Fast Online 'Next Best Offers' using Deep Learning

Abstract

In this paper we present iPrescribe, a scalable low-latency architecture for recommending 'next-best-offers' in an online setting. The paper presents the design of iPrescribe and compares its performance for implementations using different real-time streaming technology stacks. iPrescribe uses ensemble of deep learning and machine learning algorithms for prediction. We describe the scalable real-time streaming technology stack and optimised machinelearning implementations to achieve a 90th percentile recommendation latency of 38 milliseconds. Optimizations include a novel mechanism to deploy recurrent Long Short Term Memory (LSTM) deep learning networks efficiently.

Experience with PCIe streaming on FPGA for high throughput ML inferencing

“…Authors in [4] discuss the methodology of recommending products to the retail end user customers, introducing a way of how past transactional data is processed, features engineered and used to train machine learning models which could be used to serve real time recommendations. Paper [12] describes an end to end implementation of such a system by leveraging the Recobell and the Kaggle Instacart data sets. They present how use of in memory technologies help build a scalable system on CPU based server clusters, with XGBoost (with the same configuration as evaluated in the paper) as one of the inferencing algorithms.…”

“…In order to prepare the training data, we use the iPrescribe framework [12]. This framework joins all the data provided by the Recobell [6] data set and converts it into a common data format.…”

“…This model is trained with the PAKDD 2017 data set [6]. The accuracy of this model is reported in [12]. While our implementations are verified functionally to give the same level of accuracy, any further discussion on model accuracy is out of scope of the paper.…”

Customer Purchase Behavior Prediction in E-commerce: A Conceptual Framework and Research Agenda

“…x x CDM Handle of data unbalancing [28] x Feature engineering with recency and frequency of page views [29] x x CF Combination of features from clickstream and transactions for collaborative filtering [30] x Feature engineering for clickstream [31] x Combination of features from clickstream and transactions for collaborative filtering [32] x Feature engineering with product heterogeneity for collaborative filtering [33] x x x DLC Real-time predictions with ensemble and deep learning [34] x Recommendation of bundles of products considering quality and diversity criteria [35] Purchase Intent (PCI)…”

“…─ Sequential Learning: Few proposals have explored sequential ML models in this literature. Examples are recurrent neural networks, which are only adopted in three studies [25,33,40]. Such models are indicated to learn the evolving consumer behavior over time, and sequential patterns such as "She is buying a phone case after purchasing a smartphone".…”