On-demand meal delivery (ODMD) platforms such as DoorDash and Ele.me have experienced explosive growth in recent years. Effective logistics optimization strategies that could guarantee high service standards with controlled costs are crucial for the long-term sustainability of these platforms, and yet are also non-trivial due to the nature of ODMD operations. In particular, most of the orders are not known until they are placed by the customers, and any dispatching policy that only considers known requests would risk making myopic decisions in such a setting.

By providing customers with conveniences such as easy access to an extensive variety of restaurants, effortless food ordering and fast delivery, on-demand food delivery (OFD) platforms have achieved explosive growth in recent years. A crucial machine learning task performed at OFD platforms is prediction of the Order Fulfillment Cycle Time (OFCT), which refers to the amount of time elapsed between a customer places an order and he/she receives the meal. The accuracy of predicted OFCT is important for customer satisfaction, as it needs to be communicated to a customer before he/she places the order, and is considered as a service promise that should be fulfilled as well as possible. As a result, the estimated OFCT also heavily influences planning decisions such as dispatching and routing.

Quasi-cliques are an elegant way to model dense subgraphs, with each node adjacent to at least a fraction λ ∊ (0, 1] of other nodes in the subgraph. In this paper, we focus on a new graph mining problem, the query-driven maximum quasi-clique (QMQ) search, which aims to find the largest λ-quasi-clique containing a given query node set S. This problem has many applications and is proved to be NP-Hard and inapproximable. To solve the problem efficiently in practice, we propose the notion of core tree to organize dense subgraphs recursively, which reduces the search space and effectively helps find the solution within a few tree traversals. To optimize a solution to a better solution, we introduce three refinement operations: Add, Remove and Swap. We propose two iterative maximization algorithms, DIM and SUM, to approach QMQ by deterministic and stochastic means respectively. With extensive experiments on three real datasets, we demonstrate that our algorithms significantly outperform several baselines in running time and the quality.

Online social streams such as Twitter timelines, forum discussions and email threads have emerged as important channels for information propagation. Mining transient stories and their correlations implicit in social streams is a challenging task, since these streams are noisy and surge quickly. In this paper, we propose CAST, which is a context-aware story-teller that discovers new stories from social streams and tracks their structural context on the fly to build a vein of stories. More precisely, we model the social stream as a capillary network, and define stories by a new cohesive subgraph type called (k,d)-Core in the capillary network. We propose deterministic and randomized context search to support the iceberg query, which builds the story vein as social streams flow. We perform detailed experimental study on real Twitter streams and the results demonstrate the creativity and value of our approach.

Dynamic networks are commonly found in the current web age. In scenarios like social networks and social media, dynamic networks are noisy, are of large-scale and evolve quickly. In this paper, we focus on the cluster evolution tracking problem on highly dynamic networks, with clear application to event evolution tracking. There are several previous works on data stream clustering using a node-by-node approach for maintaining clusters. However, handling of bulk updates, i.e., a subgraph at a time, is critical for achieving acceptable performance over very large highly dynamic networks. We propose a subgraph-by-subgraph incremental tracking framework for cluster evolution in this paper. To effectively illustrate the techniques in our framework, we consider the event evolution tracking task in social streams as an application, where a social stream and an event are modeled as a dynamic post network and a dynamic cluster respectively. By monitoring through a fading time window, we introduce a skeletal graph to summarize the information in the dynamic network, and formalize cluster evolution patterns using a group of primitive evolution operations and their algebra. Two incremental computation algorithms are developed to maintain clusters and track evolution patterns as time rolls on and the network evolves. Our detailed experimental evaluation on large Twitter datasets demonstrates that our framework can effectively track the complete set of cluster evolution patterns from highly dynamic networks on the fly.

Recommender systems have become very important for many online activities, such as watching movies, shopping for products, and connecting with friends on social networks. User behavioral analysis and user feedback (both explicit and implicit) modeling are crucial for the improvement of any online recommender system. Widely adopted recommender systems at LinkedIn such as "People You May Know" and "Endorsements" are evolving by analyzing user behaviors on impressed recommendation items.

Online social streams such as Twitter/Facebook timelines and forum discussions have emerged as prevalent channels for information dissemination. As these social streams surge quickly, information overload has become a huge problem. Existing keyword search engines on social streams like Twitter Search are not successful in overcoming the problem, because they merely return an overwhelming list of posts, with little aggregation or semantics. In this demo, we provide a new solution called \keysee by grouping posts into events, and track the evolution patterns of events as new posts stream in and old posts fade out. Noise and redundancy problems are effectively addressed in our system. Our demo supports refined keyword query on evolving events by allowing users to specify the time span and designated evolution pattern. For each event result, we provide various analytic views such as frequency curves, word clouds and GPS distributions. We deploy \keysee on real Twitter streams and the results show that our demo outperforms existing keyword search engines on both quality and usability.

Search for objects similar to a given query object in a network has numerous applications including web search and collaborative filtering. We use the notion of structural similarity to capture the commonality of two objects in a network, e.g., if two nodes are referenced by the same node, they may be similar. Meeting-based methods including SimRank and P-Rank capture structural similarity very well. Deriving inspiration from PageRank, SimRank has gained popularity by a natural intuition and domain independence. Since it's computationally expensive, subsequent work has focused on optimizing and approximating the computation of SimRank. In this paper, we approach SimRank from a top-k querying perspective where given a query node v, we are interested in finding the top-k nodes that have the highest SimRank score w.r.t. v. The only known approaches for answering such queries are either a naive algorithm of computing the similarity matrix for all node pairs or computing the similarity vector by comparing the query node v with each other node independently, and then picking the top-k. None of these approaches can handle top-k structural similarity search efficiently by scaling to very large graphs consisting of millions of nodes. We propose an algorithmic framework called TopSim based on transforming the top-k SimRank problem on a graph G to one of finding the top-k nodes with highest authority on the product graph G G. We further accelerate Top Sim by merging similarity paths and develop a more efficient algorithm called Top Sim-SM. Two heuristic algorithms, Trun-Top Sim-SM and Prio-Top Sim-SM, are also proposed to approximate Top Sim-SM on scale-free graphs to trade accuracy for speed, based on truncated random walk and prioritizing propagation respectively. We analyze the accuracy and performance of Top Sim family algorithms and report the results of a detailed experimental study.