1.3 Computational Trust
1.3.2 Machine Learning for Trust Modeling
1.3.2.2 Machine Learning for Trust
Imagine that in a system, each agent may have a set of past transactions with other agents, we argue that by investigating useful features (e.g., con-textual information that is associated with agents and transactions) that are capable of distinguishing successful transactions from unsuccessful ones, so-phisticated machine learning algorithms can be applied to analyze past trans-actions. If these algorithms manage to model efficiently what a successful (or unsuccessful) transaction is, we can then use this to predict trustworthiness of a potential transaction. In this subsection, we discuss how machine learning can be applied to trust modeling to handle two open questions in trust
re-search: (1) estimating initial trustworthiness of the target agent in the absence of information about its past behavior (direct and/or indirect experience), and (2) capturing the target agent’s dynamic behavior in different interac-tions. The application of machine learning for other trust research questions, as well as its effectiveness in different application scenarios, is discussed in other chapters of this book.
Initial Trust Estimation
When the information about the target agent’s past behavior is not avail-able (i.e., cold start), traditional reputation based approaches cannot be di-rectly applied to assess trust. As mentioned in the earlier section, stereotyping based models such as StereoTrust [150] have advanced initial trust estimation by leveraging the trustor’s past experience with other relevant agents. How-ever, StereoTrust has its own shortcomings. For instance, it cannot tell which stereotypes are more important than others; the transaction amounts-based weight determination for stereotypes combination is straightforward, intuitive and heuristic. In order to further improve stereotyping-based trust models, machine learning algorithms are applied to more efficiently learn the trustor’s local knowledge.
MetaTrust [152, 153] is a generic machine learning framework for iden-tifying relevant features to determine trust. Specifically, a trustor uses its own previous transactions (with other agents) to build a knowledge base, and utilizes this to assess trustworthiness of a potential transaction based on associated features, which are capable of distinguishing successful transac-tions from unsuccessful ones. These features are harnessed using appropriate machine learning techniques to extract relationships between the potential transaction and previous transactions. MetaTrust is generic in the sense that various machine learning algorithms can be integrated, demonstrating that trustworthiness can be efficiently learned.3 This work uses two common but effective machine learning algorithms — linear discriminant analysis (LDA) and decision tree (DT) — as the case studies to demonstrate how MetaTrust works.
Specifically, the trustor’s past interactions (with other relevant agents) are described/characterized by a set of features. Without loss of generality, two classes are assumed: successful and unsuccessful interactions. Note that feature selection is application dependant, and all the interactions have the same feature set. For instance, in an online auction site, such features may be the price or the category of the item, or the number of items already sold by the seller (see Table 1.1 as an example in Allegro (http://allegro.pl/)). For LDA, the trustor divides his historical interactions into two groups: successful and unsuccessful. He then performs LDA on these two groups to obtain a linear classifier that allows him to estimate whether the potential interaction is likely to get classified in the successful group. For decision tree algorithms, a
3That does not mean that machine learning can always be used, since its usage still depends on the availability of appropriate information (which happens to be somewhat different from the kind of information needed in traditional trust models).
TABLE 1.1: Structure of Local Knowledge Repository
tree is first constructed based the training data. Then the algorithm classifies the potential transaction by starting from the root of the tree and moving (down) until a leaf node, i.e., get classified. Similar to MetaTrust, Burnett et al. used the M5 model tree learning algorithm to learn stereotypes [28].
With sufficient local knowledge and suitable machine learning algorithms, the trustor is able to reliably predict the trustworthiness of a potential trans-action. However, when local knowledge is insufficient, machine learning algo-rithms will perform poorly. To address this issue, a local knowledge sharing overlay network (LKSON) is constructed such that agents are able to share their local information. Different from traditional trust mechanisms where the feedback of the specific agent are shared, in this work, agents only exchange intermediate machine learning algorithm results. Such a strategy has several advantages: (1) the shared information is only the intermediate result of an algorithm so it is not easy to dig out agent’s privacy, i.e., identification; (2) since the information provider does not know whom the trustor is evaluating as well as the trustor’s local knowledge, it is difficult to send fake information to promote or to bad-mouth some specific agent; (3) a lot of computation duplication is avoided.
From a different perspective, Tang et al. [257] addressed the issue of initial trust assessment using the homophily effect. Homophily suggests that similar users are more likely to establish trust relations. For instance, people with similar tastes about fiction movies tend to trust each other. This work em-ploys low-rank matrix factorization to study trust relations. Specifically, let u = {u1, u2, ..., un} denote the set of n users. G ∈ Rn×n is a trust relation matrix where G(i, j) = 1 if ui trusts uj, and G(i, j) = 0 means no trust relation between ui and uj is observed. By performing matrix factorization (i.e., optimizing Equation 1.5), u is represented by a low-rank, user specific matrix U ∈ Rn×d, where d ≪ n is the dimensionality of latent factor vector of individual users:
L = min
U,Vk G − UVUT k2F, (1.5)
where V ∈ Rd×d is a matrix that captures correlations among low-rank rep-resentations, i.e., G(i, j) = U(i, :)VU(j, :)T. To incorporate the homophily
effect, a homophily regularization term is added to the objective function: larger ζ(i, j) is, the more likely a trust relation is established between ui and uj. ζ(i, j) can be calculated by some common measures such as Jaccard’s index and Pearson Correlation Coefficient (PCC). Equation 1.6 can be solved by applying an alternative optimization solution [55] where U and V are updated alternatingly.
We can observe that the latent representation for ui is smoothed with other users via the homophily coefficient. So even for inactive users with a few or even without any trust relations, we still can estimate their latent rep-resentations via homophily regularization, thus addressing the data sparsity issue.
Dynamic Trust Prediction
In large-scale, open systems such as online social networks, an intelligent agent may vary its behavior in different interactions with different interaction part-ners to maximize its profits. For instance, in an online auction site, a malicious seller may act honestly in selling cheap items to gather sufficient reputation and then cheat in selling an expensive item. It is thus essential to model an agent’s dynamic trust by capturing its dynamic behavior. Some early at-tempts on this issue extended the popular beta distribution-based trust models by adopting the “forgetting factor” [115, 264]. However, recent studies have showed that these approaches fail to effectively detect dynamic behavior pat-terns compared to another class of solutions that are based on Hidden Markov Model (HMM) [186].
In Moe, Tavakolifard and Knapskog [184], a trust model for multi-agent systems is developed to help the agent make optimal trust decisions over time in a dynamic environment. The target agent’s behavior is predicted according to the HMM trust estimation module following the Q-learning greedy policy.
ElSalamouny et al. [61] modeled the real dynamic behavior of an agent by HMMs. They further justified the consistency of the model by measuring the difference between real and estimated predictive probability distributions us-ing relative entropy. The works [185] and [161] demonstrate how HMM-based trust models are applied to distinct application scenarios: routing protocol design in mobile and ad-hoc networks (MANET) and Web service providers selection.
Although HMM is a promising technique to model dynamic trust, most existing models intuitively use the outcomes of the past interactions as the observation sequence. This method is effective when an agent changes its be-havior in specific patterns but is not well suited to identify implicit patterns
from the random behaviors. For instance, when most of an agent’s past inter-actions are satisfactory, it is quite challenging to detect its “sudden” behavior change. To address this issue, a context-aware HMM-based trust model is proposed [155]. Liu and Datta first argue that an agent’s dynamic behav-ior is correlated with and can be inferred (to certain extent) by interaction contextual information. Three sources where such contextual information can be extracted are identified: (1) from properties of the target agent, (2) from properties of the service/products that are provided by the target agent and (3) from properties of the target agent and other agents in the network. An HMM-based model considering such contextual information is then proposed to capture the dynamic behavior of the target agent.
Specifically, a set of interactions between the trustor and the target agent is assumed. Each interaction is associated with contextual information, as men-tioned earlier. We also assume the outcome of an interaction has l levels of trust rating: L = {L1, L2, ..., Ll} (e.g., bad, medium, good). Each level corre-sponds to a trust state of the target agent, so an l-state HMM λ is constructed.
The observation of λ is actually contextual information that is associated with each interaction. Given the model λ and a sequence of observations of past interactions Fm+1 = f0f1...fm with the target agent, the trustor is able to infer trust state sm (i.e., a certain trust rating level Li) of the target agent in the potential interaction, which is expressed by P (sm = Li|λ, Fm+1). A forward algorithm is applied to recursively derive this probability. Finally, the trustor is able to decide whether to interact with the target agent based on the trust rating which has the highest probability.
Tang et al. [259], on the other hand, tried to address the issue of dynamic trust from a very different perspective. Specifically, the authors studied online trust evolution by exploiting the dynamics of users’ preference in online review sites like Epinions.com. The proposed trust evolution framework eTrust works under the assumption that trust is strongly correlated with users’ preference similarity in rating systems [303]. That is, the more similar two users are, the greater the trust between them exists. So from this viewpoint, trust relation-ships will evolve with the dynamics of users’ preferences. For instance, when a user is interested in buying a smart phone at time t, she is likely to trust the expert reviewers in the category of “Phones”; while when she switches her interest to cars at time t + 1, she will trust the expert reviewers in the category of “Cars”.
In order to capture users’ preference evolution, and hence the dynamic trust, the rating from user i to item j is predicted by linearly combining a latent factor model, which encodes users’ preference and a neighborhood-based model, which incorporates trust information: time t, α is the weight for the two models, wv,i is the trust strength between
user i and her neighbor v (Ni is the size of user i’s neighbors). Furthermore, an exponential time function is applied to decay the influence of a past rating.
The objective function of eTrust is defined as the least square of the predicted ratings and the corresponding real ratings. Projected gradient method is used to optimize the objective function.