help

A Sentiment and Interest Based Approach for Product Recommendation

Vibhu Jawa Applied Mathematics Department Delhi Technological University

Delhi, India [email protected]

Varun Hasija Applied Mathematics Department Delhi Technological University

Delhi, India [email protected]

Abstract - The growing popularity of social networks has led to abundant availability of user sentiments, making them a crucial factor in buying decisions, public opinions, and brand reputations. Rise of real-time web has provided an opportunity to utilise time-sensitive data which is available in the form of tweets on public and private Twitter streams, as the basis for product recommendation based on Sentiment analysis. In this paper, we propose a model working on the basics of Interest graph in conjunction with Sentiment analysis to compute the correlation between different entities and provide recommendations, which range from whom to follow on Twitter to what to buy online.

Keywords - Twitter, Interest Graph, Recommendations, Algorithm, Sentiment Analysis.

I. INTRODUCTION The vast amount of user generated content in social media applications and its high rate of generation has given us the opportunity to build real-time recommendation systems. Twitter provides a medium to share our ideas and experiences with the whole world instantaneously. It acts as a channel to satisfy our curiosity, thus providing for the root need of every individual to be heard as well as a sense of worth and importance. Twitter acts as an easy and a fast medium which matches human beings at their speed of their thought [1]. The growing popularity of social networks has led to abundant availability of user sentiments, making them a crucial factor in buying decisions, public opinions, and brand reputations. “Internet-Scale Interest Graph” is a representation of the relationship between people and their interests. Having that knowledge allows us to tie two remote entities with no obvious common properties and show possible, but yet unexplored connections, thus discovering various relationships and interconnections between user groups and their interests [3]. Previous attempts on providing recommendations have been based on Interest Graphs or Sentiment analysis. The main idea behind this work is to explore the conjunction of Interest graphs and Sentiment analysis, to compute the correlations and provide recommendations, which range from whom to follow on Twitter to what to buy online.

II. LITERATURE REVIEW

A. Interest Graph Pankaj Gupta et al. [21] at Twitter, presented a service called WTF (WHO TO FOLLOW) in which they explored the implications of usage of interest graphs in their user recommendation algorithm. Other researchers have also shown affinity towards Interest Graphs over Social Graphs. “The Interest Graph is an online representation of individuals’ interests, with people and interests being the nodes of the graph. Connections exist between people and interests, as well as, between interests and derived interests. The strength of the connection depends on the level of interest or relatedness between concepts and can change over time. “[4] Interest graphs find its roots in the semantic web which is defined by Tim Berners-Lee as "web of data that can be processed directly and indirectly by machines”. Semantic Web defines the relationships between people and things, where relationships are represented as “triples” [11]. Conceptually, triples can be thought of as: “PERSON-RELATIONSHIP-THING” Such as Bob likes Moto X. There are two approaches for data collection to create interest graphs [4]: Explicit: The user provides the information. Implicit: The user’s activity is followed and the interests are derived through it. Explicit approach may create hindrances for the users and might make them uncomfortable about providing detailed information about their interests. Users may also exercise the option to search the required product or service directly. An example of explicit approach can be Pinterest. Implicit approach is best explained with the help of an example- Amazon uses implicit approach for data collection. The site collects information about one’s purchase history and search history and computes co-relations (the people who bought this also bought that or the people who searched this item also looked for that item). The bottom line is to collect information about the user's interests, irrespective of the approach followed for data collection. It has been observed that explicit approach tends to generate inaccurate profile of users. Hence, it is suggested

2015 17th UKSIM-AMSS International Conference on Modelling and Simulation

978-1-4799-8713-9/15 $31.00 © 2015 IEEE

DOI 10.1109/UKSim.2015.26

75

to use the implicit approach and collect data according to real interests of the users. Next, we find the relationship that exists between various user groups and their interests. For example, we could have a group of travel enthusiasts amongst our customers who could be interested in “camping”, “mountaineering”, and “hiking” etcetera. All the activities that they are interested in could have a common related interest of “Outdoor Sports”. Having that knowledge allows us to tie two remote entities with no obvious common properties and show possible, but yet unexplored connections. Visualization of the interest graph is a great way to discover various relationships and interconnections between user groups and interests.

Figure 1: An example of interest graph

B. Sentiment Analysis Davide Feltoni Gurini et al. [10] presented a technique based on a weighting function which takes into account users sentiments and then recommends them users based on it. Singh et al. [20] implemented a recommender system that incorporates sentiment classification in the case of movie recommendation. Bing Liu defined Sentimental analysis (or opinion mining) as the study of sentiments and opinions about an entity conveyed through text [2]. Natural language processing, text analysis, and computational linguistics are being used by researchers to identify and extract relevant information in source materials. In layman terms, the purpose of sentiment analysis is to classify the product as “good or bad”. Based on level of granularities, sentiment analysis can be investigated at three levels [2]:

1. Document level: At this level, the task is to identify whether a “document” expresses a positive or a negative opinion. This type of analysis assumes that each document expresses opinions on a single entity.

2. Sentence level: At this level, the task is to identify whether each sentence expresses a positive,

a negative or a neutral opinion. It can be thought of as a two-step process:

• Subjectivity classification • Sentiment classification

3. Entity level/ Aspect level/ Feature level: Document level and sentence level does not tell us about WHAT people liked and WHAT they did not. Entity level is based on the fact that every opinion consists of a sentiment and a target. Hence, it performs a fine-grained analysis.

Opinions can also be divided further into regular opinions and comparative opinions. A regular opinion expresses sentiments towards a particular entity, whereas a comparative opinion compares different entities based on the attributes shared by them. [19] Sentiment Lexicon: The most important indicators of sentiments are opinion words. These are the words which can be used to express positive or negative sentiments. For example, “good”, “awesome”, “amazing” are words expressing positive sentiments, whereas “bad”, “terrible”, “disgusting” express negative sentiments. Formally, opinion is defined as a Quadruple, [2] , , , , where,

• g is the opinion or sentiment target, • s is the sentiment about the target, • h is the opinion holder, • t is the time when opinion was expressed.

Objective of sentiment analysis based on the above given definition can be stated as: Given an opinion document d, discover all opinion quadruples in d and analyze those opinions to classify them as positive or negative.

III. RESEARCH LEADING TO DEVELOPMENT “Tweets are the essence of Twitter, and while they are notionally thought of as the 140 characters of text content associated with a user’s status update, there’s really quite a bit more metadata there than what meets the eye.” – Matthew Russell [5]. Tweets, in addition to textual content, come bundled with two subsidiary pieces of metadata i.e. entities and places. “Entities” consist of the user mentions, URLs, hashtags and media that might be associated with the tweet and “Place” is the location in real world that is attached to the tweet. We can easily differentiate or find similarities between the most trending topic in different parts of the world using the Yahoo! GeoPlanet’s WOE – Where on Earth, ID system [6]. Word frequency can be defined as the number of occurrences of a word per unit tweet. We can derive positive or negative sentiments from the word frequency in tweets, which gives us an insight about the public opinion of the entity being observed.

76

For example, consider the word frequ obtained when searched about a topic like “

Figure 2: Word Frequency graph for "SO

Figure 3: Sentiment Score graph for "SO

These graphs could suggest us which pr compared on the basis of the public sen from the tweets. Sentiment score is positiv entity has a positive sentimental value. Now, we take the help of interest graph analysis to recommend which product is counterparts or to rank different products f world. For example, we take two entities and c the basis of sentiment analysis of frequ words. Let the 1st and 2nd entities be “SONY respectively. On plotting the word frequency graph containing “SONY” as an entity we recogni

1. PS4 is the most popular product, 2. Tokyo is the “Place” most popul

Some words of sentimental value such as “unsatisfactory” etcetera are also observe plotting the word frequency graph of tw “Samsung” as an entity, we recorded similar

uency of words SONY”

ONY"

ONY"

roduct is better if ntiment procured ve means that the

hs and sentiment s better than its from all over the

compare them on uently occurring Y” & “Samsung”

hs of the tweets ize that: , larly linked with, “great”, “good”,

ed. Similarly, on weets containing r results.

From these examples, we infer tweets contains some important inf the kind of entity we are observing:

1. If the entity is an frequency table will con about its products.

2. If the entity is a produ table will contain some inf organization, some in counterparts and some info of the product.

3. Information about rand which the company is invo hack”.

Using interest graphs and senti provide a method that can be us certain entity or comparing differ weighted ranking system (ranking e 1).

IV. PROPOSED M We propose a model of coagul and Sentiment analysis for product r Our model can be divided into tw 1. Word processing and related en Before performing any kind of analy pre-processed in which, the redunda tags, tweet header and references to removed. [16] Our algorithm creates a data fra information for each tweet. Data fr format of the information related wi For creating interest graphs, a dat the data frame and the row sum of i.e. the number of occurrences of descending order of frequency. extracted from this list and bound to with frequencies - called cloud fra separate columns. Thus, the secondary entities i relation with our primary entity are In the interest graphs displayed connection of interest is depicted size.

that, word frequency of formation with respect to

organization, the word ntain some information

uct, the word frequency formation about its parent nformation about its ormation about the nature

dom and latest trends in olved. Ex. “SONY server

iment analysis, we now ed for recommending a rent entities based on a entities on a scale of -1 to

MODEL ation of Interest Graphs recommendation.

wo steps: ntity identification: ysis, the tweets were ant spaces, URLS, hash other screen names are

ame, which contains the rame provides a readable ith mined tweets. ta matrix is formed using

f each entity is calculated f an entity and sorted in Then the ‘entities’ are ogether into a data frame ame [15] with names in

in descending order of obtained.

d below, the strength of according to their font

77

Figure 4: Interest Graph for "Samsung"

Figure 5: Interest Graph for "SONY"

Once the secondary entities are derived we calculate the sentimental value associated with each entity using the following described approach. 2. Sentiment analysis of identified entities: In our model, once the tweets are mined and cleaned, a vector of sentences is generated and “Plyr” [8] is used to handle the list. Sentences are cleaned up using R's regex-driven global substitute, “gsub” and then, converted into lower-case words. Then, the words are compared to the dictionaries of positive & negative terms using match function and the total score is calculated by subtracting the total number of positive matches from negative matches Sentiment Scoring Scheme: Breen's algorithm is used for calculating the sentiment score for each entity [7].

• If Score > 0, we infer that the sentence has a net 'positive opinion'.

• If Score < 0, we infer that the sentence has a net 'negative opinion'

• If Score = 0, we infer that the sentence has a net a 'neutral opinion'.

We use the opinion lexicon in English for counting the positive and negative words provided by Hu and Liu. [9] After analyzing the sentiments, its graph is plotted. We depict the sentiments of 100 tweets in sets of four i.e. each data point shown in Figure: 6 represents the sentiment score for four tweets.

Figure 6: Sentiment Score graph for "Samsung" & "SONY"

After obtaining the sentiment score for the primary entity, our algorithm obtains the same for all the secondary entities obtained from the interest graph. Then, our algorithm calculates the final score by inculcating both these scores using the formula given below: Final Score 0.75 0.25 1

Where,

• Sentiment primary is the Sentiment Score of the primary entity

• Sentiment k is the Sentiment Score of the k th entity

• Frequency k is the number of occurrences of the k th

entity • Total Occurrence is the sum of all the occurrences

of the secondary entities. The resulting sentiment can be positive, negative or neutral. We can compare this score for multiple products and can recommend a product with the maximum positive social sentiment. The following table depicts the differences between the average sentiment score of “Primary Entity” and the average “Commutative sentiment score” calculated by mining 30000 tweets each of Sony and Samsung.

78

TABLE 1. SENTIMENT SCORE FOR "SONY Entity Primary Entity

Sentiment Score Comm Sentim

SONY 0.56 Samsung 0.59 The commutative score of entities calcul algorithm show that the impact of the deriv the interest graph of the primary entity play forming the public opinion of the produc ignored. Thus, our results reinforce the integrating the techniques of sentiment anal Graph we get a greater perception into the p the product. The algorithm for creating interest grap below:

The algorithm for sentiment analysis is su below:

Graph Visualization

Entitiy Identification Data

Framing Data Martix Construction

S o

Word Processing

Redundant spaces removed

Retweet Header Removed

R o

na

Final Score=Positive matc Negative matches

Scoring Matching positive and negative words in

lexicon

Data pre-processing Lowercase Transformation Spl

" & "Samsung" mutative ment Score

0.61 0.42

lated through our ved entities from ys a huge role in

ct and cannot be notion that by

lysis and Interest public opinion of

h is summarized

ummarized

V. CONCLUSION AND F We presented a novel recomm recommends products by acquirin about the product and its derived en Thus, by integrating the techniqu and Interest Graph we get a greate opinion of the product. This “one parent” approach ca derived entities for which we can d all the derived entities of the par recursive approach can be used to d correlated webs of interest grap extensive perception about product’ An alternative to allotting fixed entity can be allotting variabl representative of real world relatio These variable weights would be ba which each derived entity has on its

ACKNOWLEDGEM We would like to thank Twitter I to their service. We would also lik Russell for authoring “Mining t inspired us to work on this project.

REFERENCE [1] Issie Lapowsky (4 October 2013)

Early Years". Inc. Inc. Retrieved 5 Oc [2] Bing Liu. Sentiment Analysis and

Claypool Publishers, May 2012 [3] How the Interest Graph will shape th

Web, Mobile & Hardware ; April 1 20 [4] Cygnismedia blog “what-is-interest-gr [5] Matthew A. Russell, Mining the Soc

Facebook, Twitter, LinkedIn, and Oth Media, Inc., 2011

[6] "Yahoo! WOEID Lookup". http://Zou [7] Gary Miner (Author), John Elder IV (

Robert Nisbet (Author), Practical Analysis for Non-structured Text Dat 0123869791

[8] Hadley Wickham (2011). The Split Data Analysis. Journal of Statistica http://www.jstatsoft.org/v40/i01/.

[9] Hu and Liu, Opinion Lexicon, KDD-2 [10] Davide Feltoni Gurini, Fabio Gaspa

Giuseppe Sansonetti “A Sentiment-B Recommendation”.

[11] Berners-Lee, Tim (May 1, 2001). "T American

[12] Hang Li ,Learning to Rank for Info Language Processing, Second Edition

[13] F. Burkhardt, C. Becker-Asano, E. B Gebhard, A. Kazemzadeh, I. Steiner EmotionML. Proceedings of the 5t Emotion, Sentiment, Social Signa (ES3LOD), 2014.

[14] Paul Buitelaar, Mihael Arcan, Car Sánchez, Carlo Strapparava (2013) Sentiment Analysis 2nd Workshop o (LDL 2013): Representing and linkin other language data. Collocated with Approaches to the Lexicon, Pisa, Italy

Sorting based on frequency

Reference to other screen ame removed

ches-

n an opinion

litting

FUTURE WORK mendation system, which ng the public sentiment ntities ues of sentiment analysis er insight into the public

an be mirrored for the develop interest graphs of rent interest graph. This develop deeper and more phs giving us a more ’s social sentiment. weights to each derived

le weights which are ns to the primary entity. ased on the impact factor s parent entity.

MENT Inc. for providing an API ke to thank Matthew A. the social web” which

ES . "Ev Williams on Twitter's ctober 2013.

Opinion Mining, Morgan &

he future of the web, By rr11 in 011 raph” ial Web: Analyzing Data from

her Social Media Sites, O'Reilly

urbuth Project (Author), Thomas Hill (Author),

Text Mining and Statistical ta Applications. ISBN-13: 978-

t-Apply-Combine Strategy for l Software, 40(1), 1-29. URL

2004 aretti,Alessandro Micarelli and ased Approach to Twitter User

The Semantic Web". Scientific

ormation Retrieval and Natural nOctober 2014 Begoli, R. Cowie, G. Fobe, P. , T. Llewellyn: Application of th International Workshop on als and Linked Open Data

rlos A. Iglesias, J. Fernando ) Linguistic Linked Data for on Linked Data in Linguistics ng lexicons, terminologies and the Conference on Generative

y

79

[15] Kapanipathi, P., Jain, P., Venkataramani, C., & Sheth, A. P. (2014). Hierarchical Interest Graph from Tweets. Proceedings of the Companion Publication of the 23rd International World Wide Web Conference, 311-312.

[16] B. Sriram, D. Fuhry, E. Demir, H. Ferhatosmanoglu, and M. Demirbas. Short text classification in twitter to improve information filtering. SIGIR ’10.

[17] F. Orlandi, J. Breslin, and A. Passant. Aggregated, interoperable and multi-domain user profiles for the social web. I-SEMANTICS ’12

[18] Thet, Tun Thura, Jin-Cheon Na, Christopher S. G. Khoo, and Subbaraj Shakthikumar. 2009. Sentiment analysis of movie reviews on discussion boards using a linguistic approach. In Proceedings of the First International CIKM Workshop on Topic-Sentiment Analysis for Mass Opinion Measurement, pages 81–84, Hong Kong.

[19] Wilson, Theresa, Janyce Wiebe, and Rebecca Hwa. 2004. Just how mad are you? Finding strong and weak opinion clauses. In AAAI 2004, pages 761–767, San Jose, CA.

[20] V. K. Singh, M. Mukherjee, and G. K. Mehta. Combining collaborative filtering and sentiment classification for improved movie recommendations. In Proceedings of the 5th International Conference on Multi-Disciplinary Trends in Artificial Intelligence, MIWAI'11, pages 38{50, Berlin, Heidelberg, 2011. Springer-Verlag.

[21] P. Gupta, A. Goel, J. Lin, et al., “WTF: the who to follow service at twitter,” in Proceedings of the 22nd International World Wide Web Conference Committee (WWW '13), pp. 505–514, 2013.+

80