یادگیری مبتنی بر مفهوم رفتار انسانی برای مدیریت ارتباط با مشتری

In this paper, we apply concept learning techniques to solve a number of problems in the customer relationship management (CRM) domain. We present a concept learning technique to tackle common scenarios of interaction between conflicting human agents (such as customers and customer support representatives). Scenarios are represented by directed graphs with labeled vertices (for communicative actions) and arcs (for temporal and causal relationships between these actions and their parameters). The classification of a scenario is performed by comparing a partial matching of its graph with graphs of positive and negative examples. We illustrate machine learning of graph structures using the Nearest Neighbor approach and then proceed to JSM-based concept learning, which minimizes the number of false negatives and takes advantage of a more accurate way of matching sequences of communicative actions. Scenario representation and comparative analysis techniques developed herein are applied to the classification of textual customer complaints as a CRM component. In order to estimate complaint validity, we take advantage of the observation [19] that analyzing the structure of communicative actions without context information is frequently sufficient to judge how humans explain their behavior, in a plausible way or not. This paper demonstrates the superiority of concept learning in tackling human attitudes. Therefore, because human attitudes are domain-independent, the proposed concept learning approach is a good compliment to a wide range of CRM technologies where a formal treatment of inter-human interactions is required.

In recent years, it has become clear that it is hard to overestimate the importance of customer support and customer retention in industry. Customer relationship management (CRM) has grown into a significant industrial sector with its own series of technological advancements. A number of computer science algorithms, including optimization and scheduling, have been developed specifically targeting CRM [43], [15], [60] and [64]. More areas of Artificial Intelligence are still finding applications in CRM; the current paper addresses the simulation of human reasoning and behavior, the proper and efficient implementation of which can be vital to a series of CRM applications. A state-of-art CRM system must be capable of simulating human behavior to properly address customer needs, facilitate communication, perform customer retention and resolve conflicts should they arise. To solve these problems, a CRM application needs the capability to operate in the realm of the human thoughts, by simulating human reasoning and by learning human behavior. In this paper we propose a concept-based representation technique and an infrastructure to learn customers’ behavior. One of the main problems to be solved in facilitating customer retention and assisting inter-human conflict resolution is how to reuse previous experience in later situations with similar agents. A business rule system-based architecture is typical for CRM [27]. However, machine learning is required for handling a poorly formalized domain like human behavior [59] and [42]. Using information about customers’ prior behavior and historical patterns to understand buying activities, behaviors, and ticketing characteristics is important. Most companies are new to using such structured information about customer behavior to manage and measure relationships. Such efforts go beyond having a call center for customers to raise complaints; it requires having a modern behavior-simulation based management system that listens to the customers, documents the problem and solution, and changes the behavior of employees and call center interactions to build proper relationships with customers [53]. In a series of previous studies, we focused on various issues surrounding the practical implementation of reasoning in such domains as understanding multiagent scenarios [16] and [22], determining possible criminal behavior of mobile phone users by means of analyzing the location tracking data [16], and emotional profiling [18]. We have addressed a number of issues with graph learning, simulating reasoning about mental states and communicative actions; and introduced complaint scenarios as graphs, using argumentation-based learning [19]. We explored the contribution of specific sources of information about scenarios as communicative actions, argumentation and meta-argumentation patterns [20], and causal links [17] and [18]. In this study, we focus on scenario structures as a whole to build a concept learning framework for CRM. Referring to concept learning and concept graphs, we follow Mitchell [36] and [51]. We will observe how concept learning helps to deal with customer complaints, as well as how it assists in the interactive exploration of product features extracted from customer reviews. We select lattices and formal concept analysis (FCA, [24]) as tools for learning human behavior because they have the following properties: • flexibility, • appropriateness for poorly formalized domains like human behavior, • deterministic structures capable of explicit explanation of decisions proposed by the system, and • sensitive measure of concept similarity [14]. In the last decade, machine learning features of FCA have been leveraged by a number of industrial applications, and we believe CRM will further demonstrate its capability to handle domains with extremely complex structures. Hence, this paper contributes to the state-of-art by building a concept learning framework to operate on human attitudes for decision support and decision making and thoroughly evaluates this framework. We will demonstrate that concept-based learning is better suited for representing complex patterns of human behavior, including communication, than conventional machine learning mechanisms, such as classification of groups of words extracted from textual descriptions of a conflict or dialogue. To properly position our work in a family of CRM technologies, we mention the following classes of CRM services, following [11] and [9]: (1) aggregation of data to create a single, accessible source (whether physical or virtual), (2) analysis and presentation of that data as usable information by individuals doing strategic planning or executing strategic sales/marketing initiatives, and (3) tools and information to provide front-line personnel or systems that are interacting with customers or prospects the ability to make timely, educated decisions that benefit both the customer and company. In this paper, we focus on the tools mentioned in the third class of CRM services, specifically focusing on facilitating customer interaction through concept learning technologies. The following sequence of problems needs to be solved for predicting and classifying human behavior using a CRM system: (1) Discover how to reconstruct behavior patterns from text. It turns out that communicative actions and their subjects are essential elements of behavior discourse. (2) Construct a formal language to represent communicative actions. Find attributes of communicative actions so that similarity between them can be defined. Analyze how the mental space is ‘covered’ by communicative actions, and form a substitution matrix for them to measure similarity. (3) Build a way to extract information from natural language for communicative actions (which is relatively easy) and their subjects and parameters (which is significantly harder due to implicit references to these subjects in natural language). In the expression ‘He denied that he made an early withdrawal from his account’ communicative action = ‘deny’ and its subject = ‘he made an early withdrawal’. (4) Observe that the sequence of behavior patterns can be packaged as a scenario. Define a scenario as a graph including communicative actions and interaction between their subjects, based on causal links and relations for argumentation. (5) Define relationships between scenarios via subgraphs, with respective operations on vertices and arcs. Define similarity between scenarios based on graphs and similarities between individual communicative actions. (6) Build a machine learning framework and select a particular learning approach well suited to operate with scenario graphs. Evaluate whether concept learning is an adequate approach. One of the most important tasks in assisting negotiations and resolving inter-human conflicts in a CRM framework is the validity assessment. A scenario (in particular, a complaint) is valid if it is plausible, internally consistent, and also consistent with available domain-specific knowledge. On the contrary, a complaint scenario is invalid if there are inconsistencies in the communication discourse, so that there is doubt as to whether a problem with a product (mentioned in this complaint scenario) has actually occurred. Using concept learning, we try to discover such inconsistencies without involving domain knowledge. In the case of inter-human conflicts or negotiations, such domain-specific knowledge is frequently unavailable. In this paper, we build a CRM framework to assist companies with complaint management, assigning complaints to a class of valid or invalid scenarios. 1.1. Logical simulation of behavior An extensive body of literature addresses the problem of logical simulation of behavior of autonomous agents, taking into account their beliefs, desires and intentions [5]. A substantial advancement has been achieved in building the scenarios of multiagent interaction, given properties of agents, including their attitudes. However, the means of automated comparative analysis for interaction scenarios for human agents are still lacking. In our previous study [21], we analyzed the roles of deduction, simulation and inductive learning in application to human agents. In the current paper, we build the representation machinery and develop a concept learning technique for operating with scenarios that include a sequence of communicative actions. We propose a framework for classifying scenarios of inter-human conflicts. This framework will be implemented in a stand-alone mode and used in combination with deductive reasoning to be a part of a decision-support system. In spite of the advances in modeling conflicts and negotiations between autonomous agents and its deployment in a number of domains, a general framework to reuse the experience of conflict resolutions from earlier cases has not been developed yet. To effectively build such a framework and predict the interaction between autonomous agents, it is helpful to augment reasoning and/or simulation with machine learning [57], [44] and [52]. In the case of human agents, an adequate behavioral model that gives a plausible data structure for machine learning is essential as well. It would reduce the number of possible actions for the agents at each step, taking into account how these agents acted in previous cases. Obviously, formalizing human behavior is a much more complex task than that of automated agents. To simplify the representation, we restrict ourselves to communicative actions (plus the causal and argumentative links between them) of human agents in the course of an interaction (conflict) as a way to describe their behavior. Recently, the issue of providing BDI (Belief–Desire–Intention) agents [5] with machine learning capabilities attracted interest; an application domain of agents for intelligent information access was considered in [52]. Nevertheless, a BDI-based machine learning framework for scenarios of inter-human interactions has not yet been developed. A number of case-based reasoning approaches have been suggested to treat interaction scenarios involving BDI agents [33] and [44]; however, the description of agents’ attitudes and behaviors is reduced to their beliefs, desires and intentions in these studies. Indeed, the behavior of real-world agents in conflict is described in a richer language using a wide number of mental entities including pretending, deceiving, offending, forgiving, trusting, and others. Formalized inter-human conflict is a special case of a formal scenario where the agents have inconsistent and dynamic goals; a negotiation procedure is required to achieve a compromise [41]. In this paper, we employ the hypothesis that by following the logical structure of how a negotiation is represented in a scenario (available as a text or presented in some structured way), it is possible to judge the consistency of the scenario. We take advantage of this assumption and propose an interactive form, where the required parameters of communicative actions are specified from the viewpoint of a given agent. We believe that a useful machine learning framework for operating with scenarios of inter-human interactions should exhibit the following characteristics: • It should be capable of relating a scenario to a class of scenarios, given a number of classes specified for a given domain by experts (two classes in our case); • It should be based on a concise and effective model that represents inter-human interactions, operating with a rich set of communicative actions. • It should be domain-independent and therefore equally applicable to any domain; it should also allow the avoidance of domain-specific ontologies; • It should provide motivations for the classification decisions, because it is a component of a decision-support system for an industry sector of customer relation management. A learning model needs to be focused on a specific graph representation for these conflicts. The learning strategies used here are based on ideas similar to that of the Nearest Neighbors (see, e.g., [36]), case-based learning [30], concept-based learning [31] and [23] or the JSM-method [12] and [13]. Having defined scenarios and the operation of finding common subscenarios, we use the Nearest Neighbors algorithm as a simple illustration of our approach to relate a scenario to either the class of valid or invalid scenarios. We then proceed to JSM-based learning to avoid false positives as much as possible. JSM-based learning delivers the most cautious approach to classifying human behavior and attitudes in order to comply with the ethical and legal norms of CRM. In the current paper, we use deterministic machine learning because the explicit motivations for the decisions might be more important than the decision class itself, whenever decision support is provided. We believe concept learning is more appropriate for CRM settings where decisions have to be clearly communicated and solidly backed up, than statistical learning, even if the latter might be more accurate [18]. 1.2. Complaint validity, complaint management and CRM Complaint processing [9] has become an important issue for CRM in large companies and organizations. Complaint management is a formal process of recording and resolving a customer complaint. Even though CRM systems in general and complaint processing systems in particular are expensive, companies can extract priceless knowledge from an appropriate handling of a complaint, with significant effects on customer retention rates and word-of-mouth recommendations [60]. If complaints are transformed into knowledge about customers, they can provide valuable business intelligence for enterprises. To exploit this intelligence, companies must design, build, operate and continuously upgrade systems for managing complaints. In the last few years, several approaches have emerged to automate complaint management such as [59] and [15], among others. Retailers and service providers may profit from such software services because they allow complaints to be handled faster, providing the possibility of feedback analysis and data mining capabilities on the basis of a complaint database. A typical complaint is a report of a failure of a product or service, followed by a narrative on the customer’s attempts to resolve the issue. Complaints include both a description of the product or service failure as well as a description of the resulting interaction process (negotiation, conflict, etc.) between the customer and the company representatives. Because it is almost impossible for CRM personnel to verify the actual occurrence of such failures, company representatives must judge the adequacy of a complaint on the basis of the communicative actions provided by the customers in their narratives. Currently, most customer complaint management solutions are limited to the use of keyword processing to relate a complaint to a certain domain-specific class (e.g., banking and travel complaints, as reported in this paper), or to the application of knowledge management techniques in software platforms for workflow processing (e.g., [64] and [38]). To the best of our knowledge, existing industrial complaint management platforms do not make use of natural language processing nor machine learning techniques for quicker performance, quality assurance and lower sustainability costs; most complaint handling functionalities remain manual. Thus, for example, even advanced tools such as Oracle PeopleSoft Enterprise Customer Relationship Management (CRM) do not exploit the possible benefits of learning from available complaint data. In particular, no automated solutions have been developed to assess the validity of a customer complaint on the basis of the emerging dialogue between a customer and company representatives, with the goal of better supporting the procedure of complaint handling as a part of CRM. 1.3. Conflict scenario and communicative actions We proceed to the main definition of this study of how a behavioral scenario consists of communicative actions. A communicative action is a functor of the form verb (agent, subject, cause), where verb characterizes some kind of interaction between customer and company in a conflict scenario (e.g., explain, confirm, remind, disagree, deny), agent identifies either the customer or the company, subject refers to the information transmitted or object described, and cause refers to the motivation or explanation for the subject. Thus, for example, a communicative action associated with some customer claim such as, “I disagreed with the overdraft fee you charged me because I made a bank deposit well in advance” would be represented as disagree (customer, “overdraft fee,” “I made a bank deposit well in advance”). Scenarios are intentionally simplified as labeled directed graphs to allow for effective similarity matching among them. Each vertex in the graph will correspond to a communicative action. An arc (oriented edge) may denote either temporal precedence or an attack relationship between two actions ai and ai. In the first case, we will distinguish between consecutive actions that refer to the same subject from those that refer to different subjects. Graphically, we will distinguish these situations by means of thick arcs and thin arcs, respectively. A complaint scenario is a labeled directed graph G = (V, A), where V = {action1, action2, … , actionk} is a finite set of vertices corresponding to communicative actions, and A = Athick ∪ Athin ∪ Acausal is a finite set of labeled arcs (ordered pairs of vertices), classified as follows: • Each arc (actioni; actionj] ∈ Athick corresponds to the temporal precedence of two references to the same subject. • Each arc (actioni; actionj] ∈ Athin corresponds to the temporal precedence of two actions referring to different subjects. • Each arc (actioni; actionj] ∈ Acausal corresponds to a causal link or an attack relationship between actioni and actionj, indicating that the cause of actioni is in conflict with the subject or cause of actionj. The rest of the paper is organized as follows. We first introduce the domain of conflict scenarios and then present both a formal treatment of communicative actions and a detailed definition of conflict scenarios as graphs encoding communicative actions. Second, having defined the similarity operation on graphs as finding maximal common subgraphs, we move onto relating a scenario to a class of scenarios using the Nearest Neighbor approach, following Galitsky and Kuznetsov [19]. To improve the accuracy of the classification and to adjust the machine learning technique to real-world requirements, we use the logic programming system Jasmine, which is based on JSM-method learning [13]. The procedure of finding similarities between scenarios is then described, taking into account the aggregation of communicative actions with the same subject and causal links. We then evaluate the proposed technique in the domains of banking and travel, and also compare the technique with state-of the-art techniques in opinion mining. Towards the end of the paper, we address concept-based exploration of product features and local logic-based frameworks to deductively describe scenario discourse using non-monotonic reasoning.

In this paper, we proposed a concept learning approach to relate a human behavior pattern to classes. We used a representation language of labeled directed acyclic graph labels with vertices for communicative actions and arcs for temporal relations, causal links and attack relations on them. For the purpose of machine learning, the scenarios are represented as a sequence of communicative actions attached to agents; the communicative actions are grouped by subjects, and the order of communicative actions is retained using binary predicates after. We considered the concept lattice of communicative actions and showed how the procedure of relating a complaint to a class can be implemented by Nearest Neighbor and JSM [13] learning machinery. This approach is believed to be an innovative way to learn scenarios of inter-human interactions that are encoded as sequences of communicative actions. 8.1. The value of CRM technologies CRM is the notion that businesses should focus on the customer and reinvent themselves to deliver personalized, service-driven sales and support. We do not agree with a purely “financial” approach that the overall objectives of CRM are to drive productivity and provide measurable return on investment, while improving profitability and expanding market share. While these objectives are sometimes the result of a well-executed CRM initiative, many CRM projects help organizations achieve a specific objective such as improving customer satisfaction or retention. The importance of using predictive technologies by front-line personnel to proactively address issues has been demonstrated by Dver [11]. Using the technique presented in this paper, the front-line personnel can engage in more proactive customer contact rather than merely reacting to customer requests, as is often the case today. As has been observed with automated phone answering systems, customers with issues and problems are driven to other channels such as Web sites for self-help. The goal of CRM will be to save the time and talents of educated workers so that they may focus on gathering non-system data, such as a customer’s demeanor and future plans. Front-line workers will be transformed, becoming salespeople and/or consultants utilizing the CRM system to research historical information and add value to their personal interactions with customers, whether those interactions are on the phone, via the Internet or in person. To make this cultural change, front-line workers will need to adopt new habits (perhaps encouraged by changes in their compensation) so that they use the CRM systems and understand the value of both the system and their front-line interactions. Natural language interfaces will be used by both front-line workers and customers to facilitate easy access to information, regardless of location, channel or experience. Many analytical tasks in CRM, such as churn prognosis, risk management or targeted marketing, involve classification of customers and their behavior [6] and [42]. For example, CRM analysis of a telecommunication provider might build classification models trying to predict whether a customer presents a high, medium or low risk of switching providers, a.k.a. churn prediction. Given training and test data sets, analysts can compare different algorithms based on overall accuracy. Traditionally, classification has focused on attribute-value learning, where each example or instance can be characterized by a fixed set of attributes. The hypothesis language is propositional logic and these types of algorithms are referred to as propositional learners. Dierkes et al. [10] improve churn prediction models by leveraging network effects; in contrast to traditional classification algorithms, the information about a customer’s neighbors in the communication graph is taken into account. A Markov logic network field [49] is used, combining Markov Random Fields and Inductive Logic Programming (ILP) to define a distribution over objects’ properties and relations among them by attaching a weight (capturing the importance of the formula) to each formula in a first-order theory. Su [63] proposed a model which uses “concepts” and “associations” of concepts to model the real-world information of a database management environment. To position our proposal among the learning techniques, we state that it is related to ILP [40], which has been applied to CRM, and Explanation Based Learning, which is intended to derive as general expressions as possible from available data [47] and [37]. Our approach is distant from statistical, hybrid [49], or neurocomputing [28] approaches. In performing an automated assessment of the trustworthiness of a complaint, we follow work of Montaner et al. [39], which takes into account opinion reliability. It is worth mentioning some NLP approaches targeting extraction of implicit relationship and features from text such as complaint validity. Latent Semantic Analysis (LSA) has been applied as a model of cognitive processing and word-meaning acquisition, relying on its capacity to modulate the meaning of words by contexts, dealing successfully with polysemy and synonymy. According to Valle-Lisboa and Mizrajia [62] the underlying reason that makes LSA work well is detection of an underlying block structure (the blocks corresponding to topics and actions of involved agents) in the term-by-document matrix; in real cases this block structure is hidden because of perturbations. It seems promising to apply high scalability Latent Semantic Indexing (LSI++, [34]) which supports both serial and distributed searching of large data sets, providing the same programming interface regardless of the implementation actually executing. Zheng et al. [68] combine detection of noun phrases with the use of WordNet as background knowledge to explore better ways of representing documents semantically for clustering. Based on noun phrases as well as single-term analysis, the authors exploit different document representation methods to analyze the effectiveness of hypernymy, hyponymy, holonymy, and meronymy. Kang [29] proposed an indexing formalism that considers not only the terms in a document, but also the concepts, which are extracted by exploiting clusters of terms that are semantically related, referred to as concept clusters. When building a framework for comparative analysis of formal scenarios, one express the similarity between the main entities. In our earlier study [21], we approximated the meanings of mental entities using definitions from the basis of want-know-believe; however, we observed that this approach would be too coarse for recognizing complaints. In this study, we extend the speech act theory-based set of attributes to build an adequate concept lattice for communicative actions. This extension turned out to be more suitable for CRM than our earlier approaches to define a concept lattice for scenarios while learning. There exists a number of settings in which graph-based data mining and clustering is performed (e.g., [26] and [7]) that rely on information-theoretic or error-based measures. Concepts are the basic units of thought that underlie human intelligence and communication. The study of concept formation and learning is central to cognitive informatics. The concept-based approach has been applied to human reasoning in [61] addresses basic issues of concept formation and learning from cognitive informatics perspectives; a layered model for concept formation and learning is presented. Also, in the current paper we suggested a novel approach to building a semantic network between linguistic entities on the basis of selected attributes. The choice of attributes of communicative actions in this study is motivated by the task of scenario comparison; these attributes may vary from domain to domain. Twenty selected communicative entities are roughly at the same level of generality – there are “horizontal” semantic relations between them. In this respect, we have established a link between the theory of concept structures and the speech act theory, which has been discussed [54], but not subject to computational analysis. This work sheds light on what kind of conceptual structures communicative actions are. The necessity of extending the traditional set of attributes of communicative actions for the purposes of machine learning has been demonstrated. We believe the current work is one of the first employing machine learning in the domain of multiagent interactions described in natural language. A number of studies have shown how to enable BDI-agents that learn in a particular domain (e.g., information retrieval). However, in typical BDI settings, the description of agents’ attitudes is quite limited: only their beliefs, desires and intentions are involved. Moreover, only the automated (software) agents are addressed. In this paper, we significantly extended the expressiveness of representation language for agents’ attitudes, using twenty communicative actions linked by a concept lattice. The suggested machinery can be applied to an arbitrary domain, including inter-human conflicts, obviously characterized in natural language. The evaluation of our model shows it is an adequate technique to handle such complex objects as communicative actions of scenarios of multiagent interactions, both in terms of knowledge representation and reasoning. The JSM learning approach was found suitable to relate inter-human conflict scenarios to classes. Evaluation using two datasets of banking and travel complaints showed a satisfactory performance for the decision-support mode. The suggested approach for assessing complaint validity is appropriate for deployment in CRM settings: most typical complaints are subject to automated processing, and atypical cases are handled manually. The proposed method for formal representation of conflict scenarios allows their classification, as well as an efficient user interface for complaint submission.