For centuries, people have expressed individuality and character by their choices of clothing, furniture, and accessories as well as preferences in items such as books, songs, and movies. The advances of the Internet and mobile platforms have made new discoveries and creations easily shared and curated. Given the substantial value of such personalization of products and services, this blog will explore how AI can help with recommendations that offer value to both consumers and enterprises.
The recent developments in AI, specifically deep learning (DL) and the various topologies of deep neural networks (DNNs), are particularly suitable and more accurate in extracting reasonable representation of patterns in data. Such representation by DNNs takes place without orchestrating a number of algorithms and features of importance to create classification or prediction models. It is the nature of DNNs, and most commonly convolutional and recurrent neural networks (CNNs and RNNs), that the properties inherent in the data can be uncovered without specific formulae. Thus image and speech recognition have seen human-level performance with variants of CNNs and RNNs.
Unlike the steady time sequences of speech and video with RNN and CNN classifications, recommender systems are fundamentally different and more challenging because user activities happen at irregular time intervals that are difficult to align. Despite these challenges, at the Intel AI Lab, we research areas of commercial value where AI and DL have significant impact and can be benchmarked against previous statistical approaches of classical machine learning (ML) techniques. Therefore, personalization is a key subject of interest for exploration.
The current state-of-the-art recommendations by technology companies like Facebook, Amazon, Netflix, and Spotify have deployed types of DL models, though no official benchmarks are available to compare with ML. For instance, Amazon only published a DL library that can be used to create a combination of DNNs with matrix factorization. Because of the value and complexity of AI for personalization, we wanted to test different algorithms and share our findings.
As a baseline we used an adaption of a known model Wide and Deep (W&D) DNN published by Google. The W&D model has two streams: one is the linear part that investigates features about the data and the other uses a deep model to process the data. In our case (Fig.1), the customers and the products were embedded using a standard embedding layer. For additional model variants, other embeddings could be used if available, such as word descriptions or images of the products.
We collaborated with a retail business Lolli and Pops and their transactional time-series data of approximately 250,000 historical purchases by 50,000 customers for 3,000 products with brief descriptions. Our goal was to develop a recommendation engine that offers personalized product suggestions to shoppers by benchmarking different algorithms and identifying the revenue maximizing option for the retailer.
The result was that a DL model based on this data had 15% higher accuracy than a classical ML approach of collaborative filtering in a top K (K=3) benchmark. To illustrate, if 45% of customers are purchasing one of the top 3 suggested products with the ML system, in the same data test with the DL model, the number is 60% of customers. The significance of this finding is that customers often get much better recommendations with DL, which can lead to higher revenues for an enterprise.
Future updates to this research could apply some active learning techniques from reinforcement learning (RL), especially if a continuous stream of online data is available, and we anticipate that new methods will achieve even better results with higher accuracy and revenue impact.
Similar approaches with DL and RL would benefit not only retail or e-commerce businesses but also companies with product and service recommendations in media, finance, healthcare, education, and other industries. With the latest controversies about using consumer data and applying various AI techniques in social media, it is important to emphasize that responsible AI solutions incorporate appropriate consumer knowledge and participation, within applicable laws and regulations. In that context, DL and RL can and will provide great benefits. AI for personalization is one such area where individuals are shown to enjoy helpful and relevant assistance in their discoveries, while enterprises can derive greater commercial value.
This blog shows a specific example where DL approach delivers enhanced value proposition compared to classical ML. We expect that other businesses and their consumers will also benefit from similar AI algorithms to help with new product discovery and recommendations. Nevertheless, even with the increasing role and impact of AI for personalization, it will not replace the significance of the human individuality in shaping of tastes and preferences, given our unending pursuit of learning and exploration.