GLAD: Graph-based long-term attentive dynamic memory for sequential recommendation

Recommender systems play a crucial role in the e-commerce stores, enabling customers to explore products and facilitating the discovery of relevant items. Typical recommender systems are built using n most recent user interactions, where value of n is chosen based on trade-off between incremental gains in performance and compute/memory costs associated with processing long sequences. State-of-the-art recommendation models like Transformers, based on attention mechanism, have quadratic computation complexity with respect to sequence length, thus limiting the length of past customer interactions to be considered for recommendations. Even with the availability of compute resources, it is crucial to design an algorithm that strikes delicate balance between long term and short term information in identifying relevant products for personalised recommendation. Towards this, we propose a novel extension of Memory Networks, a neural network architecture that harnesses external memory to encapsulate information present in lengthy sequential data. The use of memory networks in recommendation use-cases remains limited in practice owing to their high memory cost, large compute requirements and relatively large inference latency, which makes them prohibitively expensive for online stores with millions of users and products. To ad- dress these limitations, we propose a novel transformer-based sequential recommendation model GLAD, with external graph-based memory that dynamically scales user memory by adjusting the memory size according to the user’s history, while facilitating the flow of information between users with similar interactions. We establish the efficacy of the proposed model by benchmarking on multiple public datasets as well as an industry dataset against state-of-the-art sequential recommendation baselines.