the methodology described here.
Two main aims can be identified: (1) given a set of potential users, finding the best targets for an advertising campaign, and (2) given a new user in input, returning the best brand for which it can be a potential new customer. Sections 3.1 and 3.2 show these two diferent formulations of the presented approach. where | | is the number of tags associated to , and (, ) is the match, intended as a sort of liking rate, of the user for the -th tag of ( = 1, … , | |).
In particular, the match (, ) is obtained by
sentiment analysis techniques [
Polarity can be expressed as a numerical rating, representing a sort of sentiment score. There are diferent approaches to identify the sentiments expressed on topics. The main algorithms proposed for polarity computation can be distinguished in two main categories: , where each rank is a set of triplets ⟨ , , (, )⟩ , sorted according to (, ) . A toy example is illustrated below.
I C 0.83 0.83 0.83 0.67 0.33 0.33 0.17 0.00
I C 1.00 0.87 0.87 0.62 0.62 0.50 0.37 0.37 Example 1 Let = { , , , , , , , } has has = {ℎ , ,
} = {ℎ, , , } and = { , , } users and brands stored in the input database, respectively. Ferrari has tags = { , , } be the , Oracle , and Walmart the features are the polarity values retrieved from the texts associated to the users in the input database, for the set of all tags associated to all brands. The K-Nearest Neighbors (KNN) classical approach can be then applied to predict the class label for each new potential customer, represented analogously by a features vector, by computing the (e.g., Euclidean) distance between this vector and those in the classes, and choosing as class label the most represented in the top neighbors. The following , respectively. example shows the KNN for the dataset in Example 1. • Supervised machine learning algorithms, trained sentiment, using for example support vector ma- tween each user and brand.
textual descriptions stored for users in the database (e.g., Example 2 Walmart, respectively. and their tweets), while Table 2 shows the three ranks ,
obtained for the three brands Ferrari, Oracle and
As the result of this toy example, Karmen, Pedro and William are the best targets for Ferrari; Marian and Andrea for Oracle; Sonia, Pedro, Janny, Karmen and Natan for Walmart. 3.2. Prediction Based on K-Nearest
Neighbors Let 1, 2, … , be classes, each associated to a brand in , respectively. Objects in the classes are vectors of features, such that each vector represents an user and
Let Ferrari = {Karmen, Pedro}, Oracle = {Marian, Andrea} and Walmart = {Sonia, Janny} be three classes, such that the users are represented by the corresponding rows in for which the class labels are to be predicted, and = 3 . For Natan, the closest users are Pedro, Sonia and Janny, therefore it will be put in the class Walmart. William has Karmen, Pedro and Sonia as closest neighbors, therefore its predicted class label is Ferrari.
Three classes have been built by choosing three tags for each, and by generating 15 users who like the corresponding brand and 15 who do not like it. The users have been obtained by generating sentences explicitly expressing positive (negative, respectively) opinions on that brand. Then, pairs of users and brands such that the user likes the brand, have been used as true positives, while pairs
4.1. Benchmarking of Sentiment Analysis
Libraries
An important problem in the considered context is that
it is often very dificult to find “true positive” and “true Here we have referred to the unsupervised lexicon-based
negative” samples in order to validate the proposed ap- approaches for the polarity computation, due to the fact
proaches. In particular, we have referred to the Twitter that they best fit with the considered context. In
parsocial network for the real data considered here, as ex- ticular, Figure 2 shows the number of negative, neutral
plained in detail in Section 4.3. If one wants to validate and positive polarity values returned by the considered
the proposed approach, that is based on textual infor- algorithms (implemented by the corresponding Python
mation stored in a database, by using independent in- libraries). It is evident that the results may be also very
formation not related to the available textual one used diferent for the evaluated sentiment analysis techniques,
for the classification, a simple method is to search for that use diferent dictionaries and approaches to calculate
those users who are also “followers” of the considered the polarity . For instance, the ’Flair’ library tends to rate
brands. Indeed, to follow a brand, the user has made an the tweets only positively or negatively, diferently than
action, and this can be considered as an explicit declara- the other ones.
tion of compatibility with the brand, that is not related In order to provide a comparison of the considered
to the user’s textual information. Therefore, followers algorithms, a small set of 20 tweets for each tag has been
can be used as true positives. However, the same cannot manually verified to understand which approach has
rebe done for the true negative samples, due to the fact turned the right polarity value. Cumulatively, the best
that negative relationships independent from the texts performing algorithm is Vader [
Before going through the description of results, we present in Section 4.1 a benchmarking analysis we have 4.2. Synthetic Dataset carried out, in order to decide the specific sentiment analysis algorithm to adopt for our experiments. of the considered brands. Most tweets contain text and embed URLs, pictures, usernames, and emoticons. Therefore a pre-processing of tweets has been performed, such that tweets are filtered, and incomplete/inconsistent data eliminated. In more detail, each tweet has been suitably cleaned, by removing: • URLs; • tagged users names; • special characters.
Moreover, all emoji symbols have been translated into text, and language contractions into their extended forms.
Figure 3: ROC curve for the synthetic dataset. Hashtags have been extended in sentences. In order to optimize the sentiment analysis process, the word lemmatization with part of speech has been applied for all extracted tweets. It is worth to point out that, in the pre-processing described above, stop words have not been removed, due to the fact that they can be important to establish the word polarity. Indeed, the sentiment analysis software libraries usually take them into account for this reason.
Tweets have been preprocessed also to eliminate
duplicate tweets and retweets from the data, which led to a
ifnal sample of 12, 585 tweets and 498 users that can be
considered significative (i.e., they have a suficient
number of tweets associated to perform the analysis).
SenFigure 4: ROC curve for the synthetic dataset with noise. timent analysis has been carried out using the Python
VADER library. The VADER Sentiment Analyzer [
Then, in order to verify also the robustness of the • negative, with score’s value 0; method, a small percentage of noise has been introduced • neutral, with score’s value −1. by exchanging the 20% of sentences between the two types of users in the dataset. Figure 4 shows the ROC curve obtained in this case, with an AUC equal to 0.84. 4.3. Real Data Validation The dataset coming from real data associated with input brands, and related targets, used for the validation is shown in Table 3. In particular, tags have been obtained from the keywords of the brands web pages.
As for the users’ opinions, they have been generated as
follows. The data of tweets for each user have been
downloaded from Twitter by the Twitter APIs [
For the purposes of our research, the values returned by
VADER have been normalized in the range [
Results obtained on real data shows the ability of the proposed approach in finding the best targets for advertising campaigns. Indeed, the 93.7% of efective followers have been correctly put in the first positions of the ranks, cumulatively. This shows that users who have provided preferences for some brands, as testified by the fact that they follow them, have scored high values of the Compatibility Index, with references to such brands, as expected.
The K-NN analysis has been performed by choosing three brands for which the tags are diferent and the resulting classes are thus well separated. The chosen brands are Adidas, Barilla and Microsoft. Only the followers of such brands have been put into the three classes.
The user-vectors have been built from the polarities obtained, for each follower in the class, for the tags of the corresponding brand. Then, the K-NN has been applied by leaving out the 20% of users for each class (test set) and keeping inside only the 80% of them (training set).
The best resulting accuracy performed by the proposed approach is of 85.71% test users correctly classified, obtained for = 6 .
We have presented a novel approach for the recommendation of new possible customers for existing brands. The approach is based on the assumption that textual information is stored in a database on both users and brands. Sentiment analysis techniques have been adopted to measure to what extent an user matches a brand, according to the retrieved users’ opinions on diferent topics. Two diferent formulations of the proposed approach have been described, and the preliminary results obtained on both synthetic and real datasets are promising, looking at the high values of accuracy reached in both cases.
In the future, we plan to finalize the implementation of all modules of the general platform illustrated in Figure 1. Moreover, further work will regard the consideration of reciproque compatibility measures, in order to predict, for example, if users with high matches with a specific brand may also likely have high compatibility with other brands.
This research has been partially supported by the projects: “AMABILE - Amarelli BIg data and bLockchain Enterprise platform” (CUP: B76G20000880005) funded by the Italian Ministry of Economic Development, and “Big knowledge graphs modelling and analysis for problem solving in the web and biological contexts” (2022, CUP: E55F22000270001), funded by INDAM GNCS. 1–7