Recommender System, an active domain of information provision, focuses on modelling user-centric information, e.g. access-logs, purchase history, rating records, and products or services reviews to predict user potentially interested items/products, preferred services or actions. The suggestions provided are aimed at improving user experience and loyalty, facilitating decision-making for users and creating more revenues for online businesses and merchants and so on. Recently applications of recommender systems go far beyond transitional prediction of movie ratings and online products, reaching out to areas of user/friend, merchant, relation and service recommendations. The input data used in recommendations varies from demographic, textual, and interactional towards multi-modalities and contextual data. The used methods transition from conventional statistical and similarity-based calculations to more learning-based algorithms, even deep learning enhanced approaches. A lot of new recommendation tasks and requirement emerge due to the diverse focuses, such as trust, explainability, fairness, group, diversity in recommendations.
DSMI has researched on related topics and published intensive papers in these areas:
- cold-start and long-tail problems
- social network-based
- trust based
- sequential based
- knowledge graph representation based
- geolocation based (POI)
- group and diversity-based
- explainable recommender systems
- music recommendation
Causality-based Explainable Machine Learning
Our research advances the state-of-the-art on causal analysis theory and its interplay with interpretable machine learning (specifically deep learning, transfer learning and supervised learning). Our research pursues the goal of enabling machine learning methods with causality and human-level intelligence. Although current interpretable models have greatly improved the interpretability landscape, but they are unable to provide causal explanations for human-level intelligence. The causality is a clear and mathematically sound mechanism for users to truly understand the core of machine learning. How to empower the interpretable models with the advanced causal explanations is largely unexploited. Our research lies in the fertile interplay between machine learning and causal analysis, with a broad array of topics including:
- Causal inference and causal discovery
- Causal reasoning for machine learning
- Causality in explainable recommendation
- Visual causality analysis
- Fairness in machine learning
By infusing our new advances in these topics, we enable our decision-makers to achieve more on customer targeting, predictive analytics and personalized pricing. We also publish our research findings at top-tier venues in computer science and beyond.
Knowledge Graph and Applications
The multi-relationship of Knowledge Graph (KG) brings new challenges for Knowledge Graph analysis, it also makes the research on KG more attractive, because, with this kind of automatically extracted structured human knowledge, we have an opportunity to reveal the human knowledge reasoning patterns with analysis methodologies. As a result of KG analysis, KG can be used as a semantic enhancement for downstream application scenarios, such as Recommender System (RS). In our research, both KG completion and KG based downstream applications are studied. Although a huge amount of human knowledge facts have been collected from multiple open resources, existed KG is still incomplete. Part of our current research focuses on following proper embedding models for KG completion, 1) entity & relation embedding model, 2) conceptual taxonomy integrated embedding model, and 3) multi-relational graph sub-structure embedding.
Besides that, we also look into Hierarchical Collaborative Embedding for KG and RS. The core technique used in the system is based on collaborative network embedding learning, a type of methods learns embedding representation for nodes and links in a heterogeneous network consisting of both multi-relational network and network in other structures.
There are following new topics available for future research in KG area.
- Complex multi-relational KG network structure analysis. Some complex structural information can be extracted from KG for several tasks, such as KG self-completion, knowledge reasoning through network propagation, etc.
- Domain-specific KG downstream applications. Currently the applications of KG are mainly in Question Answering and Recommender System categories. With the semantic enhancement ability, the KG can be applied into more scenarios including social media short-text analysis, news event detection, etc.
- Joint learning on KG and textual knowledge data. Not all the human knowledge pieces can be represented in triple structure as in current KG, in real-world scenarios, there are large amount of knowledge contained in complex textual description. An effective method to link knowledge graph with textual knowledge and achieve knowledge reasoning jointly is also a promising direction.
- Scientific Knowledge Graph construction. Scientific knowledge is contained scientific papers, it is important to find a way to let machine understand the scientific knowledge from papers and represent the scientific knowledge in a proper network structure. Once the scientific knowledge graph constructed, the graph can be used for several applications including scientific question answering, scientific information retrieval, and decision support system based on scientific knowledge.
Behaviour Modelling and Social Computing
Behaviour modelling and social computing are two fundamental research directions at DSMI. Behaviour modelling is to derive user behavioural patterns, preferences, profiles from user behaviours and user generated textual data, while social computing is to analyse the social characteristics and trends demonstrated from intra-human or human-computer interactions. We continuously publish research papers on these two topics and apply the developed techniques to help our industry partners with their business problems. Examples including anti-selection detection (Zurich One-Path), spam behaviour detection (Toutiao), and product/action recommendations (Yozo,https://www.yozo.com.au/). Specifically, our strengths include
- Identifying misbehaviours or abnormal behaviours from a population
- Discovering the underlying behaviour patterns in a population
- Recommending actions/items based on the underlying patterns
- Techniques: topic modelling, representation learning and embedding, rule mining, sequential pattern mining, anomaly detection, time series analysis.
In a project with Zurich One-Path, we analysed 10-years of insurance application forms to detect misbehaviours (anti-selection) in insurance applications. Over the years, anti-selection may happen when the applicants provide incorrect information or misleading information such that their risks appear lower, e.g., they can get covered for what they should not. In this project, we developed a system based on rule mining algorithms to detect such anti-selection behaviours. In addition to the identified misbehaviours, we also enabled transparency, e.g., all predictions come with explanations, which is vital for operationalization.
Text Mining and NLP
Our core research capability advanced the state-of-the-art in devising new algorithms in analysing meaning, topics, patterns from user generated contents, such as user comments and reviews, product descriptions, social media posts, and customer spoken language e.g. call-logs for actionable business insights. DSMI has had significant experience in develop customising and using open source methods to using free text analysis to identify risk factors in mental health and customer behaviour patterns in the R&D area, as well as applying to enterprise business operations for identifying high-risks to enable a preventative management approach (e.g. Onepath and Colonial First State). Capability areas include:
- K-gram and bag-of-word
- Term indexing, retrieval and search and ranking algorithms
- Sentiment analysis and figurative detection
- Topic modelling and summarisation
- Word and context embedding algorithms
- Latent semantic analysis and knowledge graphs
Prof Guandong Xu and Shawou Liu analysed millions of call logs to identify customer behaviour patterns (including extracting latent representations of customers from their communications). This information was used to assess the customers sentiment, as well estimate future behaviours such as churn. Where as in One-Path, they used text mining analysis of free text questions, to identify the main trigger for deterioration in mental health and risk profiles. Through incorporating free text analysis with customer cohort information, it also enabled One-Path to identify key resilience factors required for improving mental health outcomes.
DSMI is experienced in using advanced data science techniques for cohort, client and patient segmentation analysis to identify key triggers and customer cohorts that are high-risk. Application areas include insurance underwriting (Zurich One-Path), triaging for NSW emergency departments to improve emergency department outcomes (NSW Health), as well as many other data science applications for government agencies (e.g. Australian Taxation Office). Capability areas include;
Division or density-based Clustering algorithms
Advanced data visualization for non-domain experts to easily interpret
Customized algorithms to accommodate user-specified optimization objectives
In a project with NSW Health (2018), DSMI team analysed 8 million presentations to the emergency departments to understand the triggers to develop assistive triaging tools to ensure better patient and emergency department outcomes. This included reducing customer wait times and readmissions – including identifying high-risk patients to reduce mortality rates. The project included developing customised tree-based segmentation algorithms cluster patients into different groups based on NSW Emergency Department Policies and Guidelines to develop more assistive and targeted patient treatment plans for different patient cohorts in time critical conditions to reduce mortality rates, improve use of hospital resources and outcomes, in addition to reducing re-admissions. Outcomes of the project included NSW Health changing their policies and adopting different performance metrics for different hospitals based on emergency department patient cohorts.
DSMI has had significant experience in designing customised predictive analytical tools within brown filed business operating frameworks to develop automated assistive AI tools to augment performance of staff (including improving operational resource efficiency, quality assurance and fraud identification), as well as through developing more target and improved personalised customer experiences. Companies include Australian Taxation Office, One-Path, NSW Health, Colonial First State, Providence Asset Group.
- Domain-specific feature engineering
- Structured and unstructured data integration and linkage; multi-sourced data fusion, e.g. demographic, transactional, behavioural, socio-economic data
- Standard supervised learning algorithms, e.g., gradient boosting, neural networks, ensembles
- Time-series forecasting algorithms
- Multi-label supervised learning
- What-if analysis, prescriptive analytics
- Bayesian probabilistic methods
Predictive modelling is involved in almost all the projects we did. For example,
- we predicted the churn rate of customers for Colonial First State;
- we predicted the exclusion risk of life insurance applications and an automated AI underwriting risk engine for standard applications for OnePath;
- we predicted the optimal treatment time of patients for NSW Health;
- we predicted the electricity generation and dynamic price for solar farms managed by Providence Asset Group.
One-Path had a problem its underwriting process were manual and not suitable for on-line business, with customers often having to wait for 1 for manual underwriting assessment to determine in applications were declined, accepted or had a loading and exclusion applied. UTS used customised predictive analytics, within One-Path rigorous operational processes to identify triggers for high risks customers from application, underwriting through to claims, and then developed an automated AI-assistive underwriting tool to assess all the standard applications in real-time, and reduce application time, so that underwriters could focus on more complex non-standard cases.