Patent Application
20080046386
The invention described here can be realized by an analytical function:
d=D(P,F)
where d=(1, 2, . . . , M) is the index or scale of values, of one of a finite set of M actions (a1, a2, . . . , aM) to be taken at a given point during the interaction. F is a vector of features that are estimated and given an actual value during the interaction. These features may include but are not limited to, for example, the time elapsed from the beginning of the call t, the ordinal number of the current caller turn (the number of separate questions and answers solicited from the caller) NT, the number of speech recognition no-matches experienced by the caller since the beginning of the interaction NNM, the number of time-outs experienced by the system since the beginning of the interaction NTO, the time of day TD, the day of the week dw, and other parameters that are believed to influence the optimal decision. Thus, the vector F, without any loss in generality, may assume the following general form:
F=(F1,F2,F3, . . . ,FN)
and, in particular, the vector F may assume the form:
F=(t,NT,NNM,NTO,TD,dw . . . )
Without any loss in generality, the function D is expressed in a parametric form, with a number Q of free parameters described by the vector P:
P=(p1,p2,p3 . . . ,pQ)
One of many machine learning algorithms available as prior art and described in the machine learning literature can be used in conjunction with a set of training examples to estimate the parameters P of a generic parametric function in order to optimize a defined measurable criterion. For that purpose, a set T of L training examples
T=(T1,T2, . . . TL)
can be constructed from historic interaction data in the following way: for each caller interaction turn j of each call k in the set of calls that belong to the historic data, the actual feature vector Fk,j can be recorded. So, a training example Tk will be recorded as the following sequence of feature vectors:
Tk=(Fk,1Fk,2, . . . Fk,Rk,Ok)
where Rk is the actual number of turns for that call and Ok is the actual outcome of the call that is directly related to the overall cost function. For instance, in the example described above the outcome of the call can be either successful or unsuccessful.
Using an available machine learning algorithm the parameters P of the function D can be estimated in order to optimize the capacity of the technique according to the present invention to choose the best action (e.g., continue or transfer in the previous example) with respect to the chosen optimization criterion. For example, the chosen optimization criterion may include minimize the cost of all caller interactions. For that chosen optimization criterion, the automated IVR would continue without transferring the call to a live operator or transfer the call, depending on the actual values of the feature vector F obtained during the actual interaction. Once the function D is estimated, the function can be used at each turn or step of each interaction for selecting the action that would optimize the defined global criterion.
The operation of the present invention decision function, as set forth hereinbefore, as; d=D(P, F) wherein F is a vector of features and P is a vector of parameters related to each feature or combination of features may be set forth in the following exemplary descriptions. In this example the features as previously set forth may include, for example, the time elapsed from the beginning of the call t, the ordinal number of the current caller turn NT, the number of speech recognition no-matches experienced by the caller since the beginning of the interaction NNM, the number of time-outs experienced by the system since the beginning of the interaction NTO, the time of day TD, the day of the week dw, and other attributes that are believed to influence the optimal decision. The parameters P related to such features that optimize the decision according to the established optimality criterion are identified through a machine learning algorithm based on historical interactions between callers and call centers, in other words a record of prior calls between callers and an automated interactive voice response system. In the machine learning algorithm, software, which embodies the algorithm may be “trained” on test cases, in this case the historical interactions, and scored so the software knows what outcomes correspond to the chosen optimization criterion. Once trained on the historical examples, the features and parameters identified may then be used to solve real-world cases. The machine learning algorithm can identify and record distinct features for an interaction between a caller and call center as well as the outcome of the call, for example either successful or unsuccessful. The historical interactions assessed by the machine learning algorithm may be referred to as training examples. The set of features compiled during the training examples may then used to construct any number of parameters, or rules, which may be implemented to optimize the overall cost of customer care. The parameters may be logical statements relative to each feature or combination of features, and may take the form, for example, of an “if-then” or Boolean logical operation. However, the present invention is not so limited to include only “if-then” or Boolean logical operations. Other more complex relationships between the features may be developed to refine the process of optimizing the cost of customer care.
For example, a feature identified in a training example may include the time of day TD. For the time of day feature, a set of parameters or rules may then be developed based on recorded outcomes for any particular time of day. Therefore for each time of day, a parameter may be established. An example of such a parameter may be stated as; if call is initiated after 11:00 pm, then immediately direct the call to a live operator. The rule would be based on the training example interactions which may have shown that when a call is initiated after 11:00 pm, the automated call outcome is unsuccessful for a large percentage of calls, therefore, the overall cost of customer care is optimized by immediately directing the call to a live operator rather than utilizing an automated system for even a portion of the call interaction.
The machine learning algorithm may establish such a parameter by determining the optimal percentage of success versus unsuccessful outcomes for each time of day in order to optimize the overall cost of customer care. In other words as the ratio of unsuccessful to successful calls increases for each time of day measured in a training example, the decision function will identify the optimal ratio or percentage of unsuccessful outcomes where an automated call should be switched to a live operator.
Additionally, a parameter may be based on one or more than one feature, for example time of day TD. the time elapsed from the beginning of the call t, and the number of speech recognition no-matches experienced by the caller since the beginning of the interaction NNM. In this example a parameter may be established through the training examples such that calls initiated between 2:00 pm and 3:00 pm with an elapsed time greater than 5 minutes, having more than 3 speech recognition no matches, are transferred to a live operator.
Turning now to
Turning now to
As previously stated above, the machine learning algorithm, software, which embodies the algorithm may be “trained” on test cases, in this case the historical interactions, and scored so the software knows what outcomes correspond to the chosen optimization criterion. Step three (3), 204 comprises constructing with a machine learning algorithm any number of parameters, or rules, which may be implemented to optimize the overall cost of customer care from the set of features compiled during the training examples. The parameters may be logical statements relative to features or combination of features, and may take the form, for example, of an “if-then” or Boolean logical operation. However, the present invention is not so limited to include only “if-then” or Boolean logical operations. Other more complex relationships between the features may be developed to refine the process of optimizing the cost of customer care. A specific example of such a parameter established in step (3) may include: If the time elapsed from the beginning of the call is greater than 10 minutes, then transfer the call to a live operator. As previously explained any number of other parameters may be established from the features recorded from the training examples.
In step four (4) 206 a decision function in accordance with the technique of the present invention may be implemented within an automated IVR call center. The decision function d=D(P,F) utilizing the technique of the present invention may be used at each turn or step of each caller interaction for selecting the action that would optimize the defined global criterion. Step five (5) 208, includes receiving a call from a caller to an automated IVR in accordance with the present invention. Each turn or step of the interaction with the caller based on the decision function is analyzed in accordance with the decision function in step six (6) to determine if any parameter established in step (3) 204 of the decision function is invoked. Step seven (7) comprises either continuing the IVR interaction 212 or transferring 214 the call to a live operator for completion to a successful outcome 216. As is depicted in the flow chart, step (7) is an iterative process that may be repeated for each turn or step of each call. In the case where the IVR is continued, each turn is analyzed successively throughout the call in accordance with the decision function to determine if the automated IVR should be continued to a successful outcome 216, or transferred to a live operator 214.
One skilled in the art will further recognize that features other than those listed above are possible and may be obtained by common sense and by using knowledge of the application and the behavior of callers. Other exemplary features may include the originating area code of the call, the day of the week or calendar date. Likewise in addition to the exemplary parameters stated above, other logical statements relative to other features and combinations of features, defined through a machine learning algorithm may be developed. The present invention is not limited by the features and parameters disclosed herein. Any number of parameters may be developed and combined for optimizing the decision function.
While preferred embodiments of the present invention have been described using specific terms, such description is for illustrative purpose only. It is obvious that changes and variations may be made by those skilled in the art without departing from the scope of the claims. Therefore, the changes and variations are understood to be contained in the spirit or scope of the claims that follow.
The present invention claims priority to U.S. Provisional Application 60/806,483, filed Jul. 3, 2006, the entire contents of which being incorporated herein by reference.
| Number | Date | Country | |
|---|---|---|---|
| 60806483 | Jul 2006 | US |
