TECHNOLOGY OVERVIEW
Recommendation systems are an effective customer experience (CX) tool. Not only do they help to provide a personalised experience to end-users by recommending a curated list of products and services, they are also useful for businesses to comprehend customers' interests. Overall, this helps to drive customers towards the correct purchasing decision, enhances the retail experience which translates to increased sales and customer retention. However, popular techniques in recommendation systems are highly susceptible to errors from over-generalisation.
This technology offer addresses issues of existing recommendation systems associated with generalisation and the lack of exploitation of knowledge graph interlinks to infer customer preferences, in two ways:
- A new optimisation framework to enhance the robustness of popular Matrix Factorisation (MF) recommendation models towards perturbations (noise) added to model parameters for improved generalisation performance through adversarial training
- Leveraging sequential dependencies to allow effective reasoning on knowledge graph paths to infer the underlying rationale beneath a user-item interaction, leading to improved explainability
TECHNOLOGY FEATURES & SPECIFICATIONS
This technology consists of two improvements over existing recommender systems; an Adversarial Personalised Ranking (APR) framework that produces a recommender model which is robust to adversarial noise in its parameters and a knowledge-aware graph network that is able to endow the recommender model with recommendation explainability.
Both demonstrate significant improvements over existing state-of-the-art solutions.
Adversarial Personalised Ranking (APR) and Adversarial Matrix Factorisation (AMF):
An optimisation framework that enhances the robustness of existing recommender models - suitable for personalised ranking while ensuring robustness to adversarial noise:
- Similar to Bayesian Personalised Ranking (BPR), a dominant pairwise learning method
- General learning framework that is model-independent
- Instantiated as Adversarial Matrix Factorisation (AMF) - utilising a basic, yet effective recommendation model, Matrix Factorisation (MF), is trained with BPR, and further optimised with the APR framework
- Evaluated by adding perturbations (noise) to the parameters of a recommender model
- Recommendation performance does not change drastically, indicating robust and no vulnerability to adversarial noise in its parameters
Knowledge-aware Path Recurrent Network (KPRN):
- Exploits knowledge graphs to construct paths as extra user-item connectivity to complement existing user-time interactions information
- End-to-end neural network model learns knowledge graph path semantics for improved personalised recommendations
- Explicit reasoning/explanations behind recommendations e.g. "Castle on the Hill is recommended since you have listened to Shape of You sung and written by Ed Sheeran"
- Long Short-Term Memory (LSTM) is used to evaluate the sequential dependency of paths to infer user preferences
POTENTIAL APPLICATIONS
This technology enables robust personalised recommendations with viable explainability and reasoning in the following areas:
- Entertainment/Gaming
- Healthcare
- Financial Technology
- Education
- Food & Beverage
- Online commerce
- Mobile applications
Unique Value Proposition
- Explainabiility and reasoning: provides explicit reasoning to reveal the reasons behind a recommendation
- Integration with recommendation: end-to-end neural network to learn knowledge graph path interlinks to complement recommendations
- User preference understanding: enhanced customer experience and retention
- Robust to noise: recommendations are unaffected by disturbances in model parameters
- Outperforms pairwise ranking, such as Bayesian Personalised Ranking (BPR) with an 11.2% average improvement
- Outperforms state-of-the-art neural recommender models e.g. Neural Collaborative Filtering (NCF)