​

​

Big Data Security and Privacy

​

InXite: A Social Media Analytics System

​

​

Dr. Bhavani Thuraisingham

The University of Texas at Dallas

​

Guest Lecture

​

Lecture #5-3

​

August – November 2022

1

* *

Outline

• Overview of InXite
• Modules of InXite
• Cloud-based InXite
• Other InXite Applications
• Summary and Conclusion
• Future: InXite for COVID-19 and Privacy Considerations

​

• Reference: Satyen Abrol, et al Real-Time Stream Data Analytics for Multi-purpose Social Media Applications. IEEE IRI 2015: 25-30

​

​

2

* *

Overview of InXite

• We describe a cloud-based system called InXite, also called InXite-Security (Stream-based Data Analytics for Threat Detection and Prediction), that is designed to detect evolving patterns and trends in streaming data.
• InXite integrates information from a variety of online social media sites such as Twitter, Foursquare, Google+ and LinkedIn, builds people profiles through correlation, aggregation and analyses in order to identify persons of interest who pose a threat.
• Other applications include garnering user feedback on a company’s products, providing inexpensive targeted advertising, and monitoring the spread of an epidemic, among others.
• InXite is designed to detect evolving patterns and trends in streaming data including emails, blogs, sensor data and social media data such as tweets.
• InXite is designed on top of two powerful data mining/machine learning systems, namely the patented Tweethood system (location extraction for Tweets), with the explicit aim of detecting and predicting suspicious events and people and the patented SNOD system (Stream-based Novel Class Detection).
• InXite comprises four major modules: InXite Information Engine, InXite Profile Generator, InXite Psychosocial Analyzer, and InXite Threat Evaluator and Predictor

​

​

​

3

* *

Overview of InXite

• Our goal is to decipher and monitor topics in data streams as well as to detect when trends emerge.
• This includes general changes in topics such as sports or politics and also includes new, quickly-emerging trends such as hurricanes and bombings.
• The problem of correctly associating data streams (e.g., Tweet messages) with trends and topics is a challenging one.
• The challenge is best addressed with a streaming model due to the continuous and large volume of incoming messages.
• It should be noted InXite is a general purpose system that can be adapted for a variety of applications including security, marketing, law enforcement, healthcare, emergency response and finance.
• We mainly focuses on the adaptation of InXite for security applications which we call InXite-Security.
• Other adaptations of InXite include InXite-Marketing, InXite-Law, InXite-Healthcare, InXite-Emergency, and InXite-Finance.
• That is, while InXite-Security is developed mainly for counter-terrorism and Intelligence applications, most of the features can be tailored for marketing and law enforcement applications (among others) with some effort.

​

4

* *

InXite Architecture and Modules

5

* *

Modules of InXite: Information Engine

6

* *

Modules of InXite

• InXite gets the data from various social media sites including streaming data from Twitter.
• The information is integrated by the Information Engine and carry out various functions such as entity resolution and ontology alignment.
• Then the Information Analytics Engine will analyze the integrated data using data mining/machine learning techniques; the results are then given to the analyst.
• The major modules of InXite are common to our applications such as Security, Marketing and Law Enforcement; each application also has a small number of tailored components.
• InXite also follows a plug and play approach; that is, the analyst can plug his/her components if he/she has a preference for various tasks.
• At the heart of InXite is its Information Engine whose details are as follows:
• Information Engine: The first step is to extract concepts and relationships form the vast amount of data streams and categorize the messages; Then we provide semantic representation of the knowledge buried in the streams.
• This would enable an analyst to interact more directly with the hidden knowledge. Second step is to represent the concepts as ontologies, and subsequently integrate and align the ontologies.
• Once the multiple graphs extracted from the streams are integrated, then our analytics tools will analyze the graphs and subsequently predict threats.

​

​

​

7

* *

Modules of InXite

• Once the multiple graphs extracted from the streams are integrated, then our analytics tools will analyze the graphs and subsequently predict threats.
• The information engine module integrates the attributes of a user from multiple data streams including social networks (e.g., Twitter, LinkedIn, Foursquare, etc.) and performs entity resolution, ontology alignment, conflict resolution, data provenance, and reasoning under uncertain and incomplete information.
• At the heart of the Information Engine is Tweethood, a novel, patented system/algorithm to determine user attributes including, but not limited to, location, age, age group, race, ethnicity, threat, languages spoken, religion, economic status, education level, gender, hobbies or interests, based on the attribute values for the friends of the user.
• While entity resolution algorithms have been around since the mid-1990s, InXite uses a combination of content-based similarity matching and friends-based similarity matching algorithms. The Information Engine module consists of two major components: Entity Extraction and Information Integration.
• Entity Extraction – The process of extracting (mining) and/or, in certain cases, predicting user specific attributes which include demographic information, such as, age, gender, etc. and information about his/her social networks, such as friends, followers, people he/she is following, etc.
• Information Integration - The process of integrating or joining two or more user profiles from the same or different sources, such as, social networks, blogs, etc. This is done using the information obtained from the previous step.

​

​

8

* *

Modules of InXite

• Entity Extraction
• For data integration to take place properly, a certain number of content- and friend-based similarities must be found between two entities.
• Entity extraction is the first step towards that goal. Upon integrating all the known aspects of a user’s profile, additional information is added through text mining and our novel algorithms like Tweethood.
• All popular social networking websites and other data sources that provide data in either structured or semi-structured format are mined.
• Then using text mining techniques, various entities associated with the persons of interest for us are extracted.
• Next, in our iterative process, the structured data is parsed using a simple crawler and we obtain <key, value> pairs for each profile from the information, where a key is a user attribute, such as age, gender, etc. and a value is the corresponding value obtained from the profile.
• After this is done, we use our content-based and friend based (Tweethood) algorithms for prediction of attributes for which no values have been found.
• The pseudocode for Entity Extraction is as follows:
• Using Text Mining techniques in the literature, extract relevant information about the user from the multiple sources of data (including social networks, databases, etc.)

​

9

* *

Modules of InXite

• Organize the information into (key, value) pairs
• Use Tweethood to predict values that are unknown for any keys
• Information Integration
• In this step, we integrate or join two or more profiles from different data sources, which belong to the same individual based on matching of the attribute <key, value> pairs we obtained in the previous step.
• e pick one profile from data source A and try to find the closest match for it in data source B, by forming pairs and assigning them scores. This score is assigned based on the similarity of the two profiles, determined by the proximity of different user attributes in the two.
• If this score crosses a pre-determined threshold and is also the highest for that chosen profile from data source A, we link the two indicating that the two profiles point to the same user.
• Finding a partial verification of entities and friends can be a difficult process.
• The amount of information similarities needed to make a conclusive match is constantly changing.

10

* *

Modules of InXite

• These decisions therefore need to also be made from constantly changing ontologies. InXite constructs ontologies for partial entity resolution dynamically by observing patterns in complete resolutions.
• We define an ontology for each data source, such as an online social network, blog, etc.; these ontologies are then linked so that the system understands that an attribute key A, such as gender, from one data source points to the same thing as attribute B, such as sex, from another data source.
• This linkage of ontology structures constructed from different data sources is essential for the integration/disambiguation of two or more profiles. 
• The pseudocode for Information Integration is as follows:
• Construct ontologies for the entities extracted using various ontology construction techniques in the literature
• Carry out entity resolutions by determining whether two entities are the assigning scores as to how similar the entities are.
• Apply data mining/machine learning techniques to observe patterns in the entity resolution process
• For those entities that cannot be completely resolve, use the patterns observed in step 3 and resolve the entities
• Link the various ontologies constructed using the results from the entity resolution process to form a linkage of ontologies which are essentially person of interest (that is, user) profiles.

11

* *

Modules of InXite: Person of Interest (POI) Analysis

12

* *

Modules of InXite

•  Person of Interest (POI) Analysis
• Once the information engine integrates the multiple data sources, extracts entities, construct ontologies, resolves conflicts, determines similarities and links the ontologies, the next step is to analyze a person of interest based on various attributes and algorithms.
• InXite Person of Interest Profile Generation and Analysis:
• A generated profile represents one or more aggregated entities from the extraction step.
• If two profiles are determined to belong to the same person at any point before or after profile generation, then the attributes and data of the two are merged into a single profile.
• This may happen because of ontology shifts during analysis or manual discovery by an analyst.
• Even though attribute prediction happens during the entity extraction and alignment steps, SNOD is continuously used in several modules to detect novel information nodes as long as information is added or discovered in the searching process.
• This means that profiles are constantly edited, updated, and merged after profile generation.

​

13

* *

Modules of InXite: Psychosocial Analysis

14

* *

Modules of InXite: Threat Analysis

15

* *

Modules of InXite

• InXite POI Threat Analysis:
• The InXite POI Threat Analysis is a core InXite feature which combines several individual scores, some based on existing studies and some novel, to come up with a final score for evaluating the seriousness of as a potential threat.
• Each of the individual scores and the final score has a range from 0 to 100 with 0 meaning a low threat and 100 meaning a high threat. Below listed are the seven major components of the Threat Evaluation Module that contribute to the final threat evaluation score
• (i) Demographics-based score computation
• InXite predicts and aggregates user related attributes such as age, location, religion, etc. Using existing studies done on terrorists such as the one by Marc Sageman, InXite determines if the POI fits the profile of a terrorist or not.
• E.g. If (age between 22 and 35) AND (education = college) AND (ethnicity = XYZ) 🡪 High Score
• (ii) Psychological score computation
• For calculating the psychological scores, we design algorithms that analyze the language used by the POI; in particular we look at the adjectives and the nouns, to come up with five personality traits – Sociability, Evaluation, Negativity, Self-Acceptance, Fitting-in, Psychological Stability and Maturity. Based on these scores, we derive a final psychological score
• E,g, Sociability AND Negativity AND Psychological Instability AND Fitting In 🡪 High Score

​

16

* *

Modules of InXite

• Content-based score computation
• An important source of information about a user is the messages/posts that he/she puts on his/her page. In order to calculate the content based threat score for a user, we define a rule-based system that looks for suspicious nouns and verbs, analyzes their relationship and assigns a score based on that.
• E.g. I want to bomb the TallBulding 🡪 I/PRP want/VBP to/TO bomb/ VB the/DT TallBuilding/NN 🡪 High Score.
• This metric is particularly useful for identifying mal-intent users who are expressive about their intents.
• Background check score computation
• For individuals located in the US, we run background checks using existing software/websites. The integration and prediction of user attributes helps in successful user disambiguation and allows us to do an advanced search of the database.
• Based on the previous crimes committed by the individual, we assign a score which reflects the likelihood of him/her being a threat in the near future.
• E.g. If criminal and Type_of_Crime = Violent or Federal 🡪 High Score

​

17

* *

Modules of InXite

• (v) Online reputation-based score computation
• InXite analyzes various online data sources such as newspapers, blogs, and social networking sites to analyze the sentiment about the user and determines his/her involvement in Political events like rallies, riots, Scams, frauds, robberies, among others.
• As in the case for other modules, the integration and prediction of other attributes allows for successful user disambiguation.
• E.g. If user received award from president it could lead to a low score. On the other hand, an individual who is an active participant in rallies (news article in NY Times) 🡪 high score
• (vi) Social graph-based score computation
• The final module is based on our patented system/algorithm, Tweethood. We predict the threat level for all friends of the POI (based on above listed = factors) and aggregate to obtain score for the central POI.
• E.g. If Threat (friend1) = 0.9 AND Threat (friend2) = 0.1 AND Threat (friend3) = 0.8 AND Threat (friend4) = 0.7 AND Threat (friend5) = 0.5, then Threat (POI) = 0.6

​

18

* *

Modules of InXite

• Threat Assessment:
• Once the profiles of a user have been constructed, we then examine the various attributes to determine whether the given user is a potential terrorist. For example, a user’s attributes (e.g., age, location, etc.) as well as their behavioral, social and psychological properties are extracted using our analytics algorithms. We then also apply existing algorithms (e.g., Mark Sageman’s, a researcher is psychology, to enhance a user’s psychological profile.
• These profiles will be used to determine whether the current user will carry out terrorist attacks, homicides, etc. For threat assessment, here are some results that we have obtained using existing algorithms as well as using our data analytics algorithms.
• Demographics:
• Up to .2 points are assigned to fitting into ranges for the following categories: age, education, religion, politics, and hobbies. These are then added up for the final demographics score.
• The ranges for these categories are based on the research of Marc Sageman

​

19

* *

Modules of InXite

• Psychology: Verb usage is categorized into traits. Four of these traits are found to be indicative of low psychological stability. These four traits are measured by percentage of total verb usage and added together to form the psychology score. The psychology sub-module is based on techniques given in the literature.
• NLP (Natural Language Processing): Weighted average of sentiment used between verbs and high profile nouns (i.e. White House or Pentagon).
• Negative or threatening verb analyses have a weight of 1 while positive or benign verb analyses have a weight of .1. This allows strong statements such as a correlation of “bomb” and “Tallbuilding” to produce an overwhelmingly high score.
• Social Structure: Standard mean average of friends’ threat scores. A friend’s threat score is the average of their other scores (demographics, psychology, NLP, social structure, background and online reputation).
• Background Checks: Represents a DoD standard background check on the individual.
• Online Reputation: If no previous association is found with this person or all associations are positive the score will be 0.
• Any score higher than this directly represents the percentage of previous associations from mainstream media that are analyzed to have a negative sentiment.

​

20

* *

Modules of InXite

• Our design utilizes Tweethood and SNOD as well as existing algorithms to develop a comprehensive system for threat assessment/evaluation.
• InXite Psychosocial Analysis
• It consists of a variety of techniques including visualizing ontologies and generating word clouds.
• This module also uses two novel data analytics techniques that we have developed, namely micro-level location mining and sentiment mining..
• Micro-level location mining:
• This refers to a method for determining specific or fine-grained locations that may be mentioned in communications between individuals or groups of individuals.
• In addition to locations, the technique can also be used to carry out fine-grained detection of other attributes such as hobbies, places traveled and events.
• Our technique for micro-level location mining is unique and uses a crowd-sourced database, namely Foursquare and also uses the following tools/algorithms:
• WordNet is used for disambiguation of locations mentioned in communications between individuals/groups such as messages or tweets.

​

21

* *

Modules of InXite

• Tweethood is used for identifying a city-level location, which in turn is used to narrow the search for micro-level locations within the identified city.
• Prior work in the literature aims to identify zip codes/locations based on the particular language used in tweets, and is thus a different approach towards location identification.
• We have considerably enhanced current efforts and have developed a novel technique for micro-level location mining.
• The pseudocode for micro-level location mining is as follows:
• 1.Discmbiguate locations mentioned in tweets using Wordnet;
• 2. Use Foursquare to find the general locations;
• 3. When locations are missing in the tweets, use Tweethood to find the locations as well as the city-level locations;
• 4. Use variations of Tweethood mine further to pin point the exact location.

​

22

* *

Modules of InXite

• Sentiment Mining:
• We have also designed sentiment mining techniques about a certain keyword/topic.
• For example, what does “John Smith” feel about “Tax Breaks” or what does “John Smith” feel about ‘Osama bin Laden”.
• Our methods use existing as well as our own algorithms (Tweethood, SNOD) for classifying user messages as positive, negative or neutral or whether it belongs to a novel class.
• We use emotion mining and also social behavioral mining to determine sentiments. In particular, we use the following two techniques:
• User demographics-based – For example, if we know that 95% of all African Americans are pro-Obama, then our system will give a positive bias to a tweet from an African American about President Obama.
• Social factor-based (based on Tweethood) – If we know that 9 out of 10 friends of a user are pro-Obama, then we will give a positive bias to tweets from that user.
• Our training dataset is a labeled dataset – each tweet with its sentiment type positive or negative or neutral.
• We obtained the labeled training dataset from set of tweets which has emo-icons in it.

​

23

* *

Modules of InXite

• Based on the emo-icon, we determined the label of the tweet and made the training dataset. For each of the training data/tweet, we first remove the stopwords in it. Then we remove all the words starting with ‘@’ or ‘http’.
• Then we convert each token of the tweet to standard form means: we convert a token like ‘hungryyyyyy’ to ‘hungryy’.
• Then from each tweet we make the list of unigrams and bigrams for that tweet with its sentiment type. We saved the list of unigrams and bigrams in a HashSet and also convert the tweet as unigram and bigrams.
• Now, for each token in the Hashset and for each tweet, we check whether the tweet contains the token or not.
• Then we make the occurrence matrix based on their presence/absence.
• So, at this point we have a dataset of large numbers of dimensions. So to reduce the dimensionality we leverage the entropy concept.
• We choose best N attribute based on the higher information gain.
• Now, we have considerably good data and then we use WEKA for classifying purpose and we use Naïve Bayesian classifier and decision tree (J48) classifier.

​

24

* *

Modules of InXite

• The pseudocode for the Sentiment Mining Algorithm is as follows :
• Input: Set of training tweets T, Bag of stopword S, Set of testing tweets R, number of attribute N
• Output: Labels of each of the tweet of R.
• For every tweet in T
• Remove all the stopwords in S.
• Remove the words starting with ‘@’ or ‘http’ and make each token as standard form.
• Make the set G, of unigrams and bigrams and convert each tweet as a set of those and make the set W. [W contains the unigrams and bigrams of each tweet.]
• For each token g in G
• For each token in w in W
• If g matches w then encode it as 1 and fill up the occurrence matrix M
• Else encode it as 0 and fill up the occurrence matrix M
• Choose the best N attribute from M based on the information gain and make new dataset D.
• Use D as the training dataset and build the classifier NB or J48.
• Use the trained classifier to classify the instances of test set R.

​

​

25

* *

Modules of InXite

​

• InXite Threat Detection and Prediction:
• While the threat evaluation and assessment techniques described will determine whether a person is a threat or not based on some pre-determined attributes, we have also designed data mining/machine learning techniques to determine whether a person will commit future terrorist attacks.
• That is, threat prediction in InXite is carried out through a series of stages either meant to find suggested threatening behavior in a user or to eliminate individuals who are unlikely to be a current or become a future threat.
• A very broad list of users is created by first picking out those who use a list of specific nouns and verbs.
• This list is manually maintained and may include code words that are added after the discovery that they are being used.
• Persons who use enough of these words in a single statement are flagged and classified in the next stage.
• The number of statements or tweets that pass this stage contains a high number of false positives.
• In the next stage, the flagged statements are tagged by part of speech similar to the content-based score in the threat evaluation section.
• Classifiers are trained on labeled and tagged data from statements manually confirmed by analysts or engineers to indicate imminent threats.

​

26

* *

Modules of InXite

​

• Because many non-threatening statements can contain threatening words like “kill”, the classifiers are useful in removing false positives from the statements passed in from the first phase.
• Phrases like “killing time” or “people would kill for this opportunity” are eliminated in this way.
• This method of classification also serves to solidify seemingly innocent sentences that may be using code words.
• Even if the words of the sentence are replaced, the sentences’ structure and placement remain the same and compare similarly to sentences that explicitly state the obvious threatening language.
• This automated method is so effective because every word in a statement contributes an equal probability to the statement’s classification as a threat or not.
• Given enough samples of threats and non-threats, as each word is compared to its individual threat level and placement within the whole statement, the algorithm can determine where it belongs.
• When the algorithm decides that a statement is threatening, it is grouped with similar statements based on identified threatening words and grammatical structure.
• Using feedback from the analyst or user, these groups can be solidified or changed to reflect similar threatening statements.
• Predicting threats based on data content first requires that the threatening or useful data be separated from the extremely large amount of benign or useless data.
• This can be accomplished with high accuracy through the union of linear discriminate analysis and bag of words filtering.

​

27

* *

Modules of InXite

​

• This process has the benefit of breaking possibly threatening content into feature groups and dynamically detecting new threatening content categories.
• However, it also produces a large amount of false positives.
• We have multiple methods; One is based on a Naïve Bayes classifier to help eliminate false positives.
• The classifier trains threats and benign content based on individual words and their part of speech obtained from a tagger developed by CMU specifically for twitter language usage instead of published text documents.
• The other is based on our patented technology SNOD. Our integrated algorithms provide much higher accuracy than the current approaches.
• Other Features Users in our algorithms:
• WORD CLOUDS: Shows frequently used words. More frequent words are shown with a larger font.
• ENTITY CLOUDS: Shows frequently used entities. Entities of interest among the profile and their friends are shown in size by correlation to frequency of discussion.
• TWEET FREQUENCY: Line graphs of tweets over time show useful timing information. Lack of tweets are as important as writing tweets. Sleeper agents are known to cut contact and go silent before acting.
• SOCIAL GRAPH VISUALIZATION: Visually shows the threat level of the most popular friends that are associated with a given user online.
• ASSOCIATED IMAGES: Brought to life in a slide show, all the images gathered from online sources for the given user profile.

​

​

28

* *

Cloud-based Design of InXite

​

• For InXite to function in real-time and to handle massive numbers of tweets and data, we need a cloud to host all of the modules of InXite. In this section, we describe our cloud-based design for various InXite modules.
• The design uses two separate tools, namely Storm and HBase, for the development of the various InXite modules.
• A separate Storm topology is constructed for each of the InXite modules.
• Furthermore, HBase is used by all topologies for storage and retrieval of user profiles.
• We now present the design of the various topologies in the form of pseudo-code based algorithms.
• InXite Information Engine in the Cloud
• 1. Identify a user of interest using a spout.
• 2. Perform the steps included in the entity extraction and information integration module for the user selected in step 1 in a custom bolt (this also includes the implementation of Tweethood in the Cloud).
• 3. Store the identified attribute <key,value> pairs in HBase.
• 4. Perform the steps for information integration for the user selected in step 1 using the attribute <key, value> pairs obtained in step 2 in a separate custom bolt.
• 5. Update the results stored in HBase with the results of step 4.

​

29

* *

Cloud-based Design of InXite

​

• InXite Profile Generation and Prediction
• 1. Identify a user of interest using a spout.
• 2. Use the attribute <key, value> pairs created by the information engine to build a user profile in a bolt, which is stored in an HBase schema.
• 3. Update the user profile by predicting values for other attributes using the attribute <key, value> pairs in a separate custom bolt.
• 4. Conduct a threat assessment of the identified user with the help of the various scores described earlier (demographics based, psychological, etc.) using a custom bolt.
• 5. Update the user profile with the results of threat assessment.
• InXite Psychosocial Analysis and Prediction
• 1. Identify a user of interest using a spout.
• 2. Identify micro-level locations for the user and store them as a part of their profile using a custom bolt (this includes the implementation of Tweethood in the Cloud)
• 3. Perform a sentiment analysis of the user’s messages/posts/tweets using a custom bolt and store the results as a part of their profile.
• 4. Use a separate custom bolt to construct word/entity clouds, graphs for tweet frequency, determine the threat score for the top friends of this user and download images associated with this user.
• 5. Store all information obtained in step 4a as a part of the user’s profile.

​

30

* *

Cloud-based Design of InXite

​

• InXite Threat Detection and Prediction
• 1. Identify a user of interest using a spout.
• 2. Perform threat prediction for the identified user in a custom bolt using the classification algorithms described earlier.
• 3. Store the results of threat prediction as a part of the user’s profile.
• Implementation
• All of the modules of have been implemented
• These include entity extraction and information integration, profile generation and threat analysis, psychosocial analysis, as well as threat prediction.
• With respect to the cloud implementation, we have completed the implementation of Tweethood in the cloud as well as SNOD in the cloud.
• The remaining modules of InXite are yet to be implemented in the cloud.
• Other Applications:
• InXite is a multi-purpose system. In addition to InXite-Security we have also designed and developed other applications for InXite including InXite-Marketing and InXite-Law.

31

* *

Other Applications of InXite

​

• InXite-Marketing carries out entity extraction and information integration and builds profiles of persons of interest.
• In this case the person of interest is the user who is interested in a product such as iPhone or a particular kind if pizza.
• We can develop word clouds as well as associated images for this user.
• We can also carry out sentiment mining and obtain results such as John prefers iPhone to Android.
• InXite-Marketing also finds locations from Tweethood and novel classes using SNOD.
• However there are some differences between the features between InXite-Security and InXite-Marketing.
• It does not make sense to determine whether a person is a threat or whether he will commit a crime for InXite-Marketing.
• Furthermore it does not make sense to determine the psychosocial profile of a user.
• What is useful for InXite-Marketing is integrating with a recommender system; for example, we can analyze the user preferences and make recommendations
• We can predict the sentiment of tweet for a person, though it needs some further improvement like incorporating NLP techniques

​

32

* *

Other Applications of InXite

​

• Our system is easily workable for a particular person’s tweet and a particular subject. For example, if we want to determine John Smith's sentiment about iPhone-5, we can go through all his tweets about the iPhone-5 and based on the sentiment of each of the tweet, we can determine the overall sentiment of John Smith about iPhone-5.
• So now, once we know the sentiment of John Smith about iPhone-5, say it is positive, we can recommend him some more i-products or products related to the iPhone-5, for example, headphone, charger, etc. Here is another factor to consider that we call peer effect.
• The positive statements made by an individual's friends can be mined for individual products.
• Because the friends or associates of this individual talk positively about a product, especially one that the individual does not mention themselves, we can extrapolate these products for the recommender system.
• For example, if John Smith has ten friends and all of them have a positive sentiment about Android phones, then it is a good idea to recommend some Android products to John Smith.
• So, we can consider it as a weighted vector of 'personal sentiment' and 'peer sentiment’.
• Based on the weighted factor, we can decide what product and in what extent we should recommend him.
• These weights will be mostly influenced by the individual's personal sentiment and the sentiment of his peers will help fill in the gaps and more greatly expand the recommendation options for this person.

​

33

* *

Summary and Conclusion

• Summary
• We have described the design and implementation of InXite for multi-purpose applications (e.g., Security, Marketing).
• InXite will be a great asset to the analysts who have to deal with massive amounts of data streams in the form of billions of blogs and messages among others; for example, by analyzing the behavioral history of a particular group of individuals as well as details of concepts such as events, analysts will be able to predict behavioral changes in the near future and take necessary measures.
• We have also discussed our use of cloud computing in the implementation of InXite.
• Conclusion
• Analyzing data streams and updating the existing classification models is a daunting task; most existing behavior-profiling techniques are manual, which require days to analyze a single blog sample and extract its behavioral profile.
• Even existing automated techniques have been tested only on a small sample of training data.
• By integrating our approach with a cloud computing framework, as well as real-time processing techniques, we have overcome this barrier and provide a highly scalable behavior modeling tool, thereby achieving higher accuracy in rapidly detecting new patterns and trends.
• Furthermore, existing profiling (manual or automated) technique rarely address the evolving characteristics of data streams such as blogs and messages in real-time.
• Therefore, our system will have a tremendous advantage over most behavior-based approaches by quickly responding to the dynamic environment.

​

34

* *

Future: InXite for COVID-19 and Privacy Considerations

• InXite for COVID-19?
• Much of the design and development of InXite was carried out between 2011 and 2015; nevertheless as new technologies emerge (e.g. Machine Learning, Entity Resolution), they can be incorporated into InXite; InXite design is based on a plug and play approach.
• The greatest plague faced by the humans in a 100 years is the COVID-19 pandemic.
• A system such as InXite can be adapted to answer many questions such as who is likely to get COVID and where are the hotspots?
• We need to create profiles of individuals (potential patients) by collecting information about them and predicting additional attributes; Micro-level Location Mining may enable us to pinpoint the future hot spots.
• What about Privacy?
• Regardless of whether we use InXite for Security or for COVID-19, privacy is a major consideration.
• This is because we are collecting all kinds of information about individuals and analyzing the information to make predictions.
• As a result, their privacy may be violated – e.g., personal and health related information getting into the wrong hands.
• Therefore, we need to ensure privacy aware data collection, data storage, data integration, data sharing and data analytics.
• But this may interfere with the accuracy of the information.
• There is a tradeoff between using InXite for certain applications and the privacy of the individuals and this aspect needs more work.

​

35

* *