Multiple output relaxation machine learning model

[0001] The embodiments discussed herein are related to a multiple output relaxation (MOR) machine learning model.

[0002] Machine learning is a form of artificial intelligence that is employed to allow computers to evolve behaviors based on empirical data. Machine learning may take advantage of training examples to capture characteristics of interest of their unknown underlying probability distribution. Training data may be seen as examples that illustrate relations between observed variables. A major focus of machine learning research is to automatically learn to recognize complex patterns and make intelligent decisions based on data.

[0003] One main difficulty in machine learning lies in the fact that the set of all possible behaviors, given all possible inputs, is too large to be covered by a set of training data. Hence, a machine learning model must generalize from the training data so as to be able to produce a useful output in new cases.

[0004] One example of machine learning is traditional structured prediction (SP). Traditional SP is a single model approach to dependent output. With SP, once an input feature vector x is specified, a single correct output vector z can be fully specified. Thus the output vector z is fully conditioned on the input feature vector x and the different output components of output vector z (z₁, z₂, . . . ) are conditionally independent of each other given the input feature vector x. Thus, the probability of z₁ given x is equal to the probability of z₁ given x and z₂, or p(z₁|x)=p(z₁|x, z₂). However, traditional SP cannot handle an interdependent relationship between different output components. In addition, traditional SP cannot handle a problem having multiple correct output decisions for a given input.

[0005] The subject matter claimed herein is not limited to embodiments that solve any disadvantages or that operate only in environments such as those described above. Rather, this background is only provided to illustrate one example technology area where some embodiments described herein may be practiced.

[0006] In general, example embodiments described herein relate to methods of employing a multiple output relaxation (MOR) machine learning model to predict multiple interdependent output components of a multiple output dependency (MOD) output decision. The example methods disclosed herein may be employed to solve MOD problems.

[0007] In one example embodiment, a method for employing an MOR machine learning model to predict multiple interdependent output components of an MOD output decision may include training a classifier for each of multiple interdependent output components of an MOD output decision to predict the component based on an input and based on all of the other components. The method may also include initializing each possible value for each of the components to a predetermined output value. The method may further include running relaxation iterations on each of the classifiers to update the output value of each possible value for each of the components until a relaxation state reaches an equilibrium or a maximum number of relaxation iterations is reached. The method may also include retrieving an optimal component from each of the classifiers.

[0008] In another example embodiment, a method of employing an MOR machine learning model to predict multiple interdependent output components of an MOD output decision may include training a first classifier to predict a first of two interdependent output components of an MOD output decision based on an input and based on the second component. The method may also include training a second classifier to predict the second of the two components of the MOD output decision based on an input and based on the first component. The method may further include initializing each possible value for each of the components to a predetermined output value. The method may also include running relaxation iterations on each of the classifiers to update the output value of each possible value for each of the components until a relaxation state reaches an equilibrium or a maximum number of relaxation iterations is reached. The method may further include retrieving an optimal component from each of the classifiers.

[0009] In yet another example embodiment, a method of employing an MOR machine learning model to predict multiple interdependent output components of a lead response management (LRM) MOD output decision may include training a first classifier to predict a first of two interdependent output components of an LRM MOD output decision based on an input and based on the second component. The method may also include training a second classifier to predict the second of the two components of the LRM MOD output decision based on an input and based on the first component. The method may further include initializing each possible value for each of the components to a predetermined output value. The method may also include running relaxation iterations on each of the classifiers to update the output value of each possible value for each of the components until a relaxation state reaches an equilibrium or a maximum number of relaxation iterations is reached. The method may further include retrieving an optimal component from each of the classifiers.

[0010] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory and are not restrictive of the invention, as claimed.

[0024] Some embodiments described herein include methods of employing a multiple output relaxation (MOR) machine learning model to predict multiple interdependent output components of a multiple output dependency (MOD) output decision. The example methods disclosed herein may be employed to solve MOD problems.

[0025] As used herein, the term “multiple output dependency” or “MOD” refers to an output decision, or a problem having an output decision, that includes multiple output components which are interdependent in that each component is dependent not only on an input but also on the other components. Some example MOD problems include, but are not limited to: 1) which combination of stocks to purchase to balance a mutual fund given current stock market conditions, 2) which combination of players to substitute into a lineup of a sports team given the current lineup of the opposing team, and 3) which combination of shirt, pants, belt, and shoes to wear given the current weather conditions. In each of these examples, each component of the output decision depends on both the input (current stock market conditions, an opposing team lineup, or current weather conditions) and the other components (the other stocks purchased, the other substituted player, or the other clothing selected). Other examples of MOD problems may relate to hostage negotiations, retail sales, online shopping carts, web content management systems, customer service, contract negotiations, or crisis management, or any other situation that requires an output decision with multiple interdependent output components.

[0026] Another example MOD problem is lead response management (LRM). LRM is the process of responding to leads in a manner that optimizes contact or qualification rates. Leads may come from a variety of sources including, but not limited to, a web form, a referral, and a list purchased from a lead vendor. When a lead comes into an organization, the output decision of how to respond to the lead may include multiple interdependent components such as, but not limited to, who should respond to the lead, what method should be employed to respond to the lead, what content should be included in the response message, and when should the response take place. Each of these components of the output decision depends on both the input (the lead information) and the other components. For example, the timing of the response may depend on the availability of the person selected to respond. Also, the content of the message may depend on the method of response (e.g. since the length of an email message is not limited like the length of a text message). Although the example methods disclosed herein are generally explained in the context of LRM, it is understood that the example methods disclosed herein may be employed to solve any MOD problem.

[0027] Example embodiments will be explained with reference to the accompanying drawings.

[0028] FIG. 1 is a schematic block diagram illustrating an example LRM system 100. As depicted, the example LRM system 100 includes various components such as a public switched telephone network (PSTN) 110, user communication and/or computing devices 112, a TDM gateway 120 connecting the PSTN 100 to an internet 130, remote agent stations 121, workstations 128, a call center 140, an internet gateway 150 connecting a local area network 160 to the internet 130, a web server 170, a contact server 200, a lead data server 190, local agent workstations 192, and control workstations 194. The various components of the example LRM system 100 operably interconnected to collaboratively improve a process of responding to leads in a manner that optimizes contact or qualification rates.

[0029] As disclosed in FIG. 1, the remote agent stations 121 include wireless phones 122, wired phones 124, wireless computing devices 126, and workstations 128. In certain embodiments, the wireless phones 122 or the wired phones 124 may be voice over internet protocol (VOIP) phones. In some embodiments, the computing devices 126 or the workstations 128 may be equipped with a soft phone. The remote agent stations 121 enable agents to respond to lead from remote locations similar to agents stationed at the workstations 192 and directly connected to the local area network 160.

[0030] In one example embodiment, the local area network 160 resides within a call center 140 that uses VoIP and other messaging services to contact users connected to the PSTN 110 and/or the internet 130. The various servers in the call center 140 function cooperatively to acquire leads, store lead information, analyze lead information to decide how best to respond to each lead, distribute leads to agents via agent terminals such as the local agent workstations 192 and the remote agent stations 121 for example, facilitate communication between agents and leads via the PSTN 110 or the internet 130 for example, track attempted and successful agent interaction with leads, and store updated lead information.

[0031] The web server 170 may provide one or more web forms 172 to users via browser displayable web pages. The web forms may be displayed to the users via a variety of communication and/or computing devices 112 including phones, smart phones, tablet computers, laptop computers, desktop computers, media players, and the like that are equipped with a browser. The web forms 172 may prompt the user for contact data such as name, title, industry, company information, address, phone number, fax number, email address, instant messaging address, referral information, availability information, and interest information. The web server 170 may receive the lead information associated with the user in response to the user submitting the web form and provide the lead information to contact server 200 and the lead data server 190, for example.

[0032] The contact server 200 and the lead data server 190 may receive the lead information and retrieve additional data associated with the associated user such as web analytics data, reverse lookup data, credit check data, web site data, web site rank information, do-not-call registry data, data from a customer relationship management (CRM) database, and background check information. The lead data server 190 may store the collected data in a lead profile (not shown) and associate the user with an LRM plan (not shown).

[0033] The contact server 200 may contact a lead in accordance with an associated LRM plan and deliver lead information to an agent to enable the agent to respond to the lead in a manner that optimizes contact or qualification rates. The particular purpose of such contact or qualification may include, for example, establishing a relationship with the lead, thanking the lead for their interest in a product, answering questions from the lead, informing the lead of a product or service offering, selling a product or service, surveying the lead on their needs and preferences, and providing support to the lead. The contact server 200 may deliver the information to the agent using a variety of delivery services such as email services, instant messaging services, short message services, enhanced messaging services, text messaging services, telephony-based text-to-speech services, and multimedia delivery services. The agent terminals 121 or 192 may present the lead information to the agent and enable the agent to respond to the lead by communicating with the lead.

[0034] FIG. 2 is a schematic block diagram illustrating additional details of the example contact server 200 of FIG. 1. As disclosed in FIG. 2, the contact server 200 includes a contact manager 210, a dialing module 220, a messaging module 230, a PBX module 240 and termination hardware 250. In the depicted embodiment, the contact manager includes an MOR machine learning module 212, an LRM plan selection module 214, an agent selection module 216, and a lead data server access module 218. Although shown within the contact server 200, the depicted modules may reside partially or wholly on other servers such as the web server 170 and the lead data server 190 for example. The contact server 200 enables an agent to communicate with a lead in conjunction with an LRM plan.

[0035] The contact manager 210 establishes contact with users and agents and manages contact sessions where needed. The contact manager 210 may initiate contact via the dialing module 220 and/or the messaging module 230.

[0036] The MOR machine learning module 212 employs an MOR machine learning model to predict multiple interdependent output components of an MOD output decision, according to the example methods disclosed herein. In at least some example embodiments, the MOR machine learning module 212 utilizes the lead data server access module 208 to access and analyze lead information stored on the lead data server 190 of FIG. 1. Once one or more response decisions are predicted for a particular lead, the one or more response decisions may be conveyed to the LRM plan selection module 214.

[0037] The LRM plan selection module 214 presents and or selects one or more LRM plans for a particular lead and/or offering. Similarly, the agent selection module 216 selects an agent, class of agent, or agent skill set that is designated in each LRM plan.

[0038] The lead data server access module 218 enables the contact manager 210 to access lead information that is useful for contacting a lead. In one embodiment, the data storage access module 218 enables the contact manager 210 to access the lead data server 190.

[0039] The dialing module 220 establishes telephone calls including VOIP telephone calls and PSTN calls. In one embodiment, the dialing module 220 receives a unique call identifier, establishes a telephone call, and notifies the contact manager 210 that the call has been established. Various embodiments of the dialing module 220 incorporate auxiliary functions such as retrieving telephone numbers from a database, comparing telephone numbers against a restricted calling list, transferring a call, conferencing a call, monitoring a call, playing recorded messages, detecting answering machines, recording voice messages, and providing interactive voice response (IVR) capabilities. In some instances, the dialing module 220 directs the PBX module 240 to perform the auxiliary functions.

[0040] The messaging module 230 sends and receives messages to agents and leads. To send and receive messages, the messaging module 230 may leverage one or more delivery or messaging services such as email services, instant messaging services, short message services, text message services, and enhanced messaging services.

[0041] The PBX module 240 connects a private phone network to the PSTN 110. The contact manager 210 or dialing module 220 may direct the PBX module 240 to connect a line on the private phone network with a number on the PSTN 110 or internet 130. In some embodiments, the PBX module 240 provides some of the auxiliary functions invoked by the dialing module 220.

[0042] The termination hardware 250 routes calls from a local network to the PSTN 110. In one embodiment, the termination hardware 250 interfaces to conventional phone terminals. In some embodiments and instances, the termination hardware 250 provides some of the auxiliary functions invoked by the dialing module 220.

[0043] Having described a specific environment (an LRM system) and specific application (LRM) with respect to FIGS. 1 and 2, it is understood that this specific environment and application is only one of countless environments and application in which example embodiments may be employed. The scope of the example embodiments are not intended to be limited to any particular environment or application.

[0044] FIG. 3A is a schematic flow chart diagram illustrating an example MOR machine learning model 300. The model 300 is configured to be employed in sequential decision making to predict multiple interdependent output components, namely z₁, z₂, z₃, and z₄, of an MOD output decision z. Although the output decision z includes four (4) components, it is understood that an MOR machine learning model could be employed in connection with any output decision having two (2) or more interdependent components. The model 300 may be trained based on recorded historical data so that it can make optimal (or near-optimal) decisions, especially when a decision is comprised of many variables that need to be determined at the same time.

[0045] Although the model 300 may be employed in any number of applications to produce MOD output decisions, the model 300 is employed in FIG. 3A to produce an LRM MOD output decision. In particular, the model 300 is employed to decide for a given lead what response should be performed next in a sequence that will optimize the contact or qualification of the lead.

[0046] For example, the model 300 may be employed to produce an LRM MOD output decision z=(z₁, z₂, z₃, z₄), where z₁, z₂, z₃, and z₄ are four components of the output decision z, based on an input x. In this example, z₁=response agent title, z₂=response method, z₃=response message type, and z₄=response timing. The input x may be an input feature vector that includes information about a particular lead.

[0047] It is understood that the components of response agent title, response method, response message type, and response timing are only example components of an LRM MOD output decision. Other example components may include, but are not limited to, agent or lead demographic profile, agent or lead histographic profile (i.e. a profile of events in the life of the agent or the lead which could include past interactions between the agent and the lead), lead contact title (i.e. the title of a particular contact person within a lead organization), agent or lead psychographic profile (i.e. a profile of the psychological characteristics of the agent or the lead), agent or lead social network profile (i.e. the proximity of the agent to the lead in an online social network such as LinkedIn® or FaceBook® or in an offline social network such as the Entrepreneurs Organization®, civic clubs, fraternities, or religions), agent or lead geographic profile (i.e. cities, states, or other geographic designations that define current and/or past locations of the agent or the lead), response frequency (i.e. how often an agent contacts a lead), and response persistence (i.e. how long an agent persists in contacting a lead).

[0048] FIG. 3B is a text diagram illustrating an example input feature vector x. The example input feature vector x of FIG. 3B includes information about a particular lead. In particular, the example input feature vector x includes constant features about a lead, such as lead title and lead industry, and interactive features related to interactions between an agent and the lead, such as previous number of dials and previous action. The lead information provided by the example input feature vector x may be employed as input by the model 300 of FIG. 3A in order to determine what is the next sequential response that should be performed that will optimize the contact or qualification of the lead.

[0049] It is understood that the input features of lead source, lead title, lead industry, lead state, lead created date, lead company size, lead status, number of previous dials, number of previous emails, previous action, and hours since last action are only example input features to an LRM MOD output decision. Other example input features may include, but are not limited to, response agent title, response method, response message type, response timing, agent or lead demographic profile, agent or lead histographic profile, agent or lead psychographic profile, agent or lead social network profile, agent or lead geographic profile, response frequency, and response persistence. Additionally, input features could include data on current events, such as current events related to politics, economics, natural phenomena, society, and culture. It is further understood that where a particular input feature is employed as an input to a particular LRM MOD output decision, the particular input feature will not be included among the output components of the particular LRM MOD output decision.

[0050] As disclosed in FIG. 3A, there is a dependency among components z₁, z₂, z₃, and z₄. For example, a decision on the component z₂ (response method) may have an influence on the decision for the component z₄ (response timing). For example, if z₂=dial, an agent may need to consider when a lead is available to talk on a phone (e.g. usually during business hours of the time zone where the lead resides). If z₂=email, the agent may send the email at any time.

[0051] Therefore, in the example application of FIG. 3A, and as is the case with other MOD output decisions, the components of z are dependent both on an input x and on the other components of z. Thus, in this example, the probability of z₁ given x is not necessarily equal to the probability of z₁ given x and z₂, or p(z₁|x)≠p(z₁|x, z₂). In other words, it cannot be decided what value a specific component of z should take on without considering x and the values of the other components of z.

[0052] The model 300 of FIG. 3A employs a base classifier. In particular, and as disclosed in FIG. 3A, the model 300 employs multilayer perceptron (“MLP”) neural networks MLP1, MLP2, MLP3, and MLP4 as base classifiers. It is understood, however, that the model 300 could alternatively employ other types of base classifiers including, but not limited to other multilayer neural networks, decision trees, and support vector machines.

[0053] FIG. 3C is a schematic flow chart diagram illustrating the MLP neural network MLP1 that is employed to predict the first interdependent output component z₁ based on the input feature vector x of FIG. 3B and based on the predicted second interdependent output component z₂ of FIG. 3D as well as the predicted third and fourth interdependent output components z₃ and z₄. In FIG. 3C, the input feature vector x and the input components z₂, z₃, and z₄ are received by an input layer of the MLP neural network MLP1 and then processed by a hidden layer and an output layer to predict z₁ε{z₁₁, z₁₂, z₁₃}.

[0054] FIG. 3D is a schematic flow chart diagram illustrating the MLP neural network MLP2 that is employed to predict the second interdependent output component z₂ based on the input feature vector x of FIG. 3B and based on the predicted first interdependent output component z₁ of FIG. 3C as well as the predicted third and fourth interdependent output components z₃ and z₄. In FIG. 3D, the input feature vector x and the input components z₁, z₃, and z₄ are received by an input layer of the MLP neural network MLP2 and then processed by a hidden layer and an output layer to predict z₂ε{z₂₁, z₂₂, z₂₃}. As disclosed in FIG. 3A, MLP3 and MLP4 function in a similar manner to MLP1 and MLP2.

[0055] FIG. 4 is a schematic flow chart diagram of an example method 400 of employing an MOR machine learning model to predict multiple interdependent output components of an MOD output decision. The method 400 may be implemented, in at least some embodiments, by the MOR machine learning module 212 of the contact manager 210 of the contact server 210 of FIG. 1. For example, the MOR machine learning module 212 may be configured to execute computer instructions to perform operations of employing the MOR machine learning model 300 of FIG. 3A to predict multiple interdependent output components z₁, z₂, z₃, and z₄ of an LRM MOD output decision z, as represented by one or more of blocks 402, 404, 406, 408, 410, and 412 of the method 400. Although illustrated as discrete blocks, various blocks may be divided into additional blocks, combined into fewer blocks, or eliminated, depending on the desired implementation. The method 400 will now be discussed with reference to FIGS. 1-4.

[0056] The method 400 may begin at block 402, in which a classifier for each of multiple interdependent output components of an output decision is trained to predict the component based on an input and based on all of the other components. For example, the MOR machine learning module 212 may train the MLP neural networks MLP1, MLP2, MLP3, and MLP4 to predict each of the components z₁, z₂, z₃, and z₄ based on the input feature vector x of FIG. 3B and based on all of the other predicted components. In the example embodiment disclosed in FIG. 3A, z₁=response agent title, z₂=response method, z₃=response message type, and z₄=response timing. Thus, MLP1 is trained from (x, z₂, z₃, z₄; z₁) to predict response agent title z₁ using x, z₂, z₃, and z₄ as input; MLP2 is trained from (x, z₁, z₃, z₄; z₂) to predict response method z₂ using x, z₁, z₃, and z₄ as input; MLP3 is trained from (x, z₁, z₂, z₄; z₃) to predict response message type z₃ using x, z₁, z₂, and z₄ as input; and MLP4 is trained from (x, z₁, z₂, z₃; z₄) to predict response timing z₄ using x, z₁, z₂, and z₃ as input.

[0057] It is understood that since the order of components in an output decision that is produced using an MOR machine learning model can be determined simultaneously, the use herein of the subscripts 1, 2, 3, and 4 for components and the use of terms herein such as “first component” and “second component” do not denote or imply an order but rather are used herein simply for convenience in distinguishing one component from another.

[0058] At block 404, each possible value for each output component is initialized to a predetermined output value. For example, the MOR machine learning module 212 may initialize each possible value for each of the output components z₁, z₂, z₃, and z₄ to the same output value of 1/N, where N is the number of possible values for the output component, so that the sum of the initial output values of possible values for the output component is equal to 1. Alternatively, the MOR machine learning module 212 may initialize each possible value for each of the output components z₁, z₂, z₃, and z₄ to another predetermined output value including, but not limited to, an output value based on resource availability, based on a baseline, or based on Bayes priors.

[0059] In this example, assume that each of the components z₁, z₂, z₃, and z₄ has three (3) possible values as follows: z₁ε{z_1i, z₁₂, z₁₃}={sales vice president, sales manager, sales representative}; z₂ε{z₂₁, z₂₂, z₂₃}={call, email, fax}; z₃ε{z₃₁, z₃₂, z₃₃}={MT1, MT2, MT3}; and z₄ε{z₄₁, z₄₂, z₄₃}={short, medium, long}. In this example, for MLP1, MLP2, MLP3, and MLP4, there are total of twelve (12) possible input values z_ij; where iε{1, 2, 3, 4} and jε{1, 2, 3}. The inputs for the MLP neural network MLP1 are (x, z₂, z₃, z₄). There are total of nine (9) possible values for components z₂, z₃ and z₄, namely, three (3) possible values z₂₁, z₂₂, z₂₃ for z₂, three (3) possible values z₃₁, z₃₂, z₃₃ for z₃, and three (3) possible values z₄₁, z₄₂, z₄₃ for z₄.

[0060] The output value of each of the nine (9) possible values for the input components of the MLP neural network MLP1 may be initialized to ⅓ since N=3, namely, a(z₂₁), a(z₂₂), a(z₂₃), a(z₃₁), a(z₃₂), a(z₃₃), a(z₄₁), a(z₄₂), and a(z₄₃) can each be initialized to ⅓, where “a(z_ij)” is an activation that represents an output value, iε{1, 2, 3, 4}; and jε{1, 2, 3}. In a similar manner, the output value of each of the nine (9) possible values for the input components of each of the MLP neural networks MLP2, MLP3, and MLP4 may also be initialized to ⅓. For example, where z₂ε{z₂₁, z₂₂, z₂₃}={call, email, fax}, a(z₂₁) is the activation of one of the possible values, namely “call”, and may be initialized to a(z₂₁) (t)=0.33 at iteration number t=1.

[0061] Alternatively, the output value of each of the twelve (12) possible values for the input components of each of the MLP neural networks MLP1, MLP2, MLP3, and MLP4 may also be initialized to another identical output value, such as an output value less than ⅓, for example, or to non-identical output value based on resource availability, based on a baseline, or based on Bayes priors, for example.

[0062] At block 406, a relaxation iteration is run on each classifier to update the output value of each possible value for each output component. For example, the MOR machine learning module 212 may run a relaxation iteration on each of the MLP neural networks MLP1, MLP2, MLP3, and MLP4 to update the output value of each possible value for each of the output components z₁, z₂, z₃, and z₄. In this example, running a relaxation iteration on the MLP neural network MLP1 will generate three (3) output values that are retrieved directly from MLP1, namely p(z₁₁), p(z₁₂), and p(z₁₃). Similarly, running a relaxation iteration on the MLP neural networks MLP2, MLP3, and MLP4 will generate nine (9) total output values, namely p(z₂₁), p(z₂₂), p(z₂₃), p(z₃₁), p(z₃₂), p(z₃₃), p(z₄₁), p(z₄₂), and p(z₄₃). These twelve (12) output values p(z_ij) jε{1, 2, 3, 4} and jε{1, 2, 3}) may be considered as initial estimate values for the (12) output values a(z_ij), and will be used as learning targets for updating output values a(z_ij) in the next relaxation iteration, namely iteration number t+1, using the formula a(z_ij) (t+1)=a(z_ij) (t)+R·(p(z_ij) (t)−a(z_ij) (t)), which is discussed in greater detail below. For example, where iteration number t=1, a(z₂₁) (t) is updated by running a relaxation iteration on MLP2 to produce the output value p(z₂₁) (t)=0.47 of z₂₁. In this example, a(z₂₁) (t) is the output value of MLP2 at iteration number t and it is used as an input for MLP1, MLP3, and MLP4 in the next iteration, namely iteration number t+1. In this example, p(z₂₁) (t) is the output value retrieved directly from MLP2 at iteration number t and it is used as a target for updating a(z₂₁) (t+1).

[0063] At decision block 408, it is determined whether a relaxation state has reached an equilibrium. If the relaxation state has reached an equilibrium, (“Yes” at decision block 408), then the method 400 proceeds to block 412. If the relaxation state has not reached an equilibrium (“No” at decision block 408), then the method 400 proceeds to decision block 410.

[0064] For example, the MOR machine learning module 212 may determine whether a relaxation state has reached an equilibrium. Whether an equilibrium has been reached may be determined according to the following two formulas. First, a relaxation rate is applied to update the output value of each possible value for each of the output components z₁, z₂, z₃, and z₄ as follows: a(z_ij) (t+1)=a(z_ij) (t)+R·(p(z_ij) (t)−a(z_ij) (t)); where R is a relaxation rate; t is the number of iterations; iε{1, 2, 3, 4}; and jε{1, 2, 3}. Second, once the following equation is true, the relaxation may be considered to have reached an equilibrium: |a(z_ij) (t+1)−a(z_ij) (t)|≦T; where T is a threshold.

[0065] For example, where iteration number t=1, and where R=0.1, and using the example values of a(z₂₁) (t) and p(z₂₁) (t) of z₂₁ from above, namely a(z₂₁) (1)=0.33 and p(z₂₁) (1)=0.47, the formula a(z_ij) (t+1)=a(z_ij) (t)+R·(p(z_ij) (t)−a(z_ij) (t)) is first processed as follows: a(z₂₁) (2)=0.33+0.1·(0.47−0.33)=0.344. Second, where T=0.01, the formula a(z_ij) (t+1)−a(z_ij) (t)| is processed as follows: |0.344−0.33|=0.014. Since 0.014 is greater than 0.01, then the statement |a(z_ij) (t+1)−a(z_ij) (t)|≦T is false, and the relaxation is not considered to have reached an equilibrium.

[0066] At decision block 410, it is determined whether a maximum number of relaxation iterations has been reached. If the maximum number of relaxation iterations has been reached, (“Yes” at decision block 420), then the method 400 proceeds to block 412. If the maximum number of relaxation iterations has not been reached (“No” at decision block 410), then the method 400 returns to block 406 for another relaxation iteration.

[0067] For example, the MOR machine learning module 212 may determine whether a maximum number of relaxation iterations has been reached. In this example, once the following equation is false, the maximum number of relaxation iterations may be considered to have been reached: t≦M; where t is the iteration number, and M is the maximum number of relaxation iterations.

[0068] For example, where iteration number t=1, and M=100, since 1 is less than 100, then the statement t≦M is true, and the maximum number of relaxation iterations is not considered to have been reached. Therefore, the method 400 may return to block 406 for another relaxation iteration where iteration number t=2, a(z₂₁) (2)=0.344 is used as an input to MLP1, MLP3, and MLP4. Similarly, MLP2 receives input from corresponding a(z_ij) (2) values retrieved from MLP1, MLP3, and MLP4. From these inputs, MLP2 generates an output value p(z₂₁) (2). The value of a(z₂₁) (3) can then be calculated from (z₂₁) (2) and an output value p(z₂₁) (2) using the formula a(z₂₁) (3)=a(z₂₁) (2)+R·(p(z₂₁) (2)−a(z₂₁) (2)). In this example, at iteration number t=2, the other eleven (11) a(z_ij) (3) will also be updated using outputs p(z_ij) (2) of the appropriate classifiers and a(z_ij) (2) at iteration number t=2 as inputs to the formula a(z_ij) (3)=a(z_ij) (2)+R·(p(z_ij) (2)−a(z_ij) (2)).

[0069] At block 412, an optimal output component is retrieved from each classifier. For example, the MOR machine learning module 212 may retrieve an optical output component for each of the components z₁, z₂, z₃, and z₄ from the MLP neural networks MLP1, MLP2, MLP3, and MLP4, respectively.

[0070] It is understood that the above-illustrated example is but one example of employing an MOR machine learning model to predict multiple interdependent output components of an MOD output decision, and the method 400 is not limited to the particular application of this example or to the LRM MOD problem solved in this example.

[0071] FIG. 5 is a schematic flow chart diagram 500 of multiple correct MOD output decisions. As disclosed in the diagram 500, the MOR machine learning model 300 may generate multiple correct output decisions 502 and 504 for a given input feature vector x. Although in a typical decision making process it is usually assumed that there is a unique correct decision given a fixed input, for LRM MOD decisions there may be multiple correct decisions which may all produce similar favorable results. A decision may be chosen among multiple correct decisions based on available resources. For example, if a particular response agent with response agent title z₁=“sales manager” is not available at a particular time, then another correct decision with response agent title z₁=“sales representative” may be made. Where multiple output decisions are simultaneously considered to be correct, the term “correct” may refer to multiple output decisions each having a substantially similar output value. For example, each of the output decisions 502 and 504 of FIG. 5 may have an identical or substantially similar output value, which indicates that performing either output decision would produce similar favorable results. Additionally or alternatively, the term “correct” may refer to multiple output decisions each having an output value above a predetermined threshold. The threshold may be predetermined to be relatively high or relatively low, depending on the application. Although only two correct output decisions are disclosed in FIG. 5, it is understood that the MOR machine learning model 300 may generate more than two correct output decisions.

[0072] Having described example methods of employing an MOR machine learning model to predict multiple interdependent output components of an MOD output decision with respect to FIGS. 3A-5, example systems and user interfaces that enable agents to access and implement the resulting output decisions will be described with respect to FIGS. 6-8B. It is understood that these specific systems and user interfaces are only some of countless systems and user interfaces in which example embodiments may be employed. The scope of the example embodiments is not intended to be limited to any particular system or user interface.

[0073] FIG. 6 illustrates an example computer screen image of a user interface 600 of an example customer relationship management (CRM) system. The user interface 600 includes various controls that allow an agent to manage customer relationships and, in particular, manage leads that are provided by the CRM system. The user interface 600 may be presented to an agent by the web server 170 on the workstations 128 or on the local agent workstations 192 of FIG. 1, for example. The agent may use the user interface 600 to respond to leads that have been previously stored on the lead data server 190 of FIG. 1. In particular, the lead advisor display 800 may allow the agent to respond to leads in a manner that optimizes contact or qualification rates, as discussed below in connection with FIGS. 8A and 8B.

[0074] FIG. 7 illustrates an example computer screen image of a user interface 700 of an example LRM system, such as the LRM system of FIG. 1. Like the user interface 600 of FIG. 6, the user interface 700 includes various controls that allow an agent to respond to lead. The user interface 700 may be presented to an agent in a similar manner as the user interface 600. The user interface also includes a lead advisor display 800.

[0075] FIG. 8A illustrates an example computer screen image of the example lead advisor display 800 before a lead has been selected by an agent and FIG. 8B illustrates an example computer screen image of the example lead advisor display 800 after a lead has been selected by an agent. As disclosed in FIG. 8A, the lead advisor display 800 lists five leads. Each lead includes a name 802, a likelihood of success meter 804, and a likelihood of success category indicator 806. As disclosed in FIG. 8A, the leads are listed by highest likelihood of success to lowest likelihood of success. Upon inquiry by the agent, by mousing-over a lead with a mouse pointer for example, the lead may expand as shown in FIG. 8A for lead “Mark Littlefield.” Upon expansion, the lead may present the agent with additional options, such as a confirm button 808, a delete button 810, and a “more info” link 812.

[0076] Upon selection of the “more info” link 812 by the agent, by clicking on the more info link 812 with a mouse pointer for example, the agent may be presented with a pop-out display 814 as disclosed in FIG. 8B. The pop-out display 814 may present the agent with an LRM plan associated with the lead. This LRM plan may have been generated by the example methods disclosed herein and may reflect the output decision with the highest, or among the highest, output value for the lead. As disclosed in FIG. 8B, the LRM plan for the lead named “Mark Littlefield” may include employing a sales manager to send an email with message type MT1 in a short timeframe, which corresponds to the output decision 502 of FIG. 5. The agent may then simply click on the pop-out display 814 to have the lead advisor display 800 automatically generate an email to the lead with message type MT1 that will be sent by a sales manager immediately. Alternatively, the agent may manually override the response plan and manually perform a different response.

[0077] Therefore, the embodiments disclosed herein include methods of employing an MOR machine learning model to predict multiple interdependent output components of an MOD output decision. The example methods disclosed herein enable the prediction of each output component based on an input and based on all of the other output components. Therefore, the example methods disclosed herein may be employed to solve MOD problems such as LRM problems.

[0078] The embodiments described herein may include the use of a special purpose or general-purpose computer including various computer hardware or software modules, as discussed in greater detail below.

[0079] Embodiments described herein may be implemented using computer-readable media for carrying or having computer-executable instructions or data structures stored thereon. Such computer-readable media may be any available media that may be accessed by a general purpose or special purpose computer. By way of example, and not limitation, such computer-readable media may include non-transitory computer-readable storage media including RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other storage medium which may be used to carry or store desired program code in the form of computer-executable instructions or data structures and which may be accessed by a general purpose or special purpose computer. Combinations of the above may also be included within the scope of computer-readable media.

[0080] Computer-executable instructions comprise, for example, instructions and data which cause a general purpose computer, special purpose computer, or special purpose processing device to perform a certain function or group of functions. Although the subject matter has been described in language specific to structural features and/or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing the claims.

[0081] As used herein, the term “module” may refer to software objects or routines that execute on the computing system. The different modules described herein may be implemented as objects or processes that execute on the computing system (e.g., as separate threads). While the system and methods described herein are preferably implemented in software, implementations in hardware or a combination of software and hardware are also possible and contemplated.

[0082] All examples and conditional language recited herein are intended for pedagogical objects to aid the reader in understanding the example embodiments and the concepts contributed by the inventor to furthering the art, and are to be construed as being without limitation to such specifically recited examples and conditions.