Artificial Intelligence (AI) research in Canada is driving a vibrant ecosystem of startups and industry innovation, thanks significantly to the leadership of Canada CIFAR AI Chairs. Canada AI Day at KDD 2025 will showcase cutting-edge research by some of these leading experts, with a focus on building ethical, interpretable, and accessible AI systems. The event will feature a series of invited talks followed by a panel discussion bringing together academic, government, and industry researchers to address the challenges and opportunities in developing effective and responsible AI.

This paper introduces DAMOCRO, a data migration framework using online classification and tuple reordering to improve throughput and decrease the costs of data migration. The DAMOCRO workflow consists of four main steps. First, it classifies records into subgroups to maximize the similarity within each group. Next, it reorders tuples within these groups, ensuring that similar tuples are adjacent. Subsequently, column-wise compression is applied to each group. Finally, the compressed data is transferred from the source to the target machine. The initial two steps enhance the compression ratio, thereby boosting throughput and reducing costs. Our evaluations on five real-world datasets and two benchmark datasets, show that the online classification process in DAMOCRO improves throughput by more than 24% and reduces costs by over 19% compared to baselines. Besides, implementing reordering based on functional dependencies brings an additional cost reduction ranging from 10% to 60%, while also enhancing throughput.

Graph representation learning has been a very active research area in recent years. The goal of graph representation learning is to generate graph representation vectors that capture the structure and features of large graphs accurately. This is especially important because the quality of the graph representation vectors will affect the performance of these vectors in downstream tasks such as node classification, link prediction and anomaly detection. Many techniques have been proposed for generating effective graph representation vectors, which generally fall into two categories: traditional graph embedding methods and graph neural network (GNN)–based methods. These methods can be applied to both static and dynamic graphs. A static graph is a single fixed graph, whereas a dynamic graph evolves over time and its nodes and edges can be added or deleted from the graph. In this survey, we review the graph-embedding methods in both traditional and GNN-based categories for both static and dynamic graphs and include the recent papers published until the time of submission. In addition, we summarize a number of limitations of GNNs and the proposed solutions to these limitations. Such a summary has not been provided in previous surveys. Finally, we explore some open and ongoing research directions for future work.

The advent of transformers and the subsequent development of Large Language Models (LLMs) based on these technologies has revolutionized the field of Natural Language Processing (NLP). These models are able to understand and generate coherent natural language and hold conversations with humans continuously. Meanwhile, ChatGPT has become famous among many LLMs for its general-purpose characteristics and versatility. With that in mind, we investigate the capabilities of ChatGPT, which is very successful in many downstream NLP tasks on the task of Question Generation (QG). In particular, our experiments show that appropriate context through our designed prompts makes ChatGPT an appropriate tool for accurately performing the QG task. We compare ChatGPT’s question generation results with the state-of-the-art models, particularly on the SQuAD and car manual datasets. The results show that ChatGPT is able to compete with or even outperform some of the baseline models. Furthermore, we illustrate that we may improve ChatGPT through additional fine-tuning of the prompts. Finally, we also investigate the use of ChatGPT to evaluate QG models. While the use of ChatGPT for such purposes is still in its early stages, our results demonstrate that ChatGPT can potentially be a strong QG accuracy evaluator comparable to human evaluators.

Graph Neural Networks (GNNs) have been proposed to learn graph representations for various graph mining tasks such as link prediction and node classification. These methods aggregate information from neighbors of a node to generate the node representation vector. Temporal GNN models consider the temporal and neighborhood information of nodes. However, few temporal GNN methods consider network substructures such as triads and cliques. In this paper, we present a temporal GNN-based method that generates node embeddings by aggregating neighbors of a node that exist in the maximal cliques of the graph containing the node. The reason for considering neighbors that form a maximal clique with the target node is that nodes in a maximal clique are highly connected to each other and most likely share similar characteristics. In addition, we consider the time dependency of nodes by generating temporal walks on the cliques such that in these walks the time order of the nodes is respected. The node embedding is based on the aggregation of the node’s temporal walks. Our experiments on seven datasets show the effectiveness of our method in both link prediction and node classification tasks. Furthermore, our method is faster than other baselines we evaluate.

Many real-world data are considered as graphs, such as computer networks, social networks and protein-protein interaction networks. Graph embedding methods are powerful tools for representing large graphs in various domains. A graph embedding method projects the components of a graph, such as its nodes or edges, into a vector space with a lower dimensionality than the adjacency matrix of the graph, and aims to preserve the characteristics of the graph. The generated embedding vectors have been utilized in various graph mining applications such as node classification, link prediction and anomaly detection. Despite the wide success of the graph embedding methods, little study has been done to facilitate a better understanding of the graph embeddings. In this paper, inspired by advancements in interpreting word embeddings, we propose two interpretability measures to quantify the interpretability of graph embeddings by leveraging useful network centrality properties and perform comparisons of different graph embedding methods. Using these scores, we can provide insights into the representational power of graph embedding methods.

The bulk synchronous parallel (BSP) is a celebrated synchronization model for distributed training of deep learning models. A shortcoming of the BSP is that it requires workers to wait for the straggler at every iteration. Therefore, employing BSP increases the waiting time of the faster workers of a cluster and results in an overall prolonged training time. To ameliorate this shortcoming of BSP, we proposed ElasticBSP [1], a model that aims to relax its strict synchronization requirement with an elastic synchronization by allowing delayed synchronization to minimize the waiting time. ELASTICBSP is realized by the algorithm named ZipLine. In this work, we show the theoretical proof of ZipLine and further propose algorithmic and implementation optimizations of ZipLine, namely ZipLineOpt and Ziplineoptbs, which reduce the time complexity of ZipLine to linearithmic time. The experiments show that ZipLineOpt and ZipLineOptBs enable the scalability of ElasticBSP. Further experimental evaluation on large deep neural networks on large ImageNet dataset demonstrate that our proposed Elas-ticbspmodel, materialized by the proposed optimized ZipLine variants, converges faster and to a higher accuracy than the predominant BSP.

Subscription-based online newspapers usually offer non-subscribed users a certain number of free articles in a period of time, and then directs them to a page (called paywall) asking for subscription. This approach (also known as metered or fixed paywall) does not consider the user's reading history nor the articles that the user may read in the future, and consequently, it may disengage many potential subscribers. To that end, we propose adaptive paywall mechanisms to make optimal paywall decisions (i.e., showing the article or the paywall) by balancing the benefit of showing the article against that of presenting the paywall. We define the notions of utility and cost which are used to define an objective function for the optimal paywall decision problem. We propose the Lookahead policy (LAP) and QPaywall policy (QP) as two data-driven approaches to solve the adaptive paywall problem. While the LAP method makes paywall decisions on the fly by simulating trajectories of article requests using Monte Carlo sampling, the QP approach is based on reinforcement learning and learns a neural network-based action-value (Q) function for this purpose. We compare advantages of the proposed approaches and discuss the practical considerations of using them in a real environment. Empirical studies on a real dataset from a major newspaper in Canada show that the proposed methods outperform several baseline approaches in terms of various business objectives.

One popular thread of research in computational sarcasm detection involves modeling sarcasm as a contrast between positive and negative sentiment polarities or exploring more fine-grained categories of emotions such as happiness, sadness, surprise, and so on. Most current models, however, treat these affective features independently, without regard for the sequential information encoded among the affective states. In order to explore the role of transitions in affective states, we formulate the task of sarcasm detection as a sequence classification problem by leveraging the natural shifts in various emotions over the course of a piece of text. Experiments conducted on datasets from two different genres suggest that our proposed approach particularly benefits datasets with limited labeled data and longer instances of text.

Many real world networks are very large and constantly change over time. These dynamic networks exist in various domains such as social networks, traffic networks and biological interactions. To handle large dynamic networks in downstream applications such as link prediction and anomaly detection, it is essential for such networks to be transferred into a low dimensional space. Recently, network embedding, a technique that converts a large graph into a low-dimensional representation, has become increasingly popular due to its strength in preserving the structure of a network. Efficient dynamic network embedding, however, has not yet been fully explored. In this paper, we present a dynamic network embedding method that integrates the history of nodes over time into the current state of nodes. The key contribution of our work is 1) generating dynamic network embedding by combining both dynamic and static node information 2) tracking history of neighbors of nodes using LSTM 3) significantly decreasing the time and memory by training an autoencoder LSTM model using temporal walks rather than adjacency matrices of graphs which are the common practice. We evaluate our method in multiple applications such as anomaly detection, link prediction and node classification in datasets from various domains.

The bulk synchronous parallel (BSP) is a celebrated synchronization model for general-purpose parallel computing that has successfully been employed for distributed training of machine learning models. A prevalent shortcoming of the BSP is that it requires workers to wait for the straggler at every iteration. To ameliorate this shortcoming of classic BSP, we propose ELASTICBSP a model that aims to relax its strict synchronization requirement. The proposed model offers more flexibility and adaptability during the training phase, without sacrificing on the accuracy of the trained model. We also propose an efficient method that materializes the model, named ZIPLINE. The algorithm is tunable and can effectively balance the trade-off between quality of convergence and iteration throughput, in order to accommodate different environments or applications. A thorough experimental evaluation demonstrates that our proposed ELASTICBSP model converges faster and to a higher accuracy than the classic BSP. It also achieves comparable (if not higher) accuracy than the other sensible synchronization models.

Network representation learning in low dimensional vector space has attracted considerable attention in both academic and industrial domains. Most real-world networks are dynamic with addition/deletion of nodes and edges. The existing graph embedding methods are designed for static networks and they cannot capture evolving patterns in a large dynamic network. In this paper, we propose a dynamic embedding method, dynnode2vec, based on the well-known graph embedding method node2vec. Node2vec is a random walk based embedding method for static networks. Applying static network embedding in dynamic settings has two crucial problems: 1) Generating random walks for every time step is time consuming 2) Embedding vector spaces in each timestamp are different. In order to tackle these challenges, dynnode2vec uses evolving random walks and initializes the current graph embedding with previous embedding vectors. We demonstrate the advantages of the proposed dynamic network embedding by conducting empirical evaluations on several large dynamic network datasets.

Many online news agencies utilize the paywall mechanism to increase reader subscriptions. This method offers a non-subscribed reader a fixed number of free articles in a period of time (e.g., a month), and then directs the user to the subscription page for further reading. We argue that there is no direct relationship between the number of paywalls presented to readers and the number of subscriptions, and that this artificial barrier, if not used well, may disengage potential subscribers and thus may not well serve its purpose of increasing revenue. Moreover, the current paywall mechanism neither considers the user browsing history nor the potential articles which the user may visit in the future. Thus, it treats all readers equally and does not consider the potential of a reader in becoming a subscriber. In this paper, we propose an adaptive paywall mechanism to balance the benefit of showing an article against that of displaying the paywall (i.e., terminating the session). We first define the notion of cost and utility that are used to define an objective function for optimal paywall decision making. Then, we model the problem as a stochastic sequential decision process. Finally, we propose an efficient policy function for paywall decision making. The experimental results on a real dataset from a major newspaper in Canada show that the proposed model outperforms the traditional paywall mechanism as well as the other baselines.

Sarcasm detection from text has gained increasing attention. While one thread of research has emphasized the importance of affective content in sarcasm detection, another avenue of research has explored the effectiveness of word representations. In this paper, we introduce a novel model for automated sarcasm detection in text, called Affective Word Embeddings for Sarcasm (AWES), which incorporates affective information into word representations. Extensive evaluation on sarcasm detection on six datasets across three domains of text (tweets, reviews and forum posts) demonstrates the effectiveness of the proposed model. The experimental results indicate that while sentiment affective representations yield best results on datasets comprising of short length text such as tweets, richer representations derived from fine-grained emotions are more suitable for detecting sarcasm from longer length documents such as product reviews and discussion forum posts.

User acquisition is one of the most challenging problems for online news providers. In fact, due to availability of different news media, users have a lot of choices in selecting the news source. To date, most of digital news portals have tried to approach the solution indirectly by targeting the user satisfaction through the recommendation systems. In contrast, we address the problem directly by identifying valuable visitors who are likely potential subscribers in the future. First, we suggest that the decision for subscription is not a sudden, instantaneous action, but is the informed decision based on positive experience with digital medium. As such, we propose effective engagement measures and show that they are effective in building the predictive model for subscription. We design a model that not only predicts the potential subscribers but also answers queries about the subscription occurrence time. The proposed model can be used to predict the subscription time and recommend accurately the “potential users” to the current marketing campaign. We evaluate the proposed model using a real dataset from The Globe and Mail which is a major newspaper in Canada. The experimental results show that the proposed model outperforms the traditional state-of-the-art approaches significantly.

MapReduce is an important programming model for processing in distributed environments. Compared to other distributed programming models, MapReduce reduces communication overheads between computers and improves fault tolerance. However, the MapReduce model does not allow for automatic synchronization between jobs. A large number of data analytics algorithms use a recursive divide-and-conquer approach, which inherently allows for parallelism at each level of recursion. However, it is often difficult to parallelize such algorithms using the traditional MapReduce model if the process requires synchronization. In this paper we introduce Parent-Child MapReduce, a version of the MapReduce programming model that allows for MapReduce tasks to be created dynamically and synchronized in a hierarchical parent-child fashion. Using the Parallel FP-Growth (PFP) algorithm for mining frequent patterns as a reference, we show that Parent-Child MapReduce can be used to parallelize recursive divide-and-conquer algorithms using the MapReduce model and that this can lead to significant speed ups in the computational speed of such algorithms. Our evaluation shows that we can achieve 68% (or 3 times) performance gain when used with PFP.

The problem of mining high utility sequential patterns (HUSP) has been studied recently. Existing solutions are mostly memory-based, which assume that data can fit into the main memory of a computer. However, with advent of big data, such an assumption does not hold any longer. Hence, existing algorithms are not applicable to the big data environments, where data are often distributed and too large to be dealt with by a single machine. In this paper, we propose a new framework for mining HUSPs in big data. A distributed and parallel algorithm called BigHUSP is proposed to discover HUSPs efficiently. At its heart, BigHUSP uses multiple MapReduce-like steps to process data in parallel. We also propose a number of pruning strategies to minimize search space in a distributed environment, and thus decrease computational and communication costs, while still maintaining correctness. Our experiments with real life and large synthetic datasets validate the effectiveness of BigHUSP for mining HUSPs from large sequence datasets.

Using approximate inference techniques, we investigate in this paper the applicability of Bayesian Networks to the problem of ranking a large set of documents. Topology of the network is a bipartite. Network parameters (conditional probability distributions) are determined through an adoption of the weighting scheme tf-idf. Rank of a document with respect to a given query is defined as the corresponding posterior probability, which is estimated through performing Rejection Sampling. Experimental results suggest that performance of the model is at least comparable to the baseline ones such as BM25. The framework of this model potentially offers new and novel ways in weighting documents. Integrating the model with other ranking algorithms, meanwhile, is expected to bring in performance improvement in document ranking.

Forced migration is increasingly becoming a global issue of concern. In this paper, we present an effective model of targeted event detection, as an essential step towards the forced migration detection problem. To date, most of the the approaches deal with the event detection in a general setting with the main objective of detecting the presence or onset of an event. However, we focus on analyzing the magnitude of a given event from a collection of text documents such as news articles from multiple sources. We use violence as an illustration as it is one of the most critical factors of forced migration. The recent advancements in semantic similarity measures are adopted to obtain relevant violence scores for each word in the vocabulary of news articles in an unsupervised manner. The resulting scores are then used to compute the average daily violence scores over a period of three months. Evaluation of the proposed model against a manually annotated data set yields a Pearson's correlation of 0.8. We also include a case study exploring the relationship between violence and key events.

Keyword search over relational databases offers an alternative way to SQL to query and explore databases that is effective for lay users who may not be well versed in SQL or the database schema. This becomes more pertinent for databases with large and complex schemas. An answer in this context is a join tree spanning tuples containing the query's keywords. As there are potentially many answers to the query, and the user is often only interested in seeing the top-k answers, how to rank the answers based on their relevance is of paramount importance. We focus on the relevance of join as the fundamental means to rank answers. We devise means to measure relevance of relations and foreign keys in the schema over the information content of the database. This can be done offline with no need for external models. We compare the proposed measures against a gold standard we derive from a real workload over TPC-E and evaluate the effectiveness of our methods. Finally, we test the performance of our measures against existing techniques to demonstrate a marked improvement, and perform a user study to establish naturalness of the ranking of the answers.

Keyword search in relational databases was introduced in the last decade to assist users who are not familiar with a query language, the schema of the database, or the content of the data. An answer is a join tree of tuples that contains the query keywords. When searching a database with a complex schema, there are potentially many answers to the query. Therefore, ranking answers based on their relevance is crucial in this context. Prior work has addressed relevance based on the size of the answer or the IR scores of the tuples. However, this is not sufficient when searching a complex schema. We demonstrate MeanKS, a new system for meaningful keyword search over relational databases. The system first captures the user's interest by determining the roles of the keywords. Then, it uses schema-based ranking to rank join trees that cover the keyword roles. This uses the relevance of relations and foreign-key relationships in the schema over the information content of the database. In the demonstration, attendees can execute queries against the TPC-E warehouse and compare the proposed measures against a gold standard derived from a real workload over TPC-E to test the effectiveness of our methods.

Keyword search over a graph searches for a subgraph that contains a set of query keywords. A problem with most existing keyword search methods is that they may produce duplicate answers that contain the same set of content nodes (i.e., nodes containing a query keyword) although these nodes may be connected differently in different answers. Thus, users may be presented with many similar answers with trivial differences. In addition, some of the nodes in an answer may contain query keywords that are all covered by other nodes in the answer. Removing these nodes does not change the coverage of the answer but can make the answer more compact. The answers in which each content node contains at least one unique query keyword are called minimal answers in this paper. We define the problem of finding duplication-free and minimal answers, and propose algorithms for finding such answers efficiently. Extensive performance studies using two large real data sets confirm the efficiency and effectiveness of the proposed methods.

Given an expert network, we tackle the problem of finding a team of experts that covers a set of required skills and also minimizes the communication cost as well as the personnel cost of the team. Since two costs need to be minimized, this is a bicriteria optimization problem. We show that the problem of minimizing these objectives is NP-hard. We use two approaches to solve this bicriteria optimization problem. In the first approach, we propose several (α, β)-approximation algorithms that receive a budget on one objective and minimizes the other objective within the budget with guaranteed performance bounds. In the second approach, an approximation algorithm is proposed to find a set of Pareto-optimal teams, in which each team is not dominated by other feasible teams in terms of the personnel and communication costs. The proposed approximation algorithms have provable performance bounds. Extensive experiments on real datasets demonstrate the effectiveness and scalability of the proposed algorithms.

A system for efficient keyword search in graphs is demonstrated. The system has two components, a search through only the nodes containing the input keywords for a set of nodes that are close to each other and together cover the input keywords and an exploration for finding how these nodes are related to each other. The system generates all or top-k answers in polynomial delay. Answers are presented to the user according to a ranking criterion so that the answers with nodes closer to each other are presented before the ones with nodes farther away from each other. In addition, the set of answers produced by our system is duplication free. The system uses two methods for presenting the final answer to the user. The presentation methods reveal relationships among the nodes in an answer through a tree or a multi-center graph. We will show that each method has its own advantages and disadvantages. The system is demonstrated using two challenging datasets, very large DBLP and highly cyclic Mondial. Challenges and difficulties in implementing an efficient keyword search system are also demonstrated.

Emotion detection from text is a relatively new classification task. This paper proposes a novel unsupervised context-based approach to detecting emotion from text at the sentence level. The proposed methodology does not depend on any existing manually crafted affect lexicons such as Word Net-Affect, thereby rendering our model flexible enough to classify sentences beyond Ekman's model of six basic emotions. Our method computes an emotion vector for each potential affect bearing word based on the semantic relatedness between words and various emotion concepts. The scores are then fine tuned using the syntactic dependencies within the sentence structure. Extensive evaluation on various data sets shows that our framework is a more generic and practical solution to the emotion classification problem and yields significantly more accurate results than recent unsupervised approaches.

Posting reviews online has become an increasingly popular way for people to express opinions and sentiments toward the products bought or services received. Analyzing the large volume of online reviews available would produce useful actionable knowledge that could be of economic values to vendors and other interested parties. In this paper, we conduct a case study in the movie domain, and tackle the problem of mining reviews for predicting product sales performance. Our analysis shows that both the sentiments expressed in the reviews and the quality of the reviews have a significant impact on the future sales performance of products in question. For the sentiment factor, we propose Sentiment PLSA (S-PLSA), in which a review is considered as a document generated by a number of hidden sentiment factors, in order to capture the complex nature of sentiments. Training an S-PLSA model enables us to obtain a succinct summary of the sentiment information embedded in the reviews. Based on S-PLSFA, we propose ARSA, an Autoregressive Sentiment-Aware model for sales prediction. We then seek to further improve the accuracy of prediction by considering the quality factor, with a focus on predicting the quality of a review in the absence of user-supplied indicators, and present ARSQA, an Autoregressive Sentiment and Quality Aware model, to utilize sentiments and quality for predicting product sales performance. Extensive experiments conducted on a large movie data set confirm the effectiveness of the proposed approach.

We study the problem of discovering a team of experts from a social network. Given a project whose completion requires a set of skills, our goal is to find a set of experts that together have all of the required skills and also have the minimal communication cost among them. We propose two communication cost functions designed for two types of communication structures. We show that the problem of finding the team of experts that minimizes one of the proposed cost functions is NP-hard. Thus, an approximation algorithm with an approximation ratio of two is designed. We introduce the problem of finding a team of experts with a leader. The leader is responsible for monitoring and coordinating the project, and thus a different communication cost function is used in this problem. To solve this problem, an exact polynomial algorithm is proposed. We show that the total number of teams may be exponential with respect to the number of required skills. Thus, two procedures that produce top-k teams of experts with or without a leader in polynomial delay are proposed. Extensive experiments on real datasets demonstrate the effectiveness and scalability of the proposed methods.

Keyword search over a graph finds a substructure of the graph containing all or some of the input keywords. Most of previous methods in this area find connected minimal trees that cover all the query keywords. Recently, it has been shown that finding subgraphs rather than trees can be more useful and informative for the users. However, the current tree or graph based methods may produce answers in which some content nodes (i.e., nodes that contain input keywords) are not very close to each other. In addition, when searching for answers, these methods may explore the whole graph rather than only the content nodes. This may lead to poor performance in execution time. To address the above problems, we propose the problem of finding r -cliques in graphs. An r -clique is a group of content nodes that cover all the input keywords and the distance between each two nodes is less than or equal to r . An exact algorithm is proposed that finds all r -cliques in the input graph. In addition, an approximation algorithm that produces r -cliques with 2-approximation in polynomial delay is proposed. Extensive performance studies using two large real data sets confirm the efficiency and accuracy of finding r -cliques in graphs.

The major challenge in mining data streams is the issue of concept drift, the tendency of the underlying data generation process to change over time. In this paper, we propose a general rule learning framework that can efficiently handle concept-drifting data streams and maintain a highly accurate classification model. The main idea is to focus on partial drifts by allowing individual rules to monitor the stream and detect if there is a drift in the regions they cover. A rule quality measure then decides whether the affected rules are inconsistent with the concept drift. The model is accordingly updated to only include rules that are consistent with the newly arrived concept. A dynamically maintained set of instances deemed relevant to the most recent concept is also kept at memory. Learning a new concept from a larger set of instances reduces the variance of data distribution and allows for a more accurate, stable classification model. Our experiments show that this approach not only handles the drift efficiently, but it also can provide higher classification accuracy compared to other competitive approaches on a variety of real and synthetic data sets.

their patients' records from paper to computer, enormous amounts of electronic medical records (EMR) have become available for medical research. Some of the EMR data are well-structured, for which traditional database management systems can provide effective retrieval and management functions. However, most of the EMR data (such as progress notes and consultation letters) are in free text formats. How to effectively and efficiently retrieve and discover useful information from the vast amount of such semi-structured data is a challenge faced by medical professionals. Without proper tools, the rich information and knowledge buried in the medical health records are unavailable for clinical research and decision-making. The objective of our research is to develop text analytics tools that are capable of parsing clinical medical data so that predefined search subjects that correspond to a list of medical diagnoses can be extracted. In addition to this particular core functionality, it is also desired that several important assets should be present within the text-analytics tools in order to improve its overall ability to be used as recommendation tools. In this research, we work with research scientists at the Institute for Clinical Evaluative Sciences (ICES) in Toronto and examine a number of techniques for structuring and processing free text documents in order to effectively and efficiently search and analyze vast amount of medical records. We implement several powerful medical text analytics tools for clinical data searching and classification. For data classification, our tools sort through a great amount of patientrecords to identify the likelihood of a patient having myocardial infarction (MI) or hypertension (HTN), and classify the patients accordingly. Our tools can also identify the likelihood of a patient being a smoker, previous smoker or non-smoker based on the text data of medical records.

Analyzing the large volume of online reviews would produce useful knowledge that could be of economic values to vendors and other interested parties. In particular, the sentiments expressed in the online reviews have been shown to be strongly correlated with the sales performance of products. In this paper, we present an adaptive sentiment analysis model called S-PLSA+, which aims to capture the hidden sentiment factors in the reviews with the capability to be incrementally updated as more data become available. We show how S-PLSA+ can be applied to sales performance prediction using an ARSA model developed in previous literature. A case study is conducted in the movie domain, and results from preliminary experiments confirm the effectiveness of the proposed model.

Online reviews, which are getting increasingly prevalent with the rapid growth of Web 2.0, have been shown to be second only to "word-of-mouth" in terms of influencing purchase decisions. It is therefore imperative to analyze them and distill useful knowledge that could be of economic values to vendors and other interested parties. Previous studies have confirmed that the sentiments expressed in the online reviews are strongly correlated with the sales performance of products. In particular, a model called ARSA has been proposed for predicting sales performance using a model called S-PLSA. In this paper, we build upon that work, and present an adaptive sentiment analysis model called S-PLSA+, which not only can capture the hidden sentiment factors in the reviews, but has the capability to be incrementally updated as more data become available. We show how the proposed S-PLSA+model can be applied to sales performance prediction using the ARSA model. A case study is conducted in the movie domain, and results from preliminary experiments confirm the effectiveness of the proposed model.

Writing and publishing reviews online has become an increasingly popular way for people to express opinions and sentiments. Analyzing the large volume of online reviews available can produce useful knowledge that are of interest to vendors and other parties. Prior studies in the literature have shown that online reviews have a significant correlation with the sales of products, and therefore mining the reviews could help predict the sales performance of relevant products. However, those studies fail to consider one important factor that may significantly affect the accuracy of the prediction, i.e., the quality of the reviews. In this paper, we propose a regression model that explicitly takes into account the quality factor, and discusses how this quality information can be predicted when it is not readily available. Experimental results on a movie review dataset confirm the effectiveness of the proposed model.

In this paper, we present a framework for mining diverging patterns, a new type of contrast patterns whose frequency changes significantly differently in two data sets, e.g., it changes from a relatively low to a relatively high value in one dataset, but from high to low in the other. In this framework, a measure called diverging ratio is defined and used to discover diverging patterns. We use a four-dimensional vector to represent a pattern, and define the pattern's diverging ratio based on the angular difference between its vectors in two datasets. An algorithm is proposed to mine diverging patterns from a pair of datasets, which makes use of a standard frequent pattern mining algorithm to compute vector components efficiently. We demonstrate the effectiveness of our approach on real-world datasets, showing that the method can reveal novel knowledge from large databases.

A transaction database usually consists of a set of time-stamped transactions. Mining frequent patterns in transaction databases has been studied extensively in data mining research. However, most of the existing frequent pattern mining algorithms (such as Apriori and FP-growth) do not consider the time stamps associated with the transactions. In this paper, we extend the existing frequent pattern mining framework to take into account the time stamp of each transaction and discover patterns whose frequency dramatically changes over time. We define a new type of patterns, called transitional patterns, to capture the dynamic behavior of frequent patterns in a transaction database. Transitional patterns include both positive and negative transitional patterns. Their frequencies increase/decrease dramatically at some time points of a transaction database. We introduce the concept of significant milestones for a transitional pattern, which are time points at which the frequency of the pattern changes most significantly. Moreover, we develop an algorithm to mine from a transaction database the set of transitional patterns along with their significant milestones. Our experimental studies on real-world databases illustrate that mining positive and negative transitional patterns is highly promising as a practical and useful approach for discovering novel and interesting knowledge from large databases.

Online reviews provide a valuable resource for potential customers to make purchase decisions. However, the sheer volume of available reviews as well as the large variations in the review quality present a big impediment to the effective use of the reviews, as the most helpful reviews may be buried in the large amount of low quality reviews. The goal of this paper is to develop models and algorithms for predicting the helpfulness of reviews, which provides the basis for discovering the most helpful reviews for given products. We first show that the helpfulness of a review depends on three important factors: the reviewerpsilas expertise, the writing style of the review, and the timeliness of the review. Based on the analysis of those factors, we present a nonlinear regression model for helpfulness prediction. Our empirical study on the IMDB movie reviews dataset demonstrates that the proposed approach is highly effective.

With the flourish of the Internet, online review mining has attracted a lot of attention from the research community. However, compared to various well-studied sentiment analysis and opinion summarization problems, less effort has been made to analyze the quality of online reviews. The objective of this paper is to fill in this gap by automatically evaluating the "helpfulness" of reviews and consequently developing novel models to identify the most helpful reviews for a particular product. In particular, based on a thorough analysis of various factors that may affect the review quality, we propose HelpMeter, a nonlinear regression model for helpfulness prediction. Some preliminary experiments were conducted on a movie review data set, and the performance results confirm the superiority of the proposed method.

Mining frequent patterns in transaction databases has been studied extensively in data mining research. However, most of the existing frequent pattern mining algorithms do not consider the time stamps associated with the transactions. In this paper, we extend the existing frequent pattern mining framework to take into account the time stamp of each transaction and discover patterns whose frequency dramatically changes over time. We define a new type of patterns, called transitional patterns, to capture the dynamic behavior of frequent patterns in a transaction database. Transitional patterns include both positive and negative transitional patterns. Their frequencies increase/decrease dramatically at some time points of a transaction database. We introduce the concept of significant milestones for a transitional pattern, which are time points at which the frequency of the pattern changes most significantly. Moreover, we develop an algorithm to mine from a transaction database the set of transitional patterns along with their significant milestones. Our experimental studies on real-world databases illustrate that mining positive and negative transitional patterns is highly promising as a practical and useful approach to discovering novel and interesting knowledge from large databases.

We present AV-space, a new data structure for caching data set statistics for efficiently learning classification rules from large data sets. The AV-space is designed to work with sequential-covering rule induction algorithms. It is used to accelerate queries about the count of the examples in a data set that satisfy a conjunction of attribute-value pairs. With an AV-space, the learning algorithm does not have to access the training data to obtain the statistics about the data. We present the structure of an AV-space, algorithms for building and querying an AV-space, and procedures for dynamically updating the AV-space during the rule induction process. We present an experimental evaluation that compares the AV-space with a commonly-used data structure that simply loads the (encoded) training examples into memory. We show that the use of AV-space significantly improves the speed of rule induction and that it consumes less memory on large data sets.

Due to its high popularity, Weblogs (or blogs in short) present a wealth of information that can be very helpful in assessing the general public's sentiments and opinions. In this paper, we study the problem of mining sentiment information from blogs and investigate ways to use such information for predicting product sales performance. Based on an analysis of the complex nature of sentiments, we propose Sentiment PLSA (S-PLSA), in which a blog entry is viewed as a document generated by a number of hidden sentiment factors. Training an S-PLSA model on the blog data enables us to obtain a succinct summary of the sentiment information embedded in the blogs. We then present ARSA, an autoregressive sentiment-aware model, to utilize the sentiment information captured by S-PLSA for predicting product sales performance. Extensive experiments were conducted on a movie data set. We compare ARSA with alternative models that do not take into account the sentiment information, as well as a model with a different feature selection method. Experiments confirm the effectiveness and superiority of the proposed approach.

Abstract With more and more information available on the Internet, the task of making personalized recommendations to assist the user's navigation has become increasingly important. Considering there might be millions of users with different backgrounds accessing a Web site everyday, it is infeasible to build a separate recommendation system for each user. To address this problem, clustering techniques can first be employed to discover user groups. Then, user navigation patterns for each group can be discovered, to allow the adaptation of a Web site to the interest of each individual group. In this paper, we propose to model user access sequences as stochastic processes, and a mixture of Markov models based approach is taken to cluster users and to capture the sequential relationships inherent in user access histories. Several important issues that arise in constructing the Markov models are also addressed. The first issue lies in the complexity of the mixture of Markov models. To improve the efficiency of building/maintaining the mixture of Markov models, we develop a lightweight adapt‐ive algorithm to update the model parameters without recomputing model parameters from scratch. The second issue concerns the proper selection of training data for building the mixture of Markov models. We investigate two different training data selection strategies and perform extensive experiments to compare their effectiveness on a real dataset that is generated by a Web‐based knowledge management system, Livelink.

In this paper, we investigate the use of data mining, in particular the text classification and co-training techniques, to identify more relevant passages based on a small set of labeled passages obtained from the blind feedback of a retrieval system. The data mining results are used to expand query terms and to re-estimate some of the parameters used in a probabilistic weighting function. We evaluate the data mining based feedback method on the TREC HARD data set. The results show that data mining can be successfully applied to improve the text retrieval performance. We report our experimental findings in detail.

We present an extensible java-based platform for contextual retrieval based on the probabilistic information retrieval model. Modules for dual indexes, relevance feedback with blind or machine learning approaches and query expansion with context are integrated into the Okapi system to deal with the contextual information. This platform allows easy extension to include other types of contextual information.

The problem of modeling and predicting a Web surfer's browsing patterns has gained increasing attention in recent years. In this paper we present our experience in clustering Web surfers using a mixture of Markov models with a real application of Livelink log data. We propose different techniques to improve the clustering performance, and evaluate the techniques through experiments.

Abstract We present a novel session identification method based on statistical language modeling. Unlike standard timeout methods, which use fixed time thresholds for session identification, we use an information theoretic approach that yields more robust results for identifying session boundaries. We evaluate our new approach by learning interesting association rules from the segmented session files. We then compare the performance of our approach to three standard session identification methods—the standard timeout method, the reference length method, and the maximal forward reference method—and find that our statistical language modeling approach generally yields superior results. However, as with every method, the performance of our technique varies with changing parameter settings. Therefore, we also analyze the influence of the two key factors in our language‐modeling–based approach: the choice of smoothing technique and the language model order. We find that all standard smoothing techniques, save one, perform well, and that performance is robust to language model order.

We propose two algorithms for grouping and summarizing association rules. The first algorithm recursively groups rules according to the structure of the rules and generates a tree of clusters as a result. The second algorithm groups the rules according to the semantic distance between the rules by making use of an automatically tagged semantic tree-structured network of items. We provide a case study in which the proposed algorithms are evaluated. The results show that our grouping methods are effective and produce good grouping results.

Livelink is a collaborative intranet, extranet and e-business application that enables employees and business partners of an organization to capture, share and reuse business information and knowledge. The usage of the Livelink software has been recorded by the Livelink Web server in its log files. We present an application of data mining techniques to the Livelink Web usage data. In particular, we focus on how to find interesting association rules and sequential patterns from the Livelink log files. A number of interestingness measures are used in our application to identify interesting rules and patterns. We present a comparison of these measures based on the feedback from domain experts. Some of the interestingness measures are found to be better than others.

We present our experience in mining web usage patterns from a large collection of Livelink log data. Livelink is a web-based product of Open Text, which provides automatic management and retrieval of different types of information objects over an intranet or extranet. We report our experience in preprocessing raw log data and post-processing the mining results for finding interesting rules. In particular we compare and evaluate a number of rule interestingness measures and find that two of the measures that have not been used in association rule learning work very well.

We compare a number of data mining and statistical methods on the drug design problem of modeling molecular structure-activity relationships. The relationships can be used to identify active compounds based on their chemical structures from a large inventory of chemical compounds. The data set of this application has a highly skewed class distribution, in which only 2% of the compounds are considered active. We apply a number of classification methods to this extremely imbalanced data set and propose to use different performance measures to evaluate these methods. We report our findings on the characteristics of the performance measures, the effect of using pruning techniques in this application and a comparison of local learning methods with global techniques. We also investigate whether reducing the imbalance in the training data by up-sampling or down-sampling would improve the predictive performance.

Researchers have embraced a variety of machine learning (ML) techniques in their efforts to improve the quality of learning programs. The recent evolution of hybrid architectures for machine learning systems has resulted in several approaches that combine rule induction methods with case-based reasoning techniques to engender performance improvements over more traditional single-representation architectures. We briefly survey several major rule-induction and case-based reasoning ML systems. We then examine some interesting hybrid combinations of these systems and explain their strengths and weaknesses as learning systems. We present a balanced approach to constructing a hybrid architecture, along with arguments in favor of this balance and mechanisms for achieving a proper balance. Finally, we present some initial empirical results from testing our ideas and draw some conclusions based on those results.