Natural language understanding models often make use of surface patterns or idiosyncratic biases in a given dataset to make predictions pertaining to natural language inference (NLI) tasks. Unfortunately, this renders the resulting model vulnerable to out-of-distribution datasets to which the identified features are inapplicable, thereby leading to erroneous results. Many of the methods developed for out-of-distribution datasets have proven effective; however, they also tend to impose a tradeoff in performance when applied to in-distribution datasets. In this paper, we use a teacher model providing knowledge for the student ensemble model as basic information for training. The student ensemble model then integrates information of deep and shallow representations to extend learning performance to a wide range of examples. The evaluation demonstrates that the proposed model outperformed state-of-the-art models when applied to in-distribution as well as out-of-distribution datasets.

Most sequential recommendation systems, including those that employ a variety of features and state-of-the-art network models, tend to favor items that are the most popular or of greatest relevance to the historic behavior of the user. Recommendations made under these conditions tend to be repetitive; i.e., many options that might be of interest to users are entirely disregarded. This paper presents a novel algorithm that assigns a novelty score to potential recommendation items. We also present an architecture by which to incorporate this functionality in existing recommendation systems. In experiments, the proposed NASM system outperformed state-of-the-art sequential recommender systems, thereby verifying that the inclusion of novelty score can indeed improve recommendation performance.

Recommendation systems which are designed to understand and predict user interest based on user preferences play an important role in the era of information explosion. We propose the item influence embedding which adopts the social influence diffusion concept to model the item relations. We can learn the activation paths in items-item relation graph. In addition, for generating top-k items, most of recommendation systems calculate the similarity between user embedding and embedding of all items. The calculation costs too much time when number of users and items are huge. Therefore, we propose the User Preference Translation Model (UPTM) to recommend the Top-k items based on the language translation technology. UPTM directly generates the recommendation items based on translating the user preference. We can avoid to calculate the similarity of user embedding and item embedding. From the experimental results, UPTM not only outperforms the compared methods but also save the time in real large datasets.

Influence maximization refers to the process of identifying a predefined number of nodes within a given social network with the aim of maximizing the spread of influence. Most previous work has focused on unsigned networks, which means the existence of polarity relationships has largely been disregarded. In this work, we define a Sign-aware Influence Maximization (SIM) problem, which involves identification of the seed set that would simultaneously maximize positive influence and minimize negative influence. We begin by considered competitive influence under various dominance mechanisms on SCIC model, which extends the classic Independent Cascade (IC) model by incorporating binary opinions and signed relationships. We then proved that the influence of SIM under the SCIC model is non-monotonic and non-submodular, which implies that simple greedy hill-climbing would be unable to achieve an approximation ratio of 1-1/e in seeking to resolve the SIM problem. We then developed a simulation-based algorithm called Sign-aware Competitive Maximum Influence Arborescence (S-CMIA) to simulate the propagation of influence within a local region. Experiment results demonstrate the superiority of the proposed algorithm over existing methods in resolving the SIM problem in terms of reward.

In influence maximization problem, we would like to find an initial subset of nodes in a given graph, which maximizes the final number of affected nodes through "word of mouth" propagation. Measuring the influence spread of set of seed nodes and gradually selecting the node with largest marginal increase is one of the main approaches of existing algorithms. In this paper, we try to solve this problem from a different strategy - in an improvement perspective. A more complex condition is depicted as both positive and negative opinions are propagating in the social network. The objective function considers the maximization of the positive influence and minimizes the negative opinion spreading simultaneously. We propose IDR (Influence Distribution Redirection) algorithm to define initial seed nodes of influence diffusion based on redirecting the influence distribution of nodes to maximize the objective function. The influence distribution of nodes shows the potential influence trend of nodes during the influence diffusion process. The key strategy is reducing the positive influence nearby the steady nodes and increasing in the vacillate region. From the experimental results, IDR outperforms the compared method on the objective function. In addition, IDR also improves the performance of increasing the number of positive active nodes and decreasing the number of negative activated nodes respectively.

The influence maximization problem has been considered a vital problem when companies provide similar products or services. Since there are limited resources, companies must determine a strategy to occupy as much market share as possible. In this paper, we propose a general Deterministic Network Adaptive (DNA) framework to solve the multi-round multi-party influence maximization problem. To obtain the most market share, using one single strategy to determine seed nodes is not sufficient in the long term. The reason is that the network status changes during the multi-round procedure. The strategies of selecting seed nodes in each round should depend on the current status of influence diffusion in the network. DNA framework leverages the concept of reinforcement learning to maximize the expected cumulative influence. In addition, the learning process is deterministic, so that it does not take time to explore the spaces that are less important. We further design a similarity function to measure the similarity between two networks. DNA framework can avoid redundant computation when the similar networks have been trained before. Moreover, we propose the method to make the policy decision to maximize the influence spread in coopetition scenario based on DNA framework. The proposed framework is evaluated with synthetic data and real-world data. From the experimental results, DNA framework outperforms the existing works in influence maximization problems. The coopetition policy which is generated by DNA has the best performance in most cases.

Since the tremendous success of social networking websites, the related analytical research has been widely studied. Among these studies, social influence has been a significant and popular topic. We rely on the social influence model to predict and learn the influence diffusion process. However, traditional models only categorize nodes into two types of states, active and inactive. In addition, most previous models have only taken positive influences into account. Moreover, if inactive nodes are influenced successfully and turn into active nodes, these nodes cannot change their states forever. In this work, we not only break the above limitations but also propose a novel propagation method in our model. We proposes five states to represent the multiple states of influence. According to the new propagation method, the strength of the social influence may be reduced over time. Eventually, we utilize the measurement of precisions to compare with related models. The proposed multi-state model outperforms other two-state models in precisions of prediction. The experimental results show the superiority of multiple states.