System and method for learning customer journey event representations and predicting outcomes using neural sequence models
The system uses neural sequence models to convert customer journey data into vectors for real-time prediction, addressing the lack of personalization in contact center routing by optimizing interactions and agent allocation.
Patent Information
- Application Number
- JP2021559102
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2019-04-09
- Filing Date
- 2020-04-09
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2040-04-09
AI Technical Summary
Existing contact center routing systems fail to provide personalized and context-aware interactions as customers switch from web and mobile interfaces to voice communications, lacking the ability to leverage real-time data and machine learning for optimal routing and outcome prediction.
A system utilizing neural sequence models to learn customer journey event representations, converting raw data into vector representations, and applying machine learning algorithms like RNNs for real-time prediction of the next best action or outcome, integrating with WebRTC and VoIP for seamless communication.
Enables personalized and context-aware interactions by predicting optimal communication channels and agents based on customer behavior, enhancing customer experience and resource utilization in contact centers.
Smart Images

Figure 0007500604000001 
Figure 0007500604000002 
Figure 0007500604000003
Abstract
Description
[Technical field]
[0001] The present invention relates generally to telecommunications systems and methods, as well as website, mobile app, analytics, and contact center interactions through the use of technologies such as predictive analytics, machine learning, WebRTC and VoIP plug-ins for browsers. CROSS-REFERENCE TO RELATED APPLICATIONS AND PRIORITY CLAIM
[0002] This application is related to U.S. Provisional Application No. 61 / 979,479, filed April 14, 2014, entitled "System and Method for Interaction Routing by Applying Predictive Analytics and Machine Learning to Web and Mobile Application Content," which has been converted into pending U.S. Provisional Application No. 14 / 686,404, filed April 14, 2015, entitled "System and Method for Interaction Routing by Applying Predictive Analytics and Machine Learning to Web and Mobile Application Content," and pending divisional U.S. application No. 16 / 352,022, filed March 13, 2019, entitled "System and Method for Interaction Routing by Applying Predictive Analytics and Machine Learning to Web and Mobile Application Content." This application claims priority to U.S. patent application Ser. No. 16 / 379,110, also entitled “System and Method for Customer Journey Event Representation Learning and Outcome Prediction Using Neural Sequence Models,” filed Apr. 9, 2019, the contents of which are incorporated herein by reference. Summary of the Invention [Problem to be solved by the invention]
[0003] A system and method for learning customer journey event representations and outcome prediction using neural sequence models is presented. Multiple events are input to the module, each event has a schema that includes the characteristics of the event and their modality (web click, call, email, chat, etc.). Events of different modalities can be captured using different schemas, and thus the embodiments described herein are schema-agnostic. Each event is represented by the module as some number of vectors, and multiple vectors are generated in total for each customer visit. These vectors are then used in sequence learning to predict the probability of the next best action or outcome in real time in the customer journey using machine learning algorithms such as recurrent neural networks.
[0004] In one embodiment, a method is presented for transforming raw customer journey event data from browsing sessions of multiple customers on a website to obtain a machine learning ingestible dataset for obtaining vector representations of web events, the method includes: logging and storing events from the browsing sessions, the events including activity data from multiple customers; pre-processing the raw data captured on the website, further including removing or masking / encrypting values in all columns containing personally identifiable information; identifying class attributes that need to be predicted during model training; identifying and removing columns that duplicate class attributes; removing array type data; removing timestamp data; converting all Boolean columns to integer type; and replacing all null and unknown values; and obtaining an ingestible dataset that can be applied to an algorithm to obtain vector representations of web events.
[0005] The events may be stored in a common schema. The pre-processing steps may be performed in sequential order. Converting all Boolean data to integer types further includes replacing "true" with 1 and "false" with 0. Replacing all null and unknown values further includes replacing all null and unknown values with 1 for integer types, 0.0 for double types, and "unknown" for string types.
[0006] In another embodiment, a method is presented for obtaining a vector representation of web events, the method including: logging and storing events from a browsing session, the events including activity data from multiple customers; pre-processing raw data captured on a website, the pre-processing further including removing or masking / encrypting values in all columns including personally identifiable information; identifying class attributes that need to be predicted during model training; identifying and removing columns that duplicate class attributes; removing array type data; removing timestamp data; converting all Boolean type data to integer type; and replacing all null and unknown values; obtaining an ingestable dataset that can be applied to an algorithm to obtain a vector representation of web events; training a feedforward neural network with the ingestable dataset; inputting the web events into the feedforward neural network, the output including a vector representation of each activity in the web events; and appending a vector column to the dataset for a corresponding activity in the web events. Additional steps may include inputting the added dataset into a trained neural network including LSTM cell units and dense neuron units, where the input layer and hidden layer of the neural network include LSTM cell units and the output layer includes dense neuron units; obtaining the processing sequence to be input to the dense neuron layer as a single vector; applying a softmax function to the single vector; and obtaining outcome probabilities for the vector.
[0007] The vector representations include a common n-dimensional space. Training includes predicting an event class of the input. The trained feedforward neural network includes a plurality of input layer branches. The plurality of input layer branches further includes at least a first branch that accepts categorical attributes and a second branch that accepts continuous, numerical attributes.
[0008] The events are stored in a common schema. Converting all Boolean data to integer types further includes replacing "true" with 1 and "false" with 0. Replacing all null and unknown values further includes replacing all null and unknown values with 1 for integer types, 0.0 for double types, and "unknown" for string types. The pre-processing steps are performed in sequential order. Each event includes a schema that includes characteristics and modalities.
[0009] The neural network is trained using the preprocessed data, and the method for preprocessing includes identifying a set of consequence IDs for a given entity, defining a data frame schema that captures the visit ID, the time-ordered sequence of event IDs within a given visit, and the class indicator, querying the event data frame to populate the defined data frame, removing consequence events from the sequence, and replacing the event IDs with their respective event vectors.
[0010] The input layer accepts inputs in timestamp order, and a single vector contains a representation of the previous event in the sequence. [Brief description of the drawings]
[0011] [Figure 1] FIG. 1 illustrates one embodiment of a system in a networked environment.
[0012] [Figure 2A] FIG. 2A illustrates one embodiment of a computing device.
[0013] [Figure 2B] FIG. 2B illustrates one embodiment of a computing device.
[0014] [Diagram 3]FIG. 3 is a flow chart illustrating one embodiment of a process for processing data.
[0015] [Figure 4] FIG. 4 illustrates one embodiment of a system for applying predictive analytics and machine learning.
[0016] [Diagram 5] FIG. 5 illustrates one embodiment of a system for converting events into a vector representation.
[0017] [Figure 6] FIG. 6 illustrates one embodiment of a process for capturing a vector representation.
[0018] [Figure 7] FIG. 7 illustrates one embodiment of a system for continuous neural networks. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0019] For the purposes of promoting an understanding of the principles of the invention, reference will now be made to the embodiments illustrated in the drawings and specific language will be used to describe the same. Notwithstanding the above, it will be understood that no limitation of the scope of the invention is intended. Any changes and further modifications in the described embodiments, and any further applications of the principles of the invention as described herein, are contemplated as would normally occur to one skilled in the art to which the invention pertains.
[0020] Today, customers typically initiate and conduct their interactions with a product or service provider through the provider's website or mobile application, which are quickly becoming an organization's primary interface for communicating with customers, suppliers, and other stakeholders.
[0021] Traditional contact center routing to a customer care agent is done via data derived from telephone parameters such as calling line ID, or via simple selections made by the customer via DTMF tones in an interactive voice response session (e.g., press 1 for sales, 2 for support, etc.).
[0022] There is now an entirely new paradigm possible for contact centers where customers initiate business interactions via business websites or mobile applications, which are now the primary interface through which businesses support their customers. With the advent of technologies like VoIP and WebRTC, it is possible to build voice and video capabilities directly into websites and browser or mobile application based experiences. This mechanism does not even require the customer to have a phone number, as the media pathway between the customer and the customer's browser or mobile application is established peer-to-peer.
[0023] The increasing adoption of WebRTC, including telephony plug-ins for web browsers, and further Internet-based telecommunications developments by Google® (e.g., GoogleTalk®), Cisco (e.g., Jabber Guest), and the like, has made it possible to build unified communications capabilities directly within the browser, to be invoked ad-hoc by relevant built-in features of entity web sites.
[0024] However, known implementations still only use such unified communications plug-ins to initiate calls to the PSTN or enterprise PBX via telephony protocols such as SIP, and then route the call using traditional methods such as calling line ID or dialed number, along with either interactive voice response based authentication (e.g., "enter your identifier") or basic call routing (e.g., "press 1 for support, press 2 for sales"). Thus, such implementations still only use the browser as a gateway to the known traditional PSTN / PBX-based telecommunications model.
[0025] Thus, interlocutors (agents in a contact center, for one non-limiting example) receiving such calls must rely on some details about the calling user that their data systems may already hold, or even both, especially if such calls are the customer's very first real-time interaction with an entity associated with the website.
[0026] It is known to refine the captured data using analytical data processing tools that track customer behavior and use of online resources, such as, for example, Google Analytics®, which analyzes website browsing patterns, but the output of such tools is usually aggregated and anonymized data, which is not suitable for supporting an intuitive and personal experience for both calling users and their final interlocutors. Other tools (e.g., Marketo or HubSport) rank visitors based on their activity on the website. However, these tools do not apply in extending this personalization to contact centers due to the loss of context when customers switch their communication to telephone or other voice communication means. Other patent publications in the art may include U.S. Patent Application Publication No. 2004 / 039775 and WO 2014 / 071391, but the system and method of the present disclosure do not provide a satisfactory level of personalization for users interacting with websites.
[0027] FIG 1 illustrates one embodiment of a system in a networked environment, generally designated 100. The networked environment includes mobile data processing terminals 105, each capable of at least data communication with each other over a network, including a wide area network ("WAN") 111, such as the World Wide Web or the Internet. N and a static data processing terminal 110 N The present invention includes a plurality of data processing devices each comprising:
[0028] Each mobile data communications device 105 comprises a mobile telephone handset 105 capable of making and receiving wireless telecommunications over a cellular telephone network configured in accordance with Global System for Mobile Communications ("GSM"), General Packet Radio Service ("GPRS"), International Mobile Telecommunications 2000 (IMT-2000, "W-CDMA" or "3G"), International Mobile Telecommunications Protocol Advanced (ITU R-compliant, also known as "4G"), network industry standards, the telecommunications being carried out as voice, alphanumeric, or audio-video data using Short Message Service ("SMS") protocol, Wireless Application Protocol ("WAP"), Hypertext Transfer Protocol ("HTTP"), or Secure Hypertext Transfer Protocol ("HTTPS").
[0029] 105 for each mobile phone handset N The handsets 105 receive or transmit voice, text, audio and / or image data encoded as digital signals via wireless data transmission 106, which are relayed to or from the handsets, respectively, by the geographically closest communication link relay 107 or a plurality of them. The plurality of communication link relays 107 enable the digital signals to be routed between each handset 105 and their destinations by a remote gateway 108 via an MSC or base station 109. The gateway 108 is a communication network switch that couples the digital signal traffic between a wireless communication network, such as a cellular network, over which the wireless data transmission 106 takes place, and a wide area network 111. The gateway 108 further provides protocol conversion as necessary, for example, whether the handsets 105 use the WAP or HTTPS protocol to communicate data.
[0030] Alternatively, or in addition, a plurality of mobile data communication devices 105 NOne or more of the mobile data communication devices 105 may each have the capability of transmitting and receiving wired and / or wireless telecommunications via a wired local area network ("LAN") and / or a wireless LAN ("WLAN") conforming to the 802.11 standard ("Wi-Fi"). In a LAN or WLAN, telecommunications are similarly carried out as voice, alphanumeric, and / or audio-video data using Internet Protocol ("IP"), VoIP protocol, HTTP, or HTTPS, with signals relayed to and from the mobile data communication device 105, respectively, by a wired (LAN) or wireless (WLAN) router 109 that interfaces the mobile data communication device 105 to a WAN communication network 111. The mobile telephone handset 105 may have the capability of transmitting and receiving wireless telecommunications via WLAN, in addition to GSM, GPRS, W-CDMA, and / or 3G, ITU-R / 4G.
[0031] Exemplary handset 105 for use with system 100 N The mobile terminals 105 are preferably what are commonly referred to as "smartphones" (e.g., iPhones, Android phones, or equivalent handsets). In general, the mobile terminals 105 may be any portable data processing device having at least wireless communication means and audio recording and storage means. Thus, it will be readily understood from this disclosure by those skilled in the art that one or more of the mobile data communication devices 105 may alternatively be portable computers, commonly referred to as "laptops" or "notebooks," tablets, and the like.
[0032] Computer Systems
[0033] The system 100 includes one or more data processing terminals 110. NIn one embodiment, each of the various servers, controllers, switches, gateways, engines, and / or modules (collectively referred to as servers) of the illustrated figures is implemented via hardware or firmware (e.g., ASICs), as will be appreciated by those skilled in the art. Each of the various servers may be a process or thread running on one or more processors and in one or more computing devices (e.g., FIG. 2A, FIG. 2B) that executes computer program instructions and interacts with other system components to perform the various functions described herein. The computer program instructions are stored in a memory that may be implemented in the computing device using standard memory devices such as RAM. The computer program instructions may also be stored in other non-transitory computer readable media such as CD-ROMs, flash drives, etc. Those skilled in the art should appreciate that a computing device may be implemented via firmware (e.g., application specific integrated circuits), hardware, or a combination of software, firmware, and hardware. Those skilled in the art should also appreciate that the functionality of the various computing devices may be combined or integrated into a single computing device, or the functionality of a particular computing device may be distributed among one or more other computing devices without departing from the scope of the exemplary embodiments of the present invention. A server may be a software module that may also be referred to simply as a module. A set of modules within a contact center may include servers and other modules.
[0034] The various servers may be located on on-site computing devices in the same physical location as the contact center agents, or may be located off-site in different geographic locations, e.g., in a remote data center connected to the contact center via a network such as the Internet. Furthermore, some of the servers may be located in on-site computing devices at the contact center while other servers may be located in off-site computing devices, or servers providing redundant functionality may be provided via both on-site and off-site computing devices to provide better fault tolerance. In some embodiments, functionality provided by servers located on off-site computing devices may be accessed and provided via a virtual private network (VPN) as if such servers were on-site, or functionality may be provided using software as a service (SaaS) to provide functionality over the Internet using various protocols, such as, for example, by exchanging data using encoding in extensible markup language (XML) or JSON. In one embodiment, the servers may be located on computing or user devices associated with customers or users browsing the website.
[0035] 2A and 2B are diagrams illustrating an embodiment of a computing device, generally designated 200, that may be used in embodiments of the present invention. Each computing device 200 includes a CPU 205 and a main memory unit 210. As shown in FIG. 2A, computing device 200 may also include a storage device 215, a removable media interface 220, a network interface 225, an input / output (I / O) controller 230, one or more display devices 235A, a keyboard 235B, and a pointing device 235C (e.g., a mouse). Storage device 215 may include, but is not limited to, storage for an operating system and software. As shown in FIG. 2B, each computing device 200 may also include additional optional elements, such as a memory port 240, a bridge 245, one or more additional I / O devices 235D, 235E, and a cache memory 250 in communication with CPU 205. The input / output devices 235A, 235B, 235C, 235D, and 235E may be referred to collectively herein as 235.
[0036] The CPU 205 is any logic circuit that responds to and processes instructions fetched from the main memory unit 210. For example, the CPU 205 may be implemented in an integrated circuit in the form of a microprocessor, microcontroller, or graphics processing unit, or in a field-programmable gate array (FPGA) or application-specific integrated circuit (ASIC). The main memory unit 210 may be one or more memory chips that store data and allow any storage location to be directly accessed by the central processing unit 205. As shown in FIG. 2A, the central processing unit 205 communicates with the main memory 210 via a system bus 255. As shown in FIG. 2B, the central processing unit 205 may also communicate directly with the main memory 210 via a memory port 240.
[0037] In one embodiment, CPU 205 may include multiple processors and may provide functionality for concurrent execution of instructions or concurrent execution of a single instruction on one or more pieces of data. In one embodiment, computing device 200 may include a parallel processor having one or more cores. In one embodiment, computing device 200 comprises a shared memory parallel device having multiple processors and / or multiple processor cores that access all available memory as a single global address space. In another embodiment, computing device 200 is a distributed memory parallel device having multiple processors that each access only local memory. Computing device 200 may have both some memory that is shared and some memory that may be accessed only by a particular processor or subset of processors. CPU 205 may include a multi-core microprocessor that combines two or more independent processors in a single package, e.g., in a single integrated circuit (IC). For example, computing device 200 may include at least one CPU 205 and at least one graphics processing unit.
[0038] In one embodiment, CPU 205 provides single instruction multiple data (SIMD) capabilities, e.g., the ability to execute a single instruction on multiple pieces of data simultaneously. In another embodiment, some processors in CPU 205 may provide the ability to execute multiple instructions on multiple pieces of data simultaneously (multiple pieces of data (MIMD)). CPU 205 may also use any combination of SIMD and MIMD cores within a single device.
[0039] 2B illustrates an embodiment in which the CPU 205 communicates directly with the cache memory 250 via a secondary bus, sometimes referred to as a backside bus. In another embodiment, the CPU 205 communicates with the cache memory 250 using a system bus 255. The cache memory 250 typically has a faster response time than the main memory 210. As shown in FIG. 2A, the CPU 205 communicates with various I / O devices 235 via a local system bus 255. Various buses can be used as the local system bus 255, including, but not limited to, a Video Electronics Standards Association (VESA) local bus (VLB), an Industry Standard Architecture (ISA) bus, an Extended Industry Standard Architecture (EISA) bus, a Micro Channel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI Extended (PCI-X) bus, a PCI-Express bus, or a NuBus. For embodiments in which the I / O device is a display device 235A, the CPU 205 may communicate with the display device 235A via an Advanced Graphics Port (AGP). Figure 2B illustrates an embodiment of the computer 200 in which the CPU 205 communicates directly with the I / O device 235E. Figure 2B also illustrates an embodiment in which local bus and direct communication are mixed. The CPU 205 communicates with the I / O device 235D using the local system bus 255 while communicating directly with the I / O device 235E.
[0040] A wide variety of I / O devices 235 may be present in computing device 200. Input devices include one or more keyboards 235B, a mouse, a trackpad, a trackball, a microphone, and a drawing board, to name a few non-limiting examples. Output devices include a video display device 235A, speakers, and a printer. The I / O controller 230 shown in FIG. 2A may control one or more I / O devices, such as, for example, a keyboard 235B and a pointing device 235C (e.g., a mouse or optical pen).
[0041] 2A, computing device 200 may support one or more removable media interfaces 220, such as a floppy disk drive, a CD-ROM drive, a DVD-ROM drive, tape drives of various formats, a USB port, a Secure Digital or COMPACT FLASH™ memory card port, or any other device suitable for reading data from read-only media, or for reading data from or writing data to read-write media. I / O device 235 may be a bridge between system bus 255 and removable media interface 220.
[0042] The removable media interface 220 may be used, for example, to install software and programs. The computing device 200 may further include a storage device 215, such as one or more hard disk drives or hard disk drive arrays, for storing an operating system and other related software, and for storing application software programs. Optionally, the removable media interface 220 may also be used as a storage device. For example, the operating system and software may be executed from a bootable medium, such as a bootable CD.
[0043] In one embodiment, computing device 200 may include or be connected to multiple display devices 235A, each of which may be the same or different type and / or form. Thus, any of I / O devices 235 and / or I / O controller 230 may include any type and / or form of suitable hardware, software, or combination of hardware and software to support, enable, or provide for computing device 200 to connect to and use multiple display devices 235A. For example, computing device 200 may include any type and / or form of video adapter, video card, driver, and / or library for interfacing, communicating, connecting, or otherwise using display device 235A. In one embodiment, a video adapter may include multiple connectors for interfacing to multiple display devices 235A. In another embodiment, computing device 200 may include multiple video adapters, each video adapter connected to one or more of display devices 235A. In other embodiments, one or more of display devices 235A may be provided by one or more other computing devices connected to computing device 200, for example, via a network. These embodiments may include any type of software designed and constructed to use a display device of another computing device as a second display device 235A for computing device 200. Those skilled in the art will recognize and understand the various ways and embodiments in which computing device 200 may be configured to have multiple display devices 235A.
[0044] The embodiment of a computing device generally shown in Figures 2A and 2B may operate under the control of an operating system, which controls the scheduling of tasks and access to system resources. Computing device 200 may run any operating system, any embedded operating system, any real-time operating system, any open source operating system, any proprietary operating system, any operating system for mobile computing devices, or any other operating system capable of running on a computing device and performing the operations described herein.
[0045] Computing device 200 may be any workstation, desktop computer, laptop or notebook computer, server machine, computer on a steering wheel, mobile phone or other portable telecommunications device, media playback device, gaming system, mobile computing device, or any other type and / or form of computing, telecommunications, or media device capable of communicating and having sufficient processor power and memory capacity to perform the operations described herein. In some embodiments, computing device 200 may have different processors, operating systems, and input devices consistent with the device.
[0046] In other embodiments, computing device 200 is a mobile device. Examples may include a Java-enabled mobile phone or personal digital assistant (PDA), a smartphone, a digital audio player, or a portable media player. In one embodiment, computing device 200 includes a combination of devices, such as a mobile phone combined with a digital audio player or portable media player.
[0047] The computing device 200 may be one of multiple machines connected by a network or may include multiple machines so connected. A network environment may include one or more local machines, client nodes, client machines, client computers, client devices, endpoints, or endpoint nodes that communicate with one or more remote machines (which may also be generally referred to as server machines or remote machines) over one or more networks. In one embodiment, a local machine has the ability to function as both a client node seeking access to resources provided by a server machine and a server machine providing access to hosted resources for other clients. The network may be a LAN or WAN link, a broadband connection, a wireless connection, or a combination of any or all of the above. Connections may be established using a variety of communication protocols. In one embodiment, the computing device 200 communicates with other computing devices 200 through any type and / or form of gateway or tunneling protocol, such as Secure Socket Layer (SSL) or Transport Layer Security (TLS). A network interface may include a built-in network adapter, such as a network interface card, suitable for interfacing the computing device to any type of network with which it can communicate and performing the operations described herein. The I / O device may be a bridge between the system bus and an external communication bus.
[0048] In one embodiment, the network environment may be a virtual network environment in which various components of the network are virtualized. For example, the various machines may be virtual machines implemented as software-based computers running on physical machines. The virtual machines may share the same operating system. In other embodiments, different operating systems may run on each virtual machine instance. In one embodiment, a "hypervisor" type of virtualization is implemented in which multiple virtual machines run on the same host physical machine, each functioning as if they had their own dedicated box. Virtual machines may also run on different host physical machines.
[0049] Other types of virtualization are contemplated, such as, for example, the network (e.g., via Software Defined Networking (SDN)). Functions, such as the function of the Session Border Controller and other types of functions, may also be virtualized, such as, for example, via Network Functions Virtualization (NFV).
[0050] Terminal 105 N , 110 N Any user operating an entity 101 that offers a product or service accesses the entity's website 101 to obtain information about the entity and / or its products or services. Thus, the entity 101 operates static terminals 110 configured as web servers to deliver the website to requesting remote terminals, and at least one interlocator at the entity operates terminals to satisfy any ad-hoc telecommunications requirements of those users. The entity 101 also has the use of one or more interlocators at a remote contact center entity 102, which operates both static and mobile terminals for telecommunications support and assistance.
[0051] FIG. 3 shows in detail the data processing steps of an embodiment of the method performed in the environment of FIG. 1, FIG. 2A, and FIG. 2B using a user data processing terminal and an entity data processing terminal. In one embodiment, a local user application is served by an entity server to a user mobile terminal accessing an entity website, which exposes at least one application programming interface ("API") that passes user input data representing user interactions to an analysis and routing application hosted by the server 110. When the mobile terminal accesses the website on the server PC, the API is loaded, which then captures and communicates user interaction data substantially in real time to the server. The server application analyzes the communicated captured data and matches the user's interests derived from the analyzed interaction data with one or more interlocutors associated with the entity, which one or more interlocutors tend to interact with the user in any real time telecommunication that may follow the user's interactions with the website, for example, with a trained expert interlocutor related to such derived interests. Whenever a user is to initiate a real-time telecommunication, e.g., by selecting a "call", "chat", "video conference" (and equivalents) button within a web page, the server application routes any such telecommunication, preferably along with any analyzed and / or captured data, to the matched interlocutor's terminal.
[0052] Real-time telecommunications between the user and the matched interlocutor are routed as media paths using either alphanumeric (e.g., instant messaging), voice-only (e.g., telephone), or voice and video (e.g., Skype) formats within either VoIP, WebRTC, or PSTN network structures depending on how the user is accessing the website (e.g., desktop or laptop, tablet or smartphone).
[0053] Thus, in step 301, configuration parameters including interlocutor data, telecommunication formats, and routing options and rules should be entered initially at the server into a database processed by the server application. The interlocutor data includes characteristics of each interlocutor, such as name, area of training and / or topic expertise, keywords, and other logical and / or semantic distinguishing parameters, as well as the respective terminal type, communication and functional capabilities with respect to the above formats and structures, and at least one network address of each interlocutor terminal. Many such interlocutor records may be entered as there are potential interlocutors for the user to contact, and the recorded interlocutors may undergo further training to enhance their telecommunication aptitude, whereby new interlocutor records may be instantiated and the current new interlocutor records may be updated in step 302, whereby control logically returns to step 301.
[0054] In parallel with the above, at any given time, a user at the terminal may access the server in step 303 to load and peruse an entity website in a new browsing session. The server therefore creates a new user session record in the database in step 304 and makes access to the website resource conditional on a login authentication step in step 305. The user then inputs relevant authentication input data into the terminal in step 306, e.g. for a username and / or login credential data for the online social media resource. Such authentication input data is forwarded to the server in step 307 and recorded in the user session record.
[0055] The server can optionally make access to the website resource conditional on a secondary identity verification step in step 308. The user then enters relevant identity input data, for example at least a first name, into their terminal in step 309. Such identity verification input data is again forwarded to the server in step 310 and recorded again in the user session record.
[0056] The user then accesses and peruses website resources, including website data, website metadata, web page page tags, website cookies, website page headers, and / or user social media data, via traditional user page and link selections, at step 311. Thus, the user may, for example, generate a stream of selections using a mouse, known as a clickstream, and may also enter one or more search strings to locate website resources. The API captures and communicates this interaction data in substantially real-time, at step 311, as website navigation records representing historical or sequential clicks, whether as interaction event-driven updates or in aggregate form, or as a combination of both depending on, for example, bandwidth or load constraints.
[0057] The communicated captured interaction data is received by a server application at the server and analyzed in step 312 using various statistical predictive analysis techniques, such as Bayesian inference or regression models, known in the art. The user's interaction patterns are determined by the server application, and an interaction prediction is output, such as when a customer may start interacting or stop perusing a web page or website. The interaction prediction is used to trigger a matching operation in step 313 of the user with each recorded interactor, which is achieved by comparing the interaction patterns with interactor parameters, particularly areas of training and / or topic expertise and keywords, and determining the recorded interactor with the closest record to which the interaction patterns correlate.
[0058] In step 314, a matching interlocutor record is selected and a routing path between the user and the selected interlocutor is determined based on the network address of the interlocutor terminal recorded in the database, so that if the user should initiate a real-time telecommunications call from the website at substantially the time according to the prediction in step 315, the server receives a corresponding event call message in step 316 and the server application routes the call to the selected interlocutor in step 317 along with either the user's interaction pattern or the captured and communicated user interaction data, or both.
[0059] The server application in the server is a multi-threaded application that tends to execute each of steps 301-302, 304, 307, 310, 312-314, and 316-317 substantially simultaneously for multiple interacting users and multiple interlocutors, and steps 312-314 themselves form a looping subroutine of the application, whereby if the user does not initiate the interaction predicted in step 315, the user's subsequent captured and communicated interactions that move the shift point of interest will be processed in the next iteration of steps 312-314, potentially resulting in different matched interlocutors, each having a different but more relevant set of parameters that optimally correlate with the shifted point of interest.
[0060] 4 is a block diagram illustrating a high level implementation of the method of FIG. 3 within the system 100 in the environment of FIGS. 1, 2A, and 2B at run time. The steps of the method include a user mobile terminal 105 exposing one or more APIs to a website of an entity 101. N The system 100 may be implemented as a user application 401 processed in conjunction with a browser of the interlocutor 110, a server application 402 processed by the entity terminal 110 having an associated highly scalable database, and an interlocutor application, in its simplest form as a browser processed by the interlocutor's mobile or static terminal, to which user interactions are routed by the server application 402, which may be interpreted as an interactive management engine.
[0061] The user application 401 captures website data 404 such as cookies, web page headers, etc., as well as social media data 405 collected when the user logs into the website using social media-based authentication in step 306, and terminal data 406 indicative of at least the terminal type, as the user interacts with the mobile terminal 105, and transmits the captured data to the server application 402 according to step 311.
[0062] The server application 402 is an analysis and routing application, including a database, an analysis module 407, a routing module 408, and a parameterization module 409, that processes the captured user data it receives to identify relevant interlocutors associated with the entities 101 deemed most relevant, and route the user's final telecommunication thereto. The server application performs steps 301-302, 304-305, 307-308, 310, 312-314, and 316-317 (FIG. 3).
[0063] The parameterization module 409 is specifically used to execute steps 301 and 302, i.e. to define the attributes of each interlocutor associated with the website data from a logical and semantic point of view, the attributes of the interlocutor terminal, and the attributes of the telecommunication link with the user terminal, e.g. terminal type, associated communication capabilities, minimum bandwidth level to guarantee quality of service for each communication type.
[0064] The analytics module 407 specifically performs steps 312-314, i.e., receives all of the user data captured by the user application 401, which thus includes, but is not limited to, website data, website metadata, page tags for web pages, website cookies, website page headers, one or more clickstreams, one or more search strings, website navigation records representing historical or sequential user page and link selections, user social media data, and the like.
[0065] The routing module 408 specifically performs steps 316 and 317. For example, it is triggered by a telecommunication event and uses the output of the analysis module 407, and optionally also the captured user interaction data received according to step 311, to route the user interaction to a selected interlocutor terminal.
[0066] The server application 402 configures the terminal to perform at least the processing steps 301-302, 304-305, 307-308, 310, 312-314, and 316-317 as described above, which is interfaced with the device's operating system and network communication subroutines via one or more suitable APIs. The server application thus includes an analysis module 407, a routing module 408, and a parameterization module 409 as described above, and thus tends to request and obtain authentication data, identity verification data, and captured interaction data from each remote user accessing the website resource. The website resource may be stored in a database, and a user session record may be stored, including the terminal type and capacity, telecommunication rules, and interlocutor profiles for the interlocutors and users as established in steps 301 and 302, and the captured interaction data for each remote user accessing the website resource.
[0067] Server application data specific to the analysis module 407 may include user analysis data such as interaction patterns processed in step 312 and user-interlocutor pairs matched according to step 313. Server application data specific to the routing module may include user terminal-interlocutor terminal communication paths output by the analysis module 407 for perusal by the routing module. The memory 210 may further include conventional local and / or network data not related to server applications, for example, for use by one or more additional applications and / or the operating system.
[0068] The browser application data may include the downloaded support application resources and the GUI that is output to the display 235A and in which the downloaded resources are rendered. In the case of an interlocutor terminal, the browser application data may also include the analyzed user data (consisting of the output of the server application communicated to the interlocutor terminal 105 in step 317). In an alternative embodiment, the remote server 110 may further or alternatively forward some or all of the captured user data received by the remote server prior to analysis, and thus the browser application data of the interlocutor terminal may further or alternatively include some or all of this user data. The analyzed and / or captured data may be forwarded by the remote server for ad-hoc reference purposes or for other purposes, so that the interlocutor has at hand a wealth of interaction support information available to facilitate electronic communication with the user.
[0069] In one embodiment, the captured data may be analyzed based on the patterns and used by the analytics module 407 to predict a successful outcome between the first user and a selected second user, i.e., an interlocutor or agent acting on the website being browsed by the first user. A set of interaction management processes may be utilized that take a set of inputs including raw user online data (website data, behavioral data, demographic data, social data, mobile data, etc.), hard constraint configurations (agent and queue assignments, rules that ensure certain business-defined constraints are met, call routing or interaction offers based on user online data, etc.), and various forms of analytical data such as the need or likelihood to achieve a particular outcome (which may change dynamically during the customer journey) (e.g., smart attributes (need for support, propensity to complete an outcome) derived based on analysis of the raw data using machine learning algorithms, persona clusters that dynamically group users based on their behavior and demographics, and journey patterns that represent similar sets of sequences of user actions). The particular outcomes may include product purchase, booking a demo, requesting support, likelihood to accept an interaction offer, etc. These processes can be used to make interaction management decisions that may include identifying the most valuable customers within a persona set given the outcomes to be maximized, smart call routing based on agent efficiency clusters, interaction offers to agents (what type of interaction should be offered to which user and when, e.g. chat / audio / video), and automatic interaction offers focused on shaping customer journeys and optimizing business outcomes.
[0070] The analytics module 407 takes in inputs and processes the data using analytics data. The output may be machine-learned "smart attributes" that rank customers or actions against outcomes defined by the business and within persona clusters assigned by machine learning or specified by the business associated with the website. The output may be interaction recommendations. The routing module 408 may be configured to make interaction recommendations to the agent based on the provided inputs, providing the agent with contextual data learned about the user (e.g., likelihood to accept an interaction offer, likelihood to purchase a product). For example, the analytics module 407 may identify users to whom the agent should offer a chat interaction to increase product sales. This allows the agent to focus on users or customers with high predictive value for a given outcome, thus optimizing resource utilization.
[0071] The output may also include automatic interaction offers. The analytics module 407 can make informed decisions and automatically offer interactions to selected users. Such interaction offers may be made based on the analytics data to optimize business-defined outcomes. First, using the input data (as described above), the component identifies a set of high-value customers with respect to a given outcome. Then, appropriate treatments (i.e., interaction offers) are identified using predictive models built using various machine learning algorithms, including decision trees, Bayesian inference, and neural networks. The analytics module 407 can be configured to determine how many interaction offers should be automatically made to customers to maximize agent utilization and minimize user or customer wait times.
[0072] The routing module 408 makes routing decisions and performs agent selection based on the inputs provided (as described above) using machine learning inference based on persona attributes (i.e., age, location, role, gender, etc.) and activities (i.e., page views, calls, chats, etc.) Customers with similar personas and activity patterns will do better with agents like Agent X, Customer X should have an agent reserved because they are about to make a call or the system has offered them an interaction and they are likely to accept, etc.
[0073] Journey Event Expression
[0074] During the processing of user data, in order to apply state of the art machine learning algorithms to data logs corresponding to end-user interactions with a business (e.g., a user's activity stream on a website), each user activity, referred to as an event, needs to be represented as a vector. An event may also include an activity by an entity, such as a system, a user, or an agent, that triggers a front-end or back-end change on a web page. In one embodiment, events are originally recorded as objects represented by a large number of attribute-value pairs, where the attributes representing the event vary with the type of event (e.g., page view, form fill, etc.). A simple method of feeding these objects to a machine learning algorithm requires a significant amount of manual effort by data scientists in selecting important attributes (also referred to as feature engineering). In creating a representation of events as vectors (also referred to herein as "event2vec"), raw events can be input to the algorithm and output as a vector representation of the data in an end-to-end automated manner. The resulting vectors are dense and rich in terms of the information they contain. The architecture for creating a representation of events as vectors can handle multi-modal events (web clicks, calls, SMS, email, chat, etc.). Events of different modalities can also be captured using different schemas. Thus, the method for creating vectors from raw events can be schema agnostic. As a result, the same system that applies event2vec can be used to generate event vectors for a variety of machine learning based predictive systems that train on interaction data.
[0075] The vector representation of events may be used in numerous analysis and prediction systems, such as those described above in Figures 3 and 4. In one embodiment, the captured data may be analyzed based on patterns and used by the analysis module 407 to predict a successful outcome between a first user (e.g., a customer browsing a website) and a selected second user (e.g., an interlocutor or agent acting on the website being browsed by the first user). In another embodiment, the captured data may be analyzed based on patterns and used to predict a successful outcome for the first user and an action to achieve that outcome, such as automatically presenting an offer to the first user or making a recommendation to the agent. In another embodiment, the predicted outcomes may be presented to the agent as a ranked list, where the agent can manually select a desired outcome based on information provided (such as the probability and likelihood of achieving a particular outcome by the customer). The agent may also be able to select a customer from multiple customers on the website and present an interaction to the customer most likely to generate the desired outcome based on the analysis.
[0076] In yet another embodiment, outcome prediction can be used by a business to manually or automatically anticipate when a customer browsing a website will need agent support based on the rise and fall of outcome probabilities. For example, the event vectors can be used for time sequence modeling of customer journeys using machine learning algorithms such as recurrent neural networks (RNNs) (described in more detail below), which can be used to predict outcomes on a business website (e.g., sales probability during a user session). The type of outcome can vary with business type. For example, "Place an order" can be an outcome for an e-commerce business website, while "Submit a mortgage application" can be an outcome for a banking website.
[0077] In yet another embodiment, events represented as vectors can also be used in analysis of user case studies, such as observing similarities and relationships in browsing behavior by plotting events as points in a graph.
[0078] As mentioned above, customer journey data may be captured in units called events. Each event includes information such as a unique ID and a timestamp along with attributes that capture information such as, but not limited to, a customer profile, a customer's interactions with a business, online browsing behavior on a website, etc. Customer journey data may further include information such as who the customer is, where the customer is located, what hardware and / or software the customer is using, what items the customer selected on the website (or what buttons they selected), what the customer filled out on a web form, the keywords the customer used to perform a search, whether the customer interacted with an agent, and other details surrounding any such interaction (e.g., customer ranking of the interaction, transcript, outcome, etc.).
[0079] In one embodiment, the use of supervised machine learning methods for predictive analytics requires examples drawn from existing customer journeys that contain the information to be predicted (e.g., whether a business outcome was achieved). The examples are used to train a model that can make predictions for new customer journeys. The examples include raw data logs of customer journeys that need to be heavily processed before being input into the machine learning algorithms for model training and predictions. Additionally, the processing steps may vary depending on the raw data format.
[0080] In one embodiment of representation learning, real-world entities can be represented as vectors by training a predictive model on a secondary predictive task rather than the primary task of outcome prediction, where suitable event attributes are identified to be predicted. Inputs can be collected from entities, and optimal vector representations can be learned for these entities during the training phase. Neural networks can be used to learn dense vectors that are typically used in the prior art to represent real-world entities that are not otherwise recorded as multi-dimensional objects (e.g., words). In one embodiment, representation learning can also be used to define smaller and denser vector spaces for entities that are otherwise recorded using a larger number of dimensions / attributes in a very sparse vector space (e.g., events), and therefore computationally expensive and inefficient to process as input to a machine learning model.
[0081] Events are highly heterogeneous since they contain information (or attributes) of different semantic types (such as timestamps, URLs, search keywords, quantities, etc.). In one embodiment, a method for converting events to a vector representation maps events of different semantic types as vectors in a common n-dimensional space, and mathematically represents each event as an array of numbers, where the size of the array corresponds to the number of dimensions. An embodiment of the method herein includes training a predictive neural network model to predict values of selected event attributes and extracting event vectors from a hidden layer of the trained neural network.
[0082] To train a model for use in event2vec, several pre-processing steps are required to convert the raw customer journey data captured from business websites into a format that can be consumed by the algorithm during analysis of the data. Event logs from a customer's browsing session can be stored as a json object with a set of predefined attributes and the type of values recorded under these attributes. Each event is given a unique ID and timestamp. A common schema is used to capture event data from different businesses. It is within the scope of this embodiment for any number of businesses and their data to be used for analysis.
[0083] In one embodiment, the event data may be stored using public cloud storage resources, such as Amazon S3, or on private cloud resources. An example of a suitable format for storing larger volumes of events may be a dataset, which can be read and queried quickly using big data processing tools such as Spark. Large volumes of event data may be stored as multiple datasets with a common format or schema. In one embodiment, a data frame includes rows and columns, where each row represents a unique event with a unique event ID, and the columns represent event attributes. A customer journey ID may be captured in a visit ID attribute, so multiple events may have the same visit ID. The event data may be queried and sliced into time intervals (e.g., monthly) for ease of processing. Examples of event attributes may include ID, account, event type, when the visit was created (timestamp of first activity in the visit), timestamp indicating when the visitor should be considered to have left, customer or visitor cookie ID, ID of a known (authenticated) customer, URL of the last page visited, list of search terms used in the web visit, list of persona IDs to which the customer was assigned due to activities performed in the web visit, list of outcome IDs achieved due to activities performed in the web visit, search engine newsletter name or other marketing campaign source identification, user agent string, browser version, preferred language of the browser, true flag for mobile devices, operating system family, country, visitor's time zone, visitor's IP address, etc. If the attribute does not have a value, the assigned value may be null or unknown. The data type of the attribute may also be indicated, e.g., string, number, or Boolean.
[0084] Pre-processing is performed on the raw data to obtain a dataset for the event2vec algorithm. The event2vec algorithm requires the original events in the data (as opposed to outcome prediction, described further below, which requires data in sequence format). Pre-processing is customized to the problem and data. Machine learning requires the data to be reduced to a specific output, which is achieved using the pre-processing steps outlined below.
[0085] Remove or mask / encrypt all data columns that contain personally identifiable information (PII), which is a way to anonymize information to comply with applicable law.
[0086] Identify the columns (class attributes) that need to be predicted when training the model.
[0087] Identify and remove columns that duplicate class attributes. These are often duplicated with repeated information in the dataset.
[0088] The arrayed columns are removed. This simplifies the process, but in one embodiment, they do not have to be removed. If the arrayed columns are not removed, a more complicated method of pre-treatment is required.
[0089] Convert all boolean columns to integer type and replace true with 1 and false with 0. This is for formatting consistency since the algorithm works with numeric integers.
[0090] Replace all null and unknown values with 0, 0.0, and unknown for integer, double, and string types, respectively.
[0091] In one embodiment, the steps are performed in sequential order. In one embodiment, the resulting data frame contains the same number of rows as the input data frame, but fewer columns, and does not contain null values in any of the columns.
[0092] Once the pre-processing is performed, a data set is obtained that can be applied to a machine learning algorithm for training and validation. FIG. 5 illustrates an embodiment of a system (neural network) for converting events into a vector representation, shown generally. The system can be present in the analysis module 407. In an embodiment, the system shown in FIG. 5 can be obtained by training a feed-forward neural network. The system can then be used to predict the event class 507 of an input event. The event class can be originally recorded as an attribute in the raw event data. In an embodiment, the event can belong to a large number (e.g., 34) of classes, such as page views, form fills, interaction acceptance, etc. This feed-forward network includes two input layer branches. The first branch can accept categorical attributes 501, and the second branch accepts attributes that include continuous numerical attributes and / or continuous values 503. The categorical attributes can be defined as having a type of string. The categorical attributes must first be converted as vectors before being input to the system. As a result, the categorical attribute branch can include an extra layer 502 (e.g., an embedding layer). In one embodiment, the numeric attributes 503 may be used without transformation. Once the inputs 501 containing categorical attributes are transformed in the embedding layer 502, they may be combined with the numeric attribute inputs 503 in the hidden layer 504A. The neural network is then optimized through an additional hidden layer 504B to create an output layer 506, the output of which is used to predict the class of the event 507. After the network is trained, a dense event vector representation 505 can be extracted (using standard algorithms) from the hidden layer 504B. The dense event vector 505 may be used for outcome prediction (described in more detail below).
[0093] After a model has been trained in the event2vec system, it can accept any event and output an n-dimensional event vector. In one embodiment, the events are represented in the dataset as vectors in a common n-dimensional space. An additional event vector column may be added to the event data frame that holds the vector output of the event2vec model for the corresponding event in the row. The new dataset can then be used for further analysis, such as outcome prediction, which is described in more detail below.
[0094] Consequence prediction In one embodiment, the customer journey data includes sequence data, where each journey can be a sequence of events arranged in time. Not all machine learning algorithms are capable of incorporating and utilizing sequential information in such data, which is usually important for accurate predictive modeling. Sequence learning includes a class of machine learning algorithms that can incorporate sequential data and encode temporal information in the learning model. RNNs include a class of neural networks that can operate directly on sequences as input. Specifically, by way of non-limiting example, long short-term memory (LSTM) networks are best suited to handle longer sequences. RNNs can input sequential data by processing one unit of the sequence at a time and updating its encoding of the sequence state. RNN cells can hold context information in their internal memory as they process units in the sequence.
[0095] In one embodiment, an interlocutor (e.g., an agent, contact center employee, website host, application provider, etc.) can set positive and negative outcomes that they wish to achieve / avoid, where simple outcomes are "purchase" or "support ticket closure" and more complex outcomes are "openness to sales involvement". A machine learning component (e.g., analytics module 402, etc.) can determine which customers are likely to complete the desired outcome when offered an interaction (e.g., chat, call, email, proposal, video chat, etc.). The machine learning component or module captures data in real-time and predicts the probability of a successful outcome between a first user (e.g., a customer) and a selected second user (e.g., a contact center agent) in real-time. In one embodiment, the machine learning component or module includes a classification algorithm that learns a complex classification function / model on the captured event sequences, stores the learned function / model, classifies new event sequences, and provides a confidence score for the predicted class. The classes are "achieved outcomes" and "not achieved outcomes" and the confidence score is the probability of the outcome.
[0096] Consequences may be captured in binary form at the end of an interaction by either the agent or the system, which retrieves them programmatically from a CRM or ticketing system where the agent entered the consequence, or when the user completes a sequence of actions that define the consequence. This forms a feedback loop for the system to understand the persona / activity patterns that drive positive vs. negative consequences. In one embodiment, consequences may be presented in an ordered list as potential next best actions for the agent to take, and may be modified to include outcome predictions. As a result, the system, or the agent (who may manually intervene), can offer a chat or call or other action to a customer who wants to achieve a particular consequence, but may be at risk of not doing so. This is important because the pool of agents is always limited. Thus, the agent's time can be directed to the types of interactions with customers that have the highest chance of maximizing the defined consequence. In the case of an agent making an offer, machine learning can be applied to present the agent with a ranked list of customers or actions for a particular outcome regarding the discretion of which customer to present a particular interaction (whether a call, chat, etc.).
[0097] In one embodiment, a more specific type of neural network known as an RNN may be used in the outcome prediction algorithm. RNNs comprise a major class of deep learning algorithms and offer particular benefits for customer journey modeling. RNNs, and in particular LSTMs, can identify relationships between elements of long time sequences and model these as complex mathematical functions. Unlike standard feed-forward neural networks, LSTMs have feedback connections that allow processing of entire sequences of data. Manual analysis of customer journeys to identify data characteristics important for outcome prediction is consequently not necessary. RNNs also allow visualization of changes in prediction accuracy as customer journey sequences grow. In addition, interpretations and inferences may be gleaned from the predictions of individual businesses.
[0098] The RNN-based classification algorithm is composed of RNN cell units and may also be combined with dense-type neuron units. FIG. 6 illustrates an embodiment of a process for capturing an event vector representation, shown generally. In an embodiment, the event vector may be that created in FIG. 5. A customer journey 605 occurs over some length of time 606, and events 610 occur in real time. In FIG. 6, specific events provided for illustrative purposes include "Ad Click" 610a, which begins at timestamp "2019-01-11 08:23" 615a, "Homepage View" 610b, which begins at timestamp "2019-01-11 08:24" 615b, "Chat Start" 610c, which begins at timestamp "2019-01-11 09:00" 615c, and "Add to Cart" 610d, which begins at timestamp "2019-01-11 09:45" 615d. Event 610d "add to cart" may additionally be classified as a consequence event, with adding the item to the cart being the desired consequence. It is within the scope of the embodiments provided herein that any number of events may be used, and those provided herein are not intended to be limiting, but are for simplicity. Consequence predictions may also be obtained for each incremental event 610 across the customer journey 605. At each incremental event in the customer journey, the RNN classifier is invoked to make a prediction based on all sequences of events that have occurred in the customer journey up to the current time.
[0099] The input layer and hidden layer include an RNN classifier cell 625 (which is further described in detail in FIG. 7 below). An example of an RNN classifier cell may be an LSTM. The event2vec algorithm is applied to the customer journey data to obtain an event vector 620 for each event (620a, 620b, and 620c). The input layer takes in the corresponding event vectors 620a, 620b, and 620c in timestamp order. The vectors are maintained and updated to summarize the state of the customer journey. A cumulative vector 630 may summarize all previous event vectors. The RNN classifier cells 625a, 625b, and 625c take in the vectors and generate outcome probabilities 635a, 635b, and 635c at each event 610a, 610b, and 610c of the customer journey.
[0100] In order to train the outcome prediction model and obtain outcome predictions for new data, the data needs to be in the format of class-labeled sequences, so data pre-processing is required. The pre-processing of the raw event log to obtain the labeled sequence dataset is as follows:
[0101] Identifies the set of outcome IDs for a given business.
[0102] Define a dataframe schema with columns capturing the visit ID, a time-ordered sequence of event IDs within a given visit, and a class indicator.
[0103] Query the raw events dataframe (the input dataset to the event2vec system) to obtain the sequence dataset defined above, where the class indicator column holds a binary value depending on whether the outcome was achieved during the visit or not.
[0104] Remove the consequence-achieving event from the event sequence.
[0105] Use the event2vec system to replace event ids in the sequence column with their respective event vectors.
[0106] The resulting output data frame has fewer rows and a different set of columns relative to the input event data frame. In addition, the output should be in a standard format for the RNN to accept data. The output data is used to train and validate the algorithm for outcome prediction. FIG. 7 illustrates an embodiment of a system of continuous neural networks as applied in FIG. 6, shown in full. The output layer includes a dense neuron layer that inputs the RNN (or more specifically, in one embodiment, LSTM) processed sequence as a single vector. A softmax function is applied to the output of the dense layer. The network is then trained using a labeled sequence dataset. The labeled sequence dataset is obtained from the raw event dataset transformed as described above. The labels encode the information that an outcome was or was not achieved in a given visit. The system shown in FIG. 7 can be used for outcome prediction. The input data, including data of the customer journey, is represented as a set of vectors. The input layer populates, maintains, and updates vectors that are representations of the previous events in the sequence in the order of their timestamps in the event vector. As seen in FIG. 7, each event in the sequence has an RNN cell associated with it. Each RNN cell outputs an intermediate vector representation to be passed to the next RNN cell that summarizes the state of the customer journey up to the currently processed time step. Using the trained model, an event vector sequence is ingested and output in binary form (e.g., 1 or 0, respectively) corresponding to an outcome that has been achieved or not achieved.
[0107] The outcome prediction use case applies to organizations with multiple lines of business, eliminating the need for extensive manual effort to identify important event attributes for predictive modeling. The system has general-purpose capabilities, especially for new customers who may not have journey examples present in their training or historical data.
[0108] While the invention has been illustrated and described in detail in the drawings and foregoing description, it is to be considered as illustrative and not of limiting character, and it is to be understood that only the preferred embodiment has been shown and described, and that all equivalents, changes and modifications which come within the spirit of the invention as set forth in this specification and / or the following claims are desired to be protected.
[0109] Accordingly, the appropriate scope of the present invention should be determined only by the broadest interpretation of the appended claims so as to embrace all such modifications and all equivalent relationships as illustrated in the drawings and described herein.
Claims
1. A method for machine learning, by a computer system, of transforming raw customer journey event data from multiple customer browsing sessions on a website into an ingestible dataset to obtain vector representations of web events, the computer system comprising: a. logging and storing events from the browsing session, the events including activity data from the plurality of customers; b. pre-processing the raw customer journey event data captured on the website, comprising: i. removing or masking / encrypting values in all columns that contain personally identifiable information; ii. Identifying class attributes that need to be predicted during model training; iii. Identifying and removing columns that duplicate class attributes; iv. Removing the array type data; v. Removing the timestamp data; vi. converting all boolean columns to integer types; vii. replacing all null and unknown values; c) obtaining said ingestable data set that is applicable to an algorithm to obtain a vector representation of web events.
2. The method of claim 1 , wherein the events are stored in a common schema.
3. 2. The method of claim 1, wherein converting all Boolean columns to integer types further comprises replacing "true" with 1 and "false" with 0.
4. 2. The method of claim 1 , wherein replacing all null and unknown values further comprises replacing all null and unknown values with 1 for integer types, 0.0 for double types, and "unknown" for string types.
5. The method of claim 1 , wherein the pre-processing steps are performed in sequential order.
6. A method for machine learning, by a computer system, of transforming raw customer journey event data from browsing sessions of a plurality of customers on a website into an ingestible dataset for obtaining vector representations of web events, the computer system comprising: a. logging and storing events from the browsing session, the events including activity data from the plurality of customers; b. pre-processing the raw customer journey event data captured on the website, comprising: i. removing or masking / encrypting values in all columns that contain personally identifiable information; ii. Identifying class attributes that need to be predicted during model training; iii. Identifying and removing columns that duplicate class attributes; iv. Removing the array type data; v. Removing the timestamp data; vi. Converting all Boolean data to integer type; vii. replacing all null and unknown values; c. obtaining said ingestable data set that can be applied to an algorithm to obtain a vector representation of web events; d. training a feedforward neural network using the accessible data set; e. inputting the web events into the feedforward neural network, where the output includes a vector representation of each activity within the web events; f) adding a vector column to the dataset for each said activity within said web event.
7. The method of claim 6 , wherein the vector representations comprise a common n-dimensional space.
8. The method of claim 6 , wherein the training comprises predicting an event class of an input.
9. The method of claim 8 , wherein the trained feedforward neural network includes multiple input layer branches.
10. The method of claim 9 , wherein the plurality of input layer branches further comprises at least a first branch that accepts categorical attributes and a second branch that accepts continuous, numerical attributes.
11. The method of claim 6 , wherein the events are stored in a common schema.
12. 7. The method of claim 6, wherein converting all Boolean data to integer type further comprises replacing "true" with a one and "false" with a zero.
13. 7. The method of claim 6, wherein replacing all null and unknown values further comprises replacing all null and unknown values with 1 for integer types, 0.0 for double types, and "unknown" for string types.
14. The method of claim 6 , wherein the pre-processing steps are performed in sequential order.
15. The method of claim 6 , wherein each event includes a schema that includes a characteristic and a modality.
16. a. inputting the added data set into a trained neural network including LSTM cell units and high-density neuron units, the input layer and hidden layer of the neural network including LSTM cell units and the output layer including the high-density neuron units; b. obtaining a processing sequence input to the high-density neuron unit as a single vector; c. applying a softmax function to the single vector; 9. The method of claim 6, further comprising the step of: (a) obtaining a consequence probability of said vector.
17. The neural network is trained using preprocessed data, and the method for preprocessing comprises: a. identifying a set of consequential identities for a given entity; b. Defining a data frame schema that captures visit IDs, time-ordered sequences of event IDs within a given visit, and class indicators; c. querying an event data frame to populate said defined data frame; d. removing a consequent event from the time-ordered sequence; e. replacing the event ID with the respective event vector.
18. The method of claim 16 , wherein the input layer accepts inputs in timestamp order.
19. The method of claim 16 , wherein the single vector contains a representation of a previous event in the sequence.
Citation Information
Patent Citations
Associating a Web Session with a Household Member
US20140095413A1
Information processing device, information processing method, program, and system
WO2017168883A1
