Bayesian network (BN) is a well-accepted model to represent a set of random variables and their probabilistic relationships via a directed acyclic graph (DAG) [1] [2] . There is a conditional probability table (CPT) associated with each node to represent the quantitative relationships among the nodes. BN has been widely used in different aspects of intelligent applications including representation, prediction, reasoning, and so on [1] [3] . However, exact inference in an arbitrary BN is NP hard [4] , and even the approximate inference is also very difficult [5] . As a result, BN’s inferences are not very efficient, which is doomed to pose limits to BN-related applications, especially for real-time situations. Fortunately, as the qualitative abstraction of BN, QPN was proposed by Wellman [6] and can be appropriately used to remedy the above drawbacks of a BN. A QPN adopts a DAG to encode random variables and their relationships, which are not quantified by conditional probabilities, but are summarized by qualitative signs [7] . The most important thing is that there has been an efficient QPN inference (reasoning) algorithm, which is based on sign propagation and run in polynomial time [8] .

However, the high abstraction level of QPN results in the problem that there is no information left to compare two different qualitative influences in a QPN. Thus, when a node receives two inconsistent signs from its two different neighbor nodes during inferences, it is hard to know which sign is more suitable for the node, so that the “? (i.e., unknown)” sign is obtained as the result sign on the node. This means that there are inference conflicts generated on the node, which lead to less powerful expressiveness and inference capabilities than expected for real-world prediction and decision-making of uncertain knowledge in economics, health care, traffics, etc [9] - [16] . Worst of all, when the ambiguous results caused by inference conflicts are produced, they will be spread to most parts of the network with the reasoning algorithm going on.

To provide a conflict-free inference mechanism of QPN is paid much attention. Various approaches have been proposed from various perspectives. However, in some representative methods [9] [11] [13] , the original sample data or threshold values are required or predefined when the influences are weighted quantitatively. This is not consistent with the BN-based applications that focus on probabilistic inferences taking as input BN without the original sample data. Furthermore, a BN is frequently defined by experts instead of learned from data. Even though the original data are available, the efficiency of the process of deriving influence weights is sensitive to the data size. Meanwhile, the uncertainty of the influence weight has not been well incorporated, and the propagation of influence weights lacks solid theoretical foundations during QPN inferences.

Therefore, in this paper we are to consider eliminating the inference conflicts of general QPNs and developing a conflict-free inference method. We derive the quantitative QPN by adding a weight to each qualitative influence from the corresponding BN, while sample data and threshold values are not required. The weight is adopted as the information to compare two different qualitative signs. Therefore, when a node in a general QPN faces an inference conflict, a trade-off will be incorporated to avoid the conflict based on the weights.

Mutual information (MI) is a quantity that measures the mutual dependence of the two random variables [17] [18] . As well, MI is widely applied to test the association degree among the BN nodes [19] - [21] . When node A receives two inconsistent signs from its two different neighbor nodes, say B and C, we can naturally think that the result sign on A is more dependent on the sign propagated from B than that from C, if the mutual information between A and B is larger than that between A and C. Based on the above idea, we adopt MI as the weight to quantify the degree of the mutual dependence between two nodes linked by a directed edge.

Generally speaking, the main contributions of this paper can be summarized as follows:

Ÿ We define the MI-based weight of a qualitative influence, called MI weight, and extend the traditional QPN by adding a MI weight to each qualitative influence. We call the extended QPN as MI-QPN.

Ÿ We propose an efficient algorithm to derive MI weights from the conditional probability tables and prior probability distributions in the corresponding BN instead of the sample data.

Ÿ We discuss the symmetry, transitivity and composition properties of qualitative influences in the MI-QPN. Then, we extend the general QPN’s inference algorithm to achieve conflict-free inferences with the MI- QPN.

Ÿ We give preliminary experiments to verify the feasibility and correctness of our method.

BN has been successfully established as a framework to describe, manage uncertainty using the probabilistic graphical approach [2] [3] . As the qualitative abstraction of BN, QPN [6] was proposed by Wellman and widely used in decision making, industrial control, forecasting, and so on [22] - [24] . The methods to construct QPNs mainly include deriving QPNs from the corresponding existing BNs [15] or giving QPNs by the domain experts. As for the QPN’s inference, an efficient algorithm was proposed by Druzdzel et al. [8] . The inference conflict has become a common problem for QPNs and how to resolve the inference conflicts is the subject that was paid much attention in recent years. Various methods were proposed from various perspectives.

Lv et al. [9] proposed the ambiguity reduction method by associating a qualitative mutual information weight to each qualitative influence in the QPN, but the weight can only be obtained from the given or generated sample data. This is not always feasible, since the BN may be not learned from data, or the sample data cannot be available. Parsons [10] introduced the concept of categorical influence, which is either an influence that serves to increase a probability to 1 or an influence that decreases a probability to 0, regardless of any other influences. By this approach, only some of the inference conflicts can be resolved.

Renooij et al. [11] [15] proposed two methods to solve the inference conflicts by using both kappas values and pivots to zoom upon inferences. The kappas value based method distinguishes the strong and weak influences by the interval values that do not have solid theoretical basis for interval value propagation during QPN inferences. Renooij et al. [12] presented an enhanced QPN that differs from a regular QPN to distinguish between strong and weak influences. This method is only suitable for the binary-variable situation. A threshold value is required in these methods except that exploits context-specific information. Renooij et al. [13] proposed SQPN (semi-QPN) by associating a probability interval value with each qualitative influence to quantify the qualitative influences. De Campos et al. [25] discussed the complexity of inferences in polytree-shaped SQPNs. Renooij et al. [14] extended QPNs by providing the inclusion of context-specific information about influences and showed that exploiting this information upon inference had the ability to forestall unnecessarily weak results. By this method, a threshold value should be given to distinguish the strong and weak influences.

We extended QPN to solve inference conflicts by adding the weights to qualitative influences based on the rough set theory and interval probability theory respectively in [16] and [26] . By the method in [16] , the dependency degree was computed based on the threshold value, but the corresponding sample data was required to obtain the weights. By the method in [26] , the interval-probability weights were derived from the corresponding BN’s CPT, but the calculation of interval probability during inferences led to approximate results.

It is known that implies according to the symmetry property. Then, we know there exist in the MI-QPN by Definition 4.4, since exists in the corresponding QPN. We can obtain

Algorithm 1. Deriving MI weights from a BN.

In order to address the transitivity property, we consider the MI-QPN fragment in Figure 4, where X denotes the predecessors of other than, and represents the predecessors of other than. From this

network, we have and. Based on and, we

are to derive the qualitative influence of on with an MI weight, denoted as.

First, based on QPN’s transitivity property, we can obtain. Then, we derive the MI weight of qualitative influence on,. The transitivity property is denoted as

In order to discuss the composition property, we consider the MI-QPN fragment in Figure 5, where X denotes

the predecessors of other than and. From the MI-QPN, we have and. Then, we are to derive the combination influence of and, denoted as

where, denotes the composition weight and “Ú” is the composition operator.

Intuitively, the composition operator “Ú” ought to satisfy the following properties: 1) The composition weight belongs to [0, 1]; 2) The composition operation is commutative; 3) The composition operation is associative; 4) Combining two influences with the same qualitative signs (e.g., two “+” signs or two “−” signs) will result in an influence with a greater MI weight; 5) Combining two influences with different qualitative signs will result in an influence dependent on but less than the larger one. Inspired by the evidence theory and the basic idea of evidence combination [27] , we define the composition operator “Ú” based on evidence superposition.

Definition 5.1, defined as follows:

・ if, then

・ if and, then

・ if and, then

・ if and, then

Based the above properties, we give Algorithm 2 for conflict-free inferences with the MI-QPN.

Now, we discuss the time complexity of Algorithm 2 for MI-QPN inferences. First, whether a node will be visited is determined by its node sign whose changes are specified in QPN’s general inference algorithm [8] , while the influence weights are just used to decide the node sign during inferences. Second, Algorithm 2 is the same as the general QPN’s inference algorithm if there does not exist “?” during inferences [8] . However, Algorithm 2 can be used to avoid the “?” results as possible by means of weighting influences quantitatively, where each time of weight computation w.r.t. the composition property can be fuelled by Step 5 in Merge-Sign in O(1) time. Therefore, the time complexity of Algorithm 2 will be the same as that for the general QPN’s inferences.

Algorithm 2. Conflict-free inference with an MI-QPN.

To test the performance of our method in this paper, we implemented our algorithms for constructing and inferring MI-QPN. We take BN’s inference results obtained from Netica [28] as the criteria to decide whether the increase (or decrease) trends of the nodes indicated by the inference results of the corresponding MI-QPN are correct.

It is well known that Wet-Grass network is a classic BN for whether the Our Grass is wet, related to Our Wall, Rain, Our Sprinkler and Neighbor’s Grass status, containing 5 binary variables and 6 edges [2] [28] . From the Wet-Grass BN, we know the qualitative influences between node Rain and Neighbor’s Grass and the ones between Rain and Our Grass are the same. To illustrate our methods and show the results straightforwardly, we modify the knowledge in the network by removing node Neighbor’s Grass and the corresponding edge, which will not affect the rest part of the original network and we denote the possible values of the nodes wet, was on, rained as 1, and dry, was off, didn’t rain as 0. Moreover, we rename Our Wall as Wall or W, Our Grass as Grass or G and Our Sprinkler as Sprinkler or S for short. Then, the modified network is shown in Figure 6.

First, we derive the corresponding QPN and MI-QPN shown in Figure 7 and Figure 8 respectively. Then, we compare the inference results of the QPN and those of the MI-QPN to verify the feasibility of our algorithms. By using the general QPN’s inference algorithm [8] and our inference algorithm for an MI-QPN, we take each node in the QPN and the MI-QPN as evidence and record the inference results of other nodes. The comparisons are shown in Table 1, from which we can see all possible inference conflicts have been eliminated in the MI-QPN.

Then, we compare the inference results on the modified Wet-Grass BN and those on the derived MI-QPN. We take each node of the BN as evidence, and then record the inference results of other nodes shown in Table 2, where E-Node and E-Sign denotes the evidence and corresponding sign respectively. The E-sign 0 → 1 represents that the state of the evidence node changes from 0 to 1, indicating that the probability of state 1 is changed from 0 to 1. If a node in the MI-QPN takes sign “+” (“−”) as the final inference result and the increase (decrease) of the probability of state 1 of the corresponding node in the BN, we can conclude that the inference result is correct.

In this paper, we introduced mutual-information based weights (MI weights) to qualitative influences in QPNs to resolve conflicts during the inferences. We first defined the MI weights based on mutual information and MI-QPN by extending the traditional QPN. Then, we proposed the method to derive the MI weights for the MI-QPN from the corresponding BN without sample data or threshold values. By theoretic analysis, we know the method for deriving MI weights is effective. Furthermore, we discussed the symmetry, transitivity and composition properties in the MI-QPN, and extended the general influence algorithm to implement the conflict-free inferences of MI-QPN. The feasibility of our method was verified by the results of the preliminary experiment.

Our work in this paper also leaves open some other interesting research issues. We are to further consider adding the MI weights to the qualitative synergies and discussing the method to resolve the inference conflicts caused by two inconsistent signs with the same MI weight. As well, we will further resolve the conflicts that take place during the fusion or integration of multiple QPNs by adding the MI weights to the qualitative influences in the QPNs. These are exactly our future work.

This work was supported by the National Natural Science Foundation of China (61472345), the Natural Science Foundation of Yunnan Province (2014FA023, 2013FA013), the Yunnan Provincial Foundation for Leaders of Disciplines in Science and Technology (2012HB004), and the Program for Innovative Research Team in Yunnan University (XT412011).