Advancing Influencer and Sentiment Analysis on Social Networks: An introduction to IHPC s Strategic Social Systems Programme Presented by: Boon Kiat Quek, Ph.D. Scientist, Computational Social Cognition & Technical Lead, Strategic Social Systems Institute of High Performance Computing, A*STAR quekbk@ihpc.a-star.edu.sg Slide 1
Strategic Social Systems (SSS) About Us A new research programme in IHPC focusing on Translational Social Sciences R&D with the aid of advanced data analysis and computational technologies to promote sustainable business, social and urban success. People An interdisciplinary team of 15 scientists and collaborators with experience in social psychology, cognitive science, artificial intelligence, behavioral science, economics, data mining, computer engineering, and information systems. Partners SSS is currently seeking & supporting partners in ICT, financial services, transportation, and goverment agencies. We continue to seek motivated collaborators for forging strategic partnerships Slide 2
Strategic Social Systems (SSS) SMILE Business Domain Social Marketing InteLligence Enhancement Consumer facing and serviceoriented companies: banks, fashion, food & beverages, personal care SPICE Public Domain Strategic Public Information and Communication Enhancement Public facing institutions: MNCs, IHLs, government agencies SHINE Healthcare Domain Social Healthcare INformation Enhancement Health promotion agencies, infectious disease control centers, hospitals Key Enabling Technologies: Psychographic analysis (advanced profiling and segmentation) Social media analysis and monitoring (influencer tracking, fine-grained sentiment analysis, misinformation detection, opinion mining) Communication strategy recommendation and decision-support SSS is currently engaging partners in ICT, financial services, transportation, as well as govt agencies. Slide 3
IHPC s Strategic Social Systems Team Social and Behavioral Sciences Ilya Farber Cognitive Science and Philosophy Advanced Data Analysis Sebastian Feller Linguistics Noraini Rahman B.A., Sociology /Psychology Richard Shang Information Systems Yang Yinping Information Systems Computational & Software Technologies Wang Zhaoxia Computer Science Kayo Sakamoto Engineering (Cognitive Science) Quek Boon Kiat Integrative Sciences and Engineering Jerry Ping Information Systems (expected) Martin Saerbeck Industrial Design Rick Goh Electrical and Computer Engineering Lu Sifei MSc., Computer Architecture *ED, Dy ED IHPC are endorsing directors in our key initiatives. Potential contributions of health domain researchers Su Yi, Xiuju also discussed. Slide 4
Strategic Social Systems: SPICE Sentiment & Insight Analysis* Analysis & Identification of Key Influencers* 500 450 400 350 300 250 200 150 100 50 0 12 AM 2 AM 4 AM 6 AM 8 AM 10 AM 12 PM 2 PM 4 PM 6 PM 8 PM 10 PM Affective Pos Aff Neg Aff Anxiety Anger Sadness Insightful All Sentiment Dimensions Constructive feedback SPICE Team (IHPC) *work in progress Slide 5
Capability Development (SPICE) Social media monitoring and social network analysis Identifying key influencers and opinion leaders Identifying main topics, issues, and opinions Identifying purpose-specific comments (e.g., feedback) Fine-grained sentiment analysis Psychographic analysis of users Making inferences about future trends vis-à-vis the above Communication strategy recommendation Identifying users into specific groups with different goals Messages tailored to achieve goals specific to each group Slide 6
Key influencers: Why should we care? Influencers and Opinion Leaders Individuals who have the power to affect purchase decisions of others because of their (real or perceived) authority, knowledge, position, or relationship. ~ businessdictionary.com Ability to reach large segments of the online community Opinions & sentiments of influencers spread easily to others Could span across multiple social network services Identifying key influencers as a means of: Managing spread of information (including misinformation) Improving marketing of products and services Recruiting early adopters Slide 7
Key Influencers: How to identify? Step 1: Generating networks from social relationships Inferring relationships from user generated content Twitter: mentions, replies, retweets Facebook: friendship links, affiliations, interests, Likes Inferring influence as a function of these relationships Advantages: availability of network analysis methods for characterising relationships between and across nodes @miyagi @STcom Mentioned by @mrbrown @... Slide 8
Key Influencers: How to identify? Step 2: Locating key influencers via various metrics Ranking nodes based on centrality Degree - based on the number of neighbours Closeness - based on how easily a node can reach other nodes Betweeness - based on extent to which a given node lies along shortest paths between two other nodes Eigenvector score of a node based on connections to high scoring nodes e.g., PageRank (Google Inc.), Alpha Centrality (Bonacich & Lloyd, 2001) Ranking nodes based on other factors or network properties E.g, message frequency, nature of messages, psychographic features, past behavior patterns Slide 9
Visualizing Influencer Analysis Influential users play critical roles in the propagation of information throughout a social network. Identification of key influencers as a step towards better community engagement. Current influence based on twitter content collected for a given topic (blue links) Potential influence based on twitter users recent posts on twitter; to uncover latent channels through which information could be propagated. (red links) Slide 10
Potential / Latent Influence Identify latent communication channels through which information could be propagated For a given topic, some links might not have been uncovered if no actions were performed upon message receipt. But could well be implicated on a different topic in future. Thus, influence of some users could potentially be higher even though for the current topic their influence is low Topic A Topic B Dennis Rodman visits N.Korea Slide 11
Ongoing Research Sentiment Analysis Beyond simple valence polarity (e.g., positive, negative, neutral) Grounded in psychological theories of emotions Integrating influencer analysis with sentiment analysis Why? Knowing which influencers evoke which kinds of sentiment/emotions allows different actions to be tailored: Misinformation management -- getting to the source, and mitigating viral flow Recruiting early adopters of products and services Open ended issues with potentially new insights Influence x Sentiment x Topics x Opinions x Time Slide 12
Opportunities for Collaboration Human behavior analysis and modeling from multiple sources of data using computational and social sciences know-how: Consumer psychographic analysis Characterizing consumer lifestyles, personalities Deriving consumer insights for product design, targeted marketing, and personalized consumer engagement Brand-centric social media analysis Monitoring, tracking and analyzing sentiment and key influencers and opinion leaders on social media Uncovering consumers attitudes towards products and services for enhancing social media marketing and brand management Slide 13
Thank You For enquiries, please contact: Boon Kiat Quek quekbk@ihpc.a-star.edu.sg
Supplementary Information
Our approach Approaches attempted Degree centrality (InDegree, OutDegree, or Both) Weighted degree centrality Weighted PageRank Adaptation of PageRank Takes number/frequency of tweets into account Power iteration Time instance Weight of outbound link from i to j WPR of node j x ( k i + 1) = (1 α) x i 1 ( k) + α W i j w j, i ( k) x j InDeg ( k) j WPR of node i Smoothing constant Weights of all outbound links In-degree of node j Slide 16
Network Generation Starting with a Twitter query in mind, e.g., MRT Gather tweets from Twitter s API For each tweet T, Node1 T.senderName If T.text contains @, RT @ Node2 twitter handler from @ Network.addNode(Node1) Network.addNode(Node2) Network.addLink(Node1, Node2) Slide 17
Future explorations To look into alternative methods for assessing influence TwitterRank (Weng et al., 2010) k-shell decomposition (Kitsak et al., 2010) Enrich influencer analysis with other factors e.g., individual s sentiment, frequency of tweeting, number of followers, etc. Slide 18
Some Issues and Questions Issues and Questions Other metrics for assessing influence Evaluation of the above metrics Any other important but overlooked issue to be analyzed Finding the right research questions to address vs. applying available methods to meet needs? i.e., implement known methods such as TwitterRank instead of trying to come up with new ones Slide 19