The subject matter of the present invention relates to systems and methods for making statistical inferences based upon large quantities of largely unstructured data.
In social media, each day, billions of messages including texts, pictures, videos, blogs, surveys, discussions and feed-backs etc. are published on a variety of networks. A tremendous quantity of information is embedded inside social media data.
It can be understood that companies and other organizations might benefit from utilizing social media data to discover information in conversations about their brand, to engage with their customers, to conduct advisement and investment efficiency analysis, to manage and reduce potential risk and identify the factors that affect company sale and revenues. Thus, it can be understood that it would be beneficial to have systems for estimating and inferring end user buy intent from massive social media data.
The present invention relates to systems and methods for making inferences based upon social media data. More particularly, the present invention relates to systems and methods that enable usage of data embedded inside social media data to help estimate buy intent (BI), to allow companies and/or organizations to utilize this information to track conversations about their brand, to engage with their customers/users, to conduct advisement and investment efficiency analysis, to manage and reduce potential risk, and identify the factors that may affect company sales and revenues.
The accompanying drawings illustrate one or more exemplary embodiments and, together with the detailed description, serve to explain the principles and exemplary implementations of the present inventions. One of skill in the art will understand that the drawings are provided for purposes of example only.
In the Drawings:
Various exemplary embodiments of the present inventions are described herein in the context of systems and methods for making statistical inferences based upon large quantities of largely unstructured data.
Those of ordinary skill in the art will understand that the following detailed description is illustrative only and is not intended to be limiting. Other embodiments will readily suggest themselves to such skilled persons having the benefit of this disclosure, in light of what is known in the relevant arts.
In the interest of clarity, not all of the routine features of the exemplary implementations are shown and described. It will be appreciated that in the development of any such actual implementation, numerous implementation-specific decisions must be made in order to achieve the specific goals of the developer.
Throughout the present disclosure, relevant terms are to be understood consistently with their typical meanings established in the relevant art.
Generally speaking, a system and methods are herein disclosed that enable usage of data embedded inside social media data, to allow companies or organizations to utilize this information to track conversations about their brand, to engage with their customers/users, to conduct advisement and investment efficiency analysis, to manage and reduce potential risk, and identify the factors that affect company sales and revenues. The following describes ways of using and analyzing collected social media data to determine the “buy intent” of an individual.
As will be described in further detail below, Buy intent (BI) may be defined as a real number ranging from 0.0 to 1.0 representing how likely one social media user will purchase or already bought any item from given set of products from a brand. BI is estimated based on a collection of relevance social media messages from a single user over a given period time.
As used herein, the acronym POS refers to part-of-speech tagging also called grammatical tagging or word-category disambiguation. It is the process of marking up a word in a text (corpus) as corresponding to a particular part of speech, based on both its definition, as well as its context—i.e. relationship with adjacent and related words in a phrase, sentence, or paragraph.
As shown in
According to some embodiments, methods are provided for substantially automated extracting of buy intent (BI) indications of social media users, using enabling computing devices to process social media data to estimate buy intent; to enable manual grading of real social data for BI estimation training; to enable combining of exact match analysis, vector space modeling, POS tagging and message replacement, and using Bayesian classification algorithms for BI estimation; utilizing BI scores to identify potential buyers; utilizing BI scores to Identify positive and negative influences; utilizing BI scores to estimate advertising efficiency; utilizing BI scores to estimate parameters for targeted fixed effects like gender, age, education, income, region, search history and purchase patterns; utilizing BI score to identify potential buyers common properties; and utilizing BI scores to trace stimulation factors for potential buyer status changes.
As can be seen in
In step 110 a pre-defined configuration is loaded into the cluster that has been started to enable the clustering and capability establishment of the server cluster to perform the necessary functions for data collection and organization. The configuration also enables the cluster to assess its own capacity to handle the data volume and dynamically set the cluster size
In step 115 servers can be started in the cluster, to enable the configured cluster data to be processed and/or served to system and/or system users.
In step 120 a decision is to be taken whether there are enough servers to process and/or serve the users. If not, at step 125 additional server(s) may be added.
If there are enough servers at step 120, data is fetched in step 130, from multiple social media sources, in multiple data formats.
In step 135 raw data is acquired from the collected data, to be further processed.
In step 140 an index is created from the raw data, for enabling rapid categorization, sorting, fileting and searching of social media data.
In step 145 the raw data is processed by an algorithm to detect the specific user(s), and to correlate detected user(s) to the user profiles in the Social Media Data Collection System 100.
In step 150 the indexed and/or user correlation data are further processed to determine whether the collected data is to be persisted/maintained in the system 100, or is to be distributed to system users.
In step 160 the processed persistent and/or distributed data is fed to a further BI processing engine or element, for BI specific analyzing, to detect user Buy Intent.
In the embodiment illustrated in
In preferred practice, the collected data is supplied to a computational-capable server having at least a processor and at least a storage capability, to teach the processor to analyze the available data. Once taught, additional data can be provided to the server to generate a rating of the buy intent of the customers. For example, there are many millions of messages collected in the system per hour or even second, and the number is continually growing daily as more customers, accounts and different search criteria and interests are entered. The system's servers may be deployed in a cluster with virtually unlimited computing power. This process is dynamic in that more servers can be added automatically if needed.
As can be seen in
In step 205, the BI process is started by the Social Media Data Collection System 100.
In step 210 data is fetched from multiple social media sources, and in step 215, the fetched social media data is collected together in the system data storage facilities.
In step 220 the data is normalized, to convert all received data into a unified format, by system data converter element.
In step 225 the data is processed by a dictionary, optionally with multiple languages, to further normalize data from multiple languages.
In step 230 the data is processed using a POS tag Analysis Engine, to identify critical POS sale data.
In step 235 search indexes are generated, to help rapidly search and sort collected data.
In step 245 data history is used to help generate more accurate search indexes. In step 240 learned data is updated, from the search indexes and/or historical data.
In step 250 the system determines whether sufficient learned data exists to generate accurate search indexes.
If no learned data exists, then in step 260 vector space modeling analysis may be executed, to further process collected data, to complement the accuracy of the learned data. Additionally or alternatively, in step 255 message similarity analysis is executed, to further process collected data, to complement the accuracy of the learned data.
If sufficient learned data exists, then in step 265 BI related data is fetched or created into a file, for example, including data source, time, brand, product, BI score, user information, influence weight, etc.
In step 270 a decision is taken by the system as to whether the BI information acquired is accurate enough.
If not, in step 275, Bayesian analysis is executed on the processed BI data.
Alternatively or additionally, if the data is accurate enough, in step 280, if the data is determined to be persistent, the system BI statistics are updated, to include the latest BI definitions, classifications, etc.
In step 295 the processed BI data as determined by the above steps is distributed to the system's servers, for distributions to system elements or components.
In step 296, the distributed BI data is ready for usage by system users.
The flow chart in
In step 310, the collected social media data is sent to multiple data collectors 1 to n.
In step 315 the data collectors data is consolidated, for example, collected from different sources into system data collectors.
In step 320 the data is normalized, for example, to aggregate different formats and types of data.
As can be seen, in step 325 multiple processors are used to pull BI related data from a container(s) and to further process one or more data elements, and then push the resultant processed data elements to a further data container.
In step 330 data enrichment is executed, optionally including processing the pre-processed data for BI related information such as sentiment, mood, intention, spam etc.
In step 335, container 2 may be further processed by Social Media Data Collection System 100.
In step 340 a data analysis engine processes container data for further BI related metrics, such as influence, topic, trend, popular keys, buy intention, targeting analysis etc.
In step 345 container 3 may be further processed by Social Media Data Collection System 100.
In step 350 a customer role engine dispatches the message to different queues based on system requirements, such as customer rules, conditional processing etc.
In step 355 container 4 may be processed by Social Media Data Collection System 100.
In step 360 a report engine processes the data to generate BI related reports.
In step 365 container x is processed by Social Media Data Collection System 100.
In step 370 a data API for paid customers is run, to manipulate and fetch data for advanced features or functions as may be used by paid users.
In step 375 container n is processed by Social Media Data Collection System 100.
In step 380 thread updates are monitored to determine BI related modifications in data threads.
The system described herein may have many servers on a cloud that collect data from social media. The collected data is put into a container of a network distributed cache. Many processes can concurrently access same container at any time. Each process can pull one message a time and process it, and may push the modified data into another container. A user can create any number of containers at run time. The system also enables many processes to work similarly, whereby the number processes run in the system is dependent on the system configuration. For example, a user may configure from 1 to hundreds, thousands or more processes at will. Of course, more processes may require more computing power.
It should be appreciated that embodiments of the present invention may include a combination of one or more of the following elements:
In one example, each message can be classified into five categories as follows:
Let Vki be the actual values and Nki be actual occurrence of Vki for individual K. The buy intent for individual k can then be calculated as:
The BIk and Nk are two independent variables though the Nk has significant effects on BIk estimation accuracy. We have found that the accuracy of BIk estimation becomes relatively stable when the number of messages used for each individual exceeds 30. Therefore, the analysis only chooses the individuals who have 30 or more relevant messages to participate in the data analysis, thereby to increase the estimation accuracy. A key technology relates to the classification of each relevant message into specified categories based on learned data.
BI Estimation
The BI estimation process can include sample dataset manual grading, raw data preparation, classification engine dynamic data training, BI calculation and parameters estimation etc.
Dataset manual grading: A significant process is classification engine training. It requires grading a set of messages into categories, for example the five categories described above, and feeding them into classification engine learn module. Although there is no lower limits of number graded messages, the number should exceed ten thousands to satisfy minimum estimation accuracy. Generally speaking, the more graded messages for learning will resulting in more accurate estimation. The estimation accuracy obeys logistic distribution with the following properties:
Probability function:
Distribution function:
The accuracy improvement becomes less and less significant if number of graded messages reaches more than hundred thousand. The actual regression of actual accuracy improvement and number graded messages for training dependents on actual data and quality of grading.
Raw data preparation: Graded data must be normalized before either fed into classification engine's learn module or used for BI estimation. This process transfers message text to lowercase and eliminates irregular spacing. It first changes word case and adds spaces before and after each punctuation character and then parses whole message into words and punctuation array using space as delimiter. The misspelled words will be replaced based on Bayesian probability. Then three messages will be reconstructed. The first message is constructed by inserting single space between consecutive words and appending punctuation character without using any space. The second message is constructed similarly as first one except original stop words being eliminated from word array. The third message is constructed similarly as first one except all the words in the array are transferred to corresponding root words. For example, the word “sits” and “sat” are transferred to sit, the word “cars” and “happier” are transferred car and happy respectively. Three MD5 hashes are calculated and persistent along with classification category information.
POS tags input for classification engine: In order to detect POS of social media message, huge amount of graded n-gram data has been incorporated into classification engine (CE). These n-grams are generated based on the largest publicly-available, genre-balanced corpus of English—the 450 million word Corpus of Contemporary American English (COCA), 1.8 billion words data from GloWnE and 1.9 billion words from 4.4 million Wikipedia articles. The data is consisted of three pieces of information: 1, 2, 3, 4, 5-word sequences, frequency counts and corresponding individual POS tags for the word sequences. The information stored in CE memory efficiently. Original phrase will be used as identifier If actual phrase char length is smaller than 16. The MD5 hash will be used as identifier otherwise.
POS tags selection: The Classification engine adapts Penn Treebank POS tag system. Every word in social media message is associated with POS tag. All possible POS tags are listed in Table 2. The algorithm for POS tag selection is based on probability learned in previous step. The message is normalized and parsed into word and punctuation array. Slide window can have maximum five words sequence and any phrase stops whenever it encounters a punctuation. The original phrase or its MD5 hash (if phrase is longer than 16 bytes) is constructed to retrieve learned POS tag information. The corresponding tags in learned data will be used for current message if match found. Relative heavier weight is given to longer n-gram if there is conflict between different sizes of n-gram. The selection always starts in current word and tries to use longer n-gram. After exhausting sequences and making the POS choice for each word, work window is slide to right and whole process is repeated till all words in message are exhausted.
Princeton WordNet for classification and similarity calculation: The Classification engine introduces Princeton WordNet for classification index computing. WordNet® is a large lexical database of English and it groups nouns, verbs, adjectives and adverbs into 117 000 cognitive synonyms (synsets). Synsets are interlinked by means of conceptual-semantic and lexical relations. CE also incorporates more than 124 thousands manual graded phrases (two or more words) into WordNet and expands grouping to all kinds of POS types. The new generated “Synsets” are utilized intensively to compute similarity index between two messages and detect whether given message is similar to learned data. Similarity index between two messages are calculated using following steps:
Vector space model and messages similarity calculation: The vector space model is widely used for related documents retrieval and messages similarity calculation mainly because of its conceptual simplicity and the appeal of the underlying metaphor of using spatial proximity for semantic proximity. Vector space model treats message as a point in an n-dimensional spaces where n is the number of common words in the two messages or message and a category. The coordinators of given message and group are calculated based word frequency occurred in message and group of messages. The similarity coefficient is usually expressed as vectors normalized correlation coefficient as follows:
Where gi is i'th word frequency for one of learned categories and mi is the i'th word frequency for current message. The advantage of vector space model is that it uses little computer memory and computing algorithm is simple and direct. The disadvantage is that it does not use other information like word order, word combinations, word meaning and AI technology etc.
Bayesian classification and Bi estimation: When conducting Bayesian classification, BI classification information is derived from graded data and current message. To estimate the probability of a current message belonging to a certain BI category, the following can be used:
Where Ci represents different BI subclass. Since only relative values are of interest, P(W) can be ignored.
From the Chain rule, we have
We use simplest phrases weight and calculate each phrase probability as follows
Bi classification is determined based on calculated probability of current message association with each BI category.
BI Estimation
After Graded BI category data has been fed into a CE learning module, message are normalized and parsed into word and punctuation array. The POS tags are selected based on the probability model described above. Exact message information, different length word phrases are constructed and occurrence frequencies are accumulated and recorded. Vector space model and Bayesian basic statistics also established. At run time, BI value for each message is calculated as follows:
After calculating the BI score for each social media message, each message can be used as an independent variable and fitted into general mixed model equations to estimate the best linear unbiased estimators for many major factors, like author age, gender, region, income, education etc. The correlation coefficients between sentiment score and BI score can be estimated. The BLUE of advertisement efficiency can be computed by introducing targeted criteria and other fixed independent variables into model. By calculating weighted average of BI score for individual user, we can further identify potential buyers at early stage and offer guideline for company production and other activity planning. Since every user in social media network is interlinked, BI score can also be used to identify the most positive and negative influencers to a given brand. User's BI score can also be used as selection filter to generate subset of social messages which then be fed into further topic analysis model to determine common properties among potential buyers. Also, filtered data is used identify the real relationship among potential buyers and trace circumstances that stimulate user status change. Following is the basic mixed model formula for the best linear unbiased estimation (BLUE).
Y=X′β+Z′μ+e
Where:
At this juncture, it should be clear that, in other embodiments, methods are provided for enabling computing devices to use social media data to estimate to enable manual grading of real social data for BI estimation training; to enable combining of exact match analysis, vector space modeling, POS tagging and message replacement, and Bayesian classification algorithms for BI estimation; utilizing BI scores to identify potential buyers; utilizing BI scores to Identify positive and negative influences; utilizing BI scores to estimate advertising efficiency; utilizing BI scores to estimate parameters for targeted fixed effects like gender, age, education, income, region, search history and purchase patterns; utilizing BI score to identify potential buyers common properties; and utilizing BI scores to trace stimulation factors for potential buyer status changes.
Further at this juncture, it should be clear that the capability of the system to provide automated identification, analysis and use of available social media data can enable companies to enhance revenue generation and business decision making.
As mentioned above, the system may be enabled by data processing algorithms that integrate Vector space models and Bayesian basic statistics. Although the underlying mathematics theory being used herein is partially similar in Vector space and Bayesian statistics analysis, the way in which the system utilize and apply these theories are unique, particularly in terms of the parameters derivation, actual data manipulation and iteration criteria. The system described herein incorporates AI and generic algorithms into existing vector space and Bayesian analysis, thereby changing both the process and the results.
According to some embodiments, the buy intention analysis is substantially executed using a computer code (for example, a computer program which is written using C#) running on a server (for example, window server 2008) a connected to a data cloud (for example, Amazon servers cloud). There can be multiple physical servers running on the cloud on a cluster system behind a load balancer. The load balancer is configured to receive huge numbers (example, Millions per second) of social media data items that are downloaded every minute, and to distribute the data to one of the servers in the cluster system. The BI system on the server analyzes the data and derives a buy intent value between 0.0 and 1.0. The higher the number means the stronger the buy intent, wherein the maximum value is 1.0.
In summary, the following aspects are noteworthy:
Further, embodiments of the BI system described herein may be used to calculate a numeric value (between 0.0 to 1.0), to represent whether the user is likely to buy one or more products from a brand.
Organizations may use this data to implement, for example:
Exemplary embodiments have been described with reference to specific configurations. The foregoing description of specific embodiments and examples has been presented for the purpose of illustration and description only, and although the invention has been illustrated by certain of the preceding examples, it is not to be construed as being limited thereby.
Number | Name | Date | Kind |
---|---|---|---|
6094829 | Li | Apr 2000 | A |
6601234 | Bowman-Amuah | Jul 2003 | B1 |
7308418 | Malek et al. | Dec 2007 | B2 |
7761848 | Chaffin | Jul 2010 | B1 |
7930302 | Bandaru et al. | Apr 2011 | B2 |
8225271 | Eldridge et al. | Jul 2012 | B2 |
8266148 | Guha et al. | Sep 2012 | B2 |
8584082 | Baird et al. | Nov 2013 | B2 |
8635281 | Kelly | Jan 2014 | B2 |
8645912 | Welchman et al. | Feb 2014 | B2 |
8683433 | Nayak et al. | Mar 2014 | B2 |
8688726 | Mahajan et al. | Apr 2014 | B2 |
8694520 | Krishnakumar et al. | Apr 2014 | B1 |
8707259 | Trofin et al. | Apr 2014 | B2 |
8732669 | Valdiviezo et al. | May 2014 | B2 |
8819659 | Ramer et al. | Aug 2014 | B2 |
8832644 | Hirsch | Sep 2014 | B2 |
8924914 | Boeckenhauer et al. | Dec 2014 | B2 |
8978006 | Hirsch | Mar 2015 | B2 |
8996371 | Hurvitz et al. | Mar 2015 | B2 |
9002892 | Benyamin et al. | Apr 2015 | B2 |
9116983 | Ameri-Yahia et al. | Aug 2015 | B2 |
9223831 | Baker et al. | Dec 2015 | B2 |
9448776 | Sankaran | Sep 2016 | B1 |
9672032 | Zhou | Jun 2017 | B2 |
9696967 | Balasubramanian | Jul 2017 | B2 |
10073794 | Thomas et al. | Sep 2018 | B2 |
10095686 | Zhang | Oct 2018 | B2 |
20010027472 | Feng | Oct 2001 | A1 |
20060053156 | Kaushansky et al. | Mar 2006 | A1 |
20070050446 | Moore | Mar 2007 | A1 |
20070174255 | Sravanapudi | Jul 2007 | A1 |
20080249764 | Huang et al. | Oct 2008 | A1 |
20090319342 | Shilman et al. | Dec 2009 | A1 |
20100114654 | Lukose et al. | May 2010 | A1 |
20100131563 | Yin | May 2010 | A1 |
20100169304 | Hendrickson et al. | Jul 2010 | A1 |
20100235313 | Rea et al. | Sep 2010 | A1 |
20110113447 | Jung | May 2011 | A1 |
20110191372 | Kaushansky et al. | Aug 2011 | A1 |
20110213655 | Henkin | Sep 2011 | A1 |
20120036007 | Robertson et al. | Feb 2012 | A1 |
20120116868 | Chin et al. | May 2012 | A1 |
20120158489 | Benyamin et al. | Jun 2012 | A1 |
20120203584 | Mishor | Aug 2012 | A1 |
20120215625 | Ramer et al. | Aug 2012 | A1 |
20120246033 | Keiser | Sep 2012 | A1 |
20120246054 | Sastri et al. | Sep 2012 | A1 |
20120290340 | Ramanujam et al. | Nov 2012 | A1 |
20120290938 | Subbarao et al. | Nov 2012 | A1 |
20120296845 | Andrews et al. | Nov 2012 | A1 |
20130080264 | Umeda | Mar 2013 | A1 |
20130103385 | Ghosh et al. | Apr 2013 | A1 |
20130124653 | Vick | May 2013 | A1 |
20130132851 | Konopniki et al. | May 2013 | A1 |
20130166379 | Ehindero et al. | Jun 2013 | A1 |
20130238356 | Torrii | Sep 2013 | A1 |
20130246430 | Szucs et al. | Sep 2013 | A1 |
20130263019 | Castellanos et al. | Oct 2013 | A1 |
20130311323 | Stoll et al. | Nov 2013 | A1 |
20130339126 | Cui et al. | Dec 2013 | A1 |
20140025749 | Glowacki | Jan 2014 | A1 |
20140025758 | Glowacki | Jan 2014 | A1 |
20140056526 | Scipioni | Feb 2014 | A1 |
20140136323 | Zhang et al. | May 2014 | A1 |
20140188897 | Baker et al. | Jul 2014 | A1 |
20140254454 | Calcev | Sep 2014 | A1 |
20140337320 | Hernandez et al. | Nov 2014 | A1 |
20150106304 | Gupta | Apr 2015 | A1 |
20150112753 | Suvarna | Apr 2015 | A1 |
20150186790 | Ehlen | Jul 2015 | A1 |
20160087933 | Johnson | Mar 2016 | A1 |
20170109141 | Thomas et al. | Apr 2017 | A1 |
20170243125 | Thomas et al. | Aug 2017 | A1 |
20170300564 | Feng et al. | Oct 2017 | A1 |
20170315996 | Fothergill | Nov 2017 | A1 |
20180165706 | Feng et al. | Jun 2018 | A1 |
Number | Date | Country |
---|---|---|
2698955 | Feb 2014 | EP |
2010008685 | Jan 2010 | WO |
2013101812 | Jul 2013 | WO |
2013158790 | Oct 2013 | WO |
2013158840 | Oct 2013 | WO |
2013177230 | Nov 2013 | WO |
2014143018 | Sep 2014 | WO |
Entry |
---|
S. Buckley et al., “Social media and customer behavior analytics for personalized customer engagements,” in IBM Journal of Research and Development, vol. 58, No. 5/6, pp. 7:1-7:12, Sep.-Nov. 2014, doi: 10.1147/JRD.2014.2344515. (Year: 2014). |
Christl et al., Using Mobile Technology for Inter-Organisational Collaboration and End-Customer Integration, ACM, 2013, pp. 1-8. |
Rosa et al., Smart Composition of Reusable Software Components in Mobile Application Product Lines, ACM, 2011, pp. 45-49. |
Sung et al., Mobile Application Development Classes for the Mobile Era, ACM, 2014, pp. 141-146. |
Bharat et al., Building Distributed, Multi-User Applications by Direct Manipulation, ACM, 1994, pp. 71-81. |
Koziokas et al., Usability Testing of Mobile Applications: Web vs. Hybrid Apps, ACM, 2017, pp. 1-2. |
Meskens et al., Plug-and-Design: Embracing Mobile Devices as Part of the Design Environment, ACM, 2009, pp. 149-153. |
Yang et al., Personalized Mobile Application Discovery, ACM, 2014, pp. 49-54. |
Zhan et al., MCKit: A Mobile App for Conferences, ACM, 2014, pp. 59-64. |
Kirk, M., Thoughtful Machine Learning: A Test-Driven Approach, O'Reilly Media, Inc., Ch. 4, Naïve Bayesian Classification, 2014. |
Number | Date | Country | |
---|---|---|---|
20170148038 A1 | May 2017 | US |