Customer service is how companies interface with their customers. It can contribute heavily towards the overall customer satisfaction. However, high-quality service can become expensive, creating an incentive to make it as cost efficient as possible and prompting most companies to utilize AI-powered assistants, or "chat bots". On the other hand, human-to-human interaction is still desired by customers, especially when it comes to complex scenarios such as disputes and sensitive topics like bill payment.

Session-based recommendation aims to predict the next item in an anonymous session. Recent advances have shown the importance of exploiting inter-session dependencies, such as item-item transitions and session-session similarities. However, the existing methods either ignore the relative order of item co-occurrences or assign the same importance to co-occurrence patterns at all distances. Besides, they are prone to extracting wrong signals to learn user preferences from dependencies between sessions. To solve these problems, we propose a model called FOCOL to better exploit the intersession dependencies by considering Fine-grained item co-Occurrences and applying the COntrastive Learning framework. Specifically, to capture inter-session item-item dependencies, we propose a component called FOGCN (Fine-grained co-Occurrence Graph Convolution Network) to automatically learn the importance of item co-occurrence patterns from a global graph that encodes the detailed information about item co-occurrences such as relative order and distance. To directly capture dependencies between sessions, we view the recommendation task as a clustering problem, and propose a component called CSRL (Contrastive Session Representation Learning) to implicitly group similar sessions (i.e., sessions with the same next item) into the same cluster and push apart sessions at different clusters. Extensive experiments conducted on three public datasets show that the proposed model is superior to the state-of-the-art methods and the proposed two components can learn more informative item and session representations by considering the fine-grained item co-occurrences and directly capturing dependencies between sessions.

In many applications of session-based recommendation, social networks are usually available. Since users' interests are influenced by their friends, recommender systems can leverage social networks to better understand their users' preferences and thus provide more accurate recommendations. However, existing methods for session-based social recommendation are not efficient. To predict the next item of a user's ongoing session, the methods need to process many additional sessions of the user's friends to capture social influences, while non-social-aware methods (i.e., those without using social networks) only need to process one single session. To solve the efficiency issue, we propose an efficient framework for session-based social recommendation. In the framework, first, a heterogeneous graph neural network is used to learn user and item representations that integrate the knowledge from social networks. Then, to generate predictions, only the user and item representations relevant to the current session are passed to a non-social-aware model. During inference, since the user and item representations can be precomputed, the overall model runs as fast as the original non-social-aware model, while it can achieve better performance by leveraging the knowledge from social networks. Apart from being efficient, our framework has two additional advantages. First, the framework is flexible because it is compatible with any existing non-social-aware models and can easily incorporate more knowledge other than social networks. Second, our framework can capture cross-session item transitions while existing methods can only capture intra-session item transitions. Extensive experiments conducted on three public datasets demonstrate the effectiveness and efficiency of the proposed framework. Our code is available at https://github.com/twchen/SEFrame.

Recently, graph neural networks (GNNs) have gained increasing popularity due to their convincing performance in various applications. Many previous studies also attempted to apply GNNs to session-based recommendation and obtained promising results. However, we spot that there are two information loss problems in these GNN-based methods for session-based recommendation, namely the lossy session encoding problem and the ineffective long-range dependency capturing problem. The first problem is the lossy session encoding problem. Some sequential information about item transitions is ignored because of the lossy encoding from sessions to graphs and the permutation-invariant aggregation during message passing. The second problem is the ineffective long-range dependency capturing problem. Some long-range dependencies within sessions cannot be captured due to the limited number of layers. To solve the first problem, we propose a lossless encoding scheme and an edge-order preserving aggregation layer based on GRU that is dedicatedly designed to process the losslessly encoded graphs. To solve the second problem, we propose a shortcut graph attention layer that effectively captures long-range dependencies by propagating information along shortcut connections. By combining the two kinds of layers, we are able to build a model that does not have the information loss problems and outperforms the state-of-the-art models on three public datasets.

Session-based recommendation is a task to predict users' next actions given a sequence of previous actions in the same session. Existing methods either encode the previous actions in a strict order or completely ignore the order. However, sometimes the order of actions in a short sub-sequence, called the detailed order, may not be important, e.g., when a user is just comparing the same kind of products from different brands. Nevertheless, the high-level ordering information is still useful because the data is sequential in nature. Therefore, a good session-based recommender should pay different attention to the sequential information in different levels of granularity. To this end, we propose a novel model to automatically ignore the insignificant detailed ordering information in some sub-sessions, while keeping the high-level sequential information of the whole sessions. In the model, we first use a full self-attention layer with Gaussian weighting to extract features of sub-sessions, and then we apply a recurrent neural network to capture the high-level sequential information. Extensive experiments on two real-world datasets show that our method outperforms or matches the state-of-the-art methods.

When faced with a database containing millions of tuples, an end user might be only interested in finding his/her favorite tuple in the database. In this paper, we study how to help an end user to find such a favorite tuple with a few user interactions. In each interaction, a user is presented with a small number of tuples (which can be artificial tuples outside the database or true tuples inside the database) and s/he is asked to indicate the tuple s/he favors the most among them. Different from the previous work which displays artificial tuples to users during the interaction and requires heavy user interactions, we achieve a stronger result. Specifically, we use a concept, called the utility hyperplane, to model the user preference and an effective pruning strategy to locate the favorite tuple for a user in the whole database. Based on these techniques, we developed an interactive system, called FindYourFavorite, and demonstrate that the system could identify the favorite tuple for a user with a few user interactions by always displaying true tuples in the database.

Little is currently known about dynamic brain networks involved in high-level cognition and their ontological basis. Here we develop a novel Variational Bayesian Hidden Markov Model (VB-HMM) to investigate dynamic temporal properties of interactions between salience (SN), default mode (DMN), and central executive (CEN) networks-three brain systems that play a critical role in human cognition. In contrast to conventional models, VB-HMM revealed multiple short-lived states characterized by rapid switching and transient connectivity between SN, CEN, and DMN. Furthermore, the three "static" networks occurred in a segregated state only intermittently. Findings were replicated in two adult cohorts from the Human Connectome Project. VB-HMM further revealed immature dynamic interactions between SN, CEN, and DMN in children, characterized by higher mean lifetimes in individual states, reduced switching probability between states and less differentiated connectivity across states. Our computational techniques provide new insights into human brain network dynamics and its maturation with development.