Over the past two decades, consumers have become accustomed to receiving personalized recommendations in all facets of their online life. Whether that be recommended products while shopping on Amazon, a curated list of apps in the Google Play store, or relevant videos to watch next on YouTube. In fact, in a Verge article “​​How YouTube perfected the feed: Google Brain gave YouTube new life,” the Google Brain team reveals how their recommendation engine has impacted the platform with “more than 70 percent of the time people spend watching videos on the site being driven by YouTube’s algorithmic recommendations” thereby increasing time spent on the platform by 20X in three years. 

It’s become clear that personalized recommendations are no longer a differentiator for an organization but rather something consumers have come to expect in their day-to-day experiences online. So what should you do if you are behind the curve and want to get started or simply want to improve upon what you already have? While there are all sorts of techniques, from content-based systems to deep learning methods, our goal in this recommender-focused blog series is to demystify three available approaches to building recommendation systems on Google Cloud: Matrix Factorization in BigQuery Machine Learning (BQML),  Recommendations AI, and deep retrieval techniques available via the Two-Tower built-in algorithm.

One of these approaches can be used to meet you where you are in your personalization journey, no matter if you are just starting or if you are well into it. This first blog post will introduce our three approaches and when to use them. 

What is Matrix Factorization and how does it work?

Collaborative filtering is a foundational model for building a recommendation system as the input dataset is simple and the embeddings are learned for you. How does Matrix factorization fit into the mix you might be wondering?  Matrix factorization is simply the model that applies collaborative filtering. BQML enables users to create and execute a matrix factorization model by using standard SQL directly in the data warehouse. 

Collaborative filtering begins by creating an interaction matrix. The interaction matrix represents users as a row and items as columns in your dataset. This interaction matrix often is sparse in nature as not all users will have interacted with many items in your catalog. This is where embeddings come into play. Generating embeddings for users and items not only allows you to collapse many sparse features into a lower dimensional space but they also allow you to derive a similarity measure so that similar users/items fall nearby in the embedding space. These similarity measures are key as collaborative filtering uses similarities between users and items to make the end recommendations. The underlying assumption being that similar users will like similar items whether that be movies or handbags. 

What’s required to get started?

To train a matrix factorization model you need a table that includes three input columns: user(s), item(s), and an implicit or explicit feedback variable (e.g., ratings is an example of explicit feedback). With the base input dataset in place, you can then easily run your model in BigQuery after specifying several hyperparameters in your CREATE MODEL SQL statement. Hyperparameters are available to specify the number of embeddings, the feedback type, the amount of L2 regularization applied and so on.    

Why use this approach and who is it a good fit for?  

As mentioned earlier, Matrix Factorization in BQML is a great way for those new to recommendation systems to get started. Matrix factorization has many benefits: 

• Little ML Expertise: Leveraging SQL to build the model lowers the level of ML expertise needed
• Few Input Features: Data inputs are straightforward, requiring a simple interaction matrix
• Additional Insight: Collaborative filtering is adept at discovering new interests or products for users 

While Matrix Factorization is a great tool for deriving recommendations it does come with additional considerations and potential drawbacks depending upon the use case. 

• Not Amenable to Large Feature Sets: The input table can only contain two feature columns (e.g., user(s), item(s)). If there is a need to include additional features such as contextual signals, Matrix factorization may not be the right method for you.  
• New Items: If an item is not available in the training data, the system can’t create an embedding for it and will have difficulty recommending similar items. While there are some workarounds available to address this cold-start issue, if your item catalog often includes new items, Matrix factorization may not be a good fit.  
• Input Data Limitations: While the input matrix is expected to be sparse, training examples without feedback can cause problems. Filtering for items and users that have at least a handful of feedback (e.g., ratings) examples can improve the model. More information on limitations can be found here. 

In summary, for users with a simplified dataset looking to iterate quickly and develop a baseline recommendation system, Matrix Factorization is a great approach to begin your personalization AI journey. 

What is Recommendations AI and how does it work?

Recommendations AI is a fully managed service which helps organizations deploy scalable recommendation systems that use state-of-the-art deep learning techniques, including cutting-edge architectures such as two-tower encoders, to serve personalized and contextually relevant recommendations throughout the customer journey.

Deep learning models are able to improve the context and relevance of recommendations in part because they can easily address the previously mentioned limitations of Matrix Factorization. They incorporate a wide set of user and item features, and by definition they emphasize learning successive layers of increasingly meaningful representations from these features. This flexibility and expressivity allows them to capture complex relationships like short-lived fashion trends and niche user behaviors. However, this increased relevance comes at a cost, as deep learning recommenders can be difficult to train and expensive to serve at scale. 

Recommendations AI helps organizations take advantage of serving these deep learning models and handles the MLOps required to serve these models globally with low latency. Models are automatically retrained daily and tuned quarterly to capture changes in customer behavior, product assortment, pricing, and promotions. Newly trained models follow a resilient CI/CD routine which validates they are fit to serve and promotes them to production without service interruption. The models achieve low serving latency by using a scalable approximate nearest neighbors (ANN) service for efficient item retrieval at inference time. And, to maintain consistency between online and offline tasks, a scalable feature store is used, preventing common production challenges such as data leakage and training-serving skew.  

What’s required to get started?

To get started with Recommendations AI we first need to ingest product and user data into the API:

• Import product catalog: For large product catalog updates, ingest catalog items in bulk using the catalogItems.import method. Frequent catalog updates can be schedule with Google Merchant Center or BigQuery
• Record user events: User events track actions such as clicking on a product, adding items to cart, or even purchasing an item. These events need to be ingested in real time to reflect the latest user behavior and then joined to items imported in the product catalog 
• Import historical user events: The models need sufficient training data before they can provide accurate predictions. The recommended user event data requirements are different across model types (learn more here)

Once the data requirements are met, we are able to create one or multiple models to serve recommendations: 

• Determine your recommendation types and placements:  The location of the recommendation panel and the objective for that panel impact model training and tuning. Review the available recommendations types, optimization objectives, and other model tuning options to determine the best options for your business objectives.
• Create model(s): Initial model training and tuning can take 2-5 days depending on the number of user events and size of the product catalog 
• Create serving configurations and preview recommendations: After the model is activated, create serving configurations and preview the recommendations to ensure your setup is functioning as expected before serving to production traffic

Once models are ready to serve, consider setting up A/B experiments to understand how newly trained models impact your customer experience before serving them to 100% of your traffic. In the Recommendations AI console, see the Monitoring & Analytics  page for summary and placement-specific metrics (e.g., recommender-engaged revenue, click-through-rate, conversion rate, and more).

Why use this approach and who is it a good fit for?  

Recommendations AI is a great way to engage customers and grow your online presence through personalization. It’s used by teams who lack technical experience with production recommendation systems, as well as customers who have this technical depth but want to allocate their team’s effort towards other priorities and challenges. No matter your team’s technical experience or bandwidth, you can expect several benefits with Recommendations AI: 

• Fully managed service: no need to preprocess data, train or hypertune machine learning models, load balance or manually provision you infrastructure – this is all taken care of for you. The recommendation API also provides a user-friendly console to monitor performance over time. 
• State-of-the-art AI: take advantage of the same modeling techniques used to serve recommendations across Google Ads, Google Search, and YouTube. These models excel in scenarios with long-tail products and cold-starts users and items
• Deliver at any touchpoint: serve high-quality recommendations to both first-time users and loyal customers anywhere in their journey via web, mobile, email, and more
• Deliver globally: serve recommendations in any language anywhere in the world at low-latency with a fully automated global serving infrastructure
• Your data, your models: Your data and models are yours. They’ll never be used for any other Google product nor shown to any other Google customer

For users looking to leverage state of the art AI to fuel their recommendation systems but need an existing solution to get up and running more quickly, Recommendations AI is the right solution for you. 

What are Two Tower encoders and how do they work?

As a reminder, in recommendation system design, our objective is to surface the most relevant set of items for a given user or set of users. The items are usually referred to as the candidate(s) where we might include information about the items such as the title or description of the item, other metadata about the item like language, number of views, or even clicks on the item over time. User(s) are often represented in the form of a query to a recommendation system where we might provide details about the user such as the location of the user, preferred languages, and what they have searched for in the past.   

Let’s start with a common example. Imagine that you are creating a movie recommendation system. The input candidates for such a system would be thousands of movies and the query set can consist of millions of viewers. The goal of the retrieval stage is to select a smaller subset of movies(candidates) for each user and then score and rank order them before presenting the final recommended list to the query/user.

The retrieval stage is able to refine our list of candidates by encoding both the candidate and the query data so they share the same embedding space. A good embedding space will place candidates which are similar to one another closer together and dissimilar items/queries farther apart in the embedding space.

Once we have a database of query and candidate embeddings we can then use an approximate nearest neighbor search method to then generate a list of final “like” candidates, i.e. find a certain number of nearest neighbors for a given query/user and surface final recommendations.

What’s required to get started?

At the most basic level, in order to train a two-tower model you need the following inputs:

• Training Data: Training data is created by combining your query/user data with data about the candidates/items. The data must include matched pairs, cases where both user and item information is available. Data in the training set can include many formats from text, numeric data, or even images. 
• Input Schema: The input schema describes the schema of the combined training data along with any specific feature configurations.

Several services within Vertex AI have come available that complement the existing Two-Tower built-in algorithm and can be leveraged in your execution: 

• Nearest Neighbor (ANN) Service: Vertex AI Matching Engine and  ScANN provide a high-scale and low-latency Approximate Nearest Neighbor (ANN) service so you can more easily identify similar embeddings.
• Hyperparameter Tuning Service: A hyperparameter tuning service such as Vizier can help you identify the optimal hyperparameters such as the number of hidden layers, the size of the hidden layers, and the learning rate in fewer trials. 
• Hardware Accelerators: Specialized hardware, such as GPUs or TPUs, can be valuable in your recommendation system to help accelerate experiments and improve the speed of training cycles. 

Why use this approach and who is it a good fit for?  

The Two-Tower built-in algorithm can be considered the “custom sports car” of recommendation systems and comes with several benefits: 

• Greater Control: While Recommendations AI uses the two-tower architecture as one of the available architectures it doesn’t provide granular control or visibility into model training, example generation, and model validation details. In comparison, the Two-Tower built in algorithm provides a more customizable approach as you are training a model directly in a notebook environment. 

• More Feature Options: The Two Tower approach can handle additional contextual signals ranging from text to images. 
• Cold Start Cases: Leveraging a rich set of features not only enhances performance but also allows the candidate generation to work for new users or new candidates.

While the Two-Tower built in algorithm is an excellent and best-in class solution for deriving   recommendations, it does come with additional considerations and potential drawbacks depending upon the use case.

• Technical ML Expertise Required: Two tower encoders are not a “plug and play” solution like the other approaches mentioned above. In order to effectively leverage this approach, appropriate coding and ML expertise is required. 
• Speed to Insight: Building out a custom solution via two-tower encoders may require additional time as the solution is not pre-built for the user.   

For users looking for greater control, increased flexibility, and have the technical chops to easily work within a managed notebook environment – the two-tower built in algorithm is the right solution for them. 

What’s next?

In this article, we explored three common methods for building recommendation systems on Google Cloud Platform. As you can see thus far, there are alot of considerations to take into account before choosing a final approach. In an effort to help you align more quickly we have distilled the decision criteria down to a few simple steps (see below for more details).

In the next installments of this series, we will dive more deeply into each method, explore how hardware accelerators can play a key role in recommendation system design, and discuss how recommendation systems may be leveraged in key verticals. Stay tuned for future posts in our recommendation systems series. Thank you for reading! Have a question or want to chat? Find authors here – R.E. [Twitter | LinkedIn], Jordan [LinkedIn], and Vaibhav [LinkedIn].

Acknowledgements

Special thanks to Pallav Mehta, Henry Tappen,Abhinav Khushraj, and Nicholas Edelman for helping to review this post. 

References

• Announcing ScaNN: Efficient Vector Similarity Search
• Factorization Machines for Item Recommendation with Implicit Feedback Data

With the novel coronavirus COVID-19 sweeping across continents and fatalities climbing every day, it was only a matter of time before countries closed their borders to contain its spread.

In the wake of this decision, travel and tourism, the linchpins of many economies, were among the worst affected.

According to the United Nations World Tourism Organization (UNWTO), the COVID-19 pandemic caused a 22 percent fall in international tourist arrivals during the first quarter of 2020, and could see an annual decline of between 60 percent and 80 percent when compared with 2019.

The impact on the economy and to livelihoods that are dependent on tourism and hospitality has been significant. Prior to the pandemic, the outlook was quite different. Research by UNWTO in 2018 estimated that India would have 50 million outbound tourists by 2020.

Technology was also set to play a huge part in that growth. A Google Travel study showed that 74 percent of travellers were planning their trips on the Internet, and technologies like AI, IoT and VR were all set to be key trends this year.

Now, as borders slowly reopen and travel restrictions are gradually lifted, technology could once again be the game-changer. Manoj Sharma CTO, Cleartrip.com spoke to YourStory about the industry’s road to recovery and how technology will aid that journey.

Read the Full Story on YourStory

Today we’re hosting our Retail & Consumer Goods Summit, a digital event dedicated to helping leading retailers and brands digitally transform their business. For me, this is a personally exciting moment, as I see tremendous opportunities for those companies that choose to focus on their customers and leverage technology to elevate experiences.

Our event includes breakout sessions to help retailers and brands become customer centric, embrace the digital moment and transform their operations. Some of my favorite sessions include: 

• Why Search Abandonment is the metric that matters, highlighting Retail Search, and featuring a conversation with Macy’s
• Driving Consumer Closeness in a Privacy-Centric World, discussing how retailers and brands can create a successful first party data strategy, featuring a conversation with P&G
• The Modern Store: 7 Innovation Hotspots, sharing how retailers and brands can approach store transformation to unlock the most value from technology, featuring a conversation with The Home Depot.   

I’ll be speaking in our Retail Spotlight session, discussing the current retail landscape and our industry approach, followed by conversations with Albert Bertilsson, Head of Engineering – Edge at IKEA Retail (Indga Group) and Neelima Sharma, Senior Vice President, Technology Ecommerce, Marketing and Merchandising at Lowe’s. 

Let me share a bit more about the topics we’ll discuss in that session.

In retail specifically, digital-first shopping journeys are blurring the lines between the physical and digital brand experience. Shoppers want to know what’s available before they visit your stores, and they expect fulfillment options like curbside pickup. We see this when tracking trends for interest in curbside pickup or in-stock items.

This has left many retailers asking how they can get smarter with their data, tackle the $300 billion dollar problem of “search abandonment,” move faster to create new customer experiences, and do a better job of connecting their employees and customers – with confidence.

Our team has been spending time thinking about how we can rise and succeed in this new era together. We continue to focus on areas where we can bring the best of our capabilities to our retail customers around the world. And we’re focused on ways we can bring the best of what Google has to offer through cloud integrations.

Our goal is to help retailers become customer-centric and data-driven, capture digital and omni-channel revenue growth, create the modern store and drive operational improvement.

Let’s dig into each of these strategic pillars in a bit more detail. 

Become customer centric and data driven

Customers today expect experiences that are timely, targeted, and tailored for them and their needs, and reject experiences that can’t deliver these features. Data modeling, legacy technology, and siloed systems often prevent retailers from providing that level of personalized experience. 

At Google Cloud, we work with global retailers and our ecosystem partners to activate and bring value from first-party data, particularly in the field of customer data platforms (CDPs). This includes integrations from Google Cloud, such as our business intelligence platform Looker and other popular platforms to power one source of customer data through the organization. We also help retailers modernize their data warehouse with Looker for gathering business intelligence across their organization. This is important not just for consumer data, but inventory, supply chain, and store operations as well. 

Capture Digital and Omnichannel Growth 

We power some of the largest e-commerce sites in the world, helping them scale for Black Friday, Cyber Monday, and other holiday events. While scale is critically important, it’s also important to consider the quality of the online experience. How do your customers find products? How can you help deliver seamless online and omnichannel experiences? 

To help, we’re building product discovery solutions that bring together the best of our technologies that help retailers drive engagement with their consumers. Retail Search, for example, gives retailers the ability to provide Google-quality search on their own digital properties – search that is customizable for their unique business needs and built upon Google’s advanced understanding of user intent & context. 

The imperative is clear. Recent research found that retailers lose more than $300 billion to search abandonment — when purchase intent is not converted into a sale due to bad search results — every year in the US alone. 

Today, we announced that Retail Search is available to a larger set of retailers. If you are interested in learning more about Retail Search you can contact your sales representative for additional details.

Create the modern store  

With the rise of buying trends like curbside pickup and proximity-based search, our Google Maps Platform team is working on new products and features to help raise inventory awareness for your shoppers. We want to help you make it easier for them to understand what’s available to purchase in their channel of choice.

With Product Locator, each product page connects customers with information they need for local pickup and delivery options. This ensures customers are aware of pickup and delivery options throughout the buying journey—not just checkout. 

Awareness of local inventory can boost a wide range of key metrics for your business. Shopify recently shared that shoppers who opt for local pickup over delivery had a +13% higher conversion rate and that 45% of local pickup customers make an additional purchase upon arrival.

This is just one quick example of how our Google Maps Platform team can improve experiences for your shoppers.

Operational improvement

It can be challenging to operate in a world and at a time when consumer behavior and supply chains are so disrupted and volatile, and where entire retail teams had to go remote during the pandemic and beyond. 

We’re working with retailers to leverage artificial intelligence (AI) to improve consumer experience through chat bots or conversational commerce that solves problems for customers from anywhere. You can learn more about these offerings in our Conversational Commerce with Google breakout session, featuring Albertsons.

As the need for digital transformation continues to accelerate, Google Cloud is helping retailers stay ahead of the curve with solutions for digital and omnichannel growth, data-driven and customer-focused experiences, and operational improvement. For every era of cloud technologies, from the past into the future, Google Cloud is committed to providing solutions to retailers.

Read more about our solutions for retail, and check out additional sessions, including the CPG Industry Spotlight Session How To Grow Brands in Times of Rapid Change – Featuring L’Oréal at our Retail & Consumer Goods Summit.

You are what you eat. We’ve all been told this, but the truth is what we eat is often more complex than we are – genetically at least. Take a grain of rice. The plant that produces rice has 40,000 to 50,000 genes, double that of humans, yet we know far more about the composition of the human genome than of plant life. We need to close this knowledge gap quickly if we are to answer the urgent challenge of feeding 8 billion people, especially as food security around the globe is likely to worsen with climate change.

For this reason, AI company InstaDeep has teamed up with Google Cloud to train a large AI model with more than 20 billion parameters on a dataset of reference genomes for cereal crops and edible vegetables, using the latest generation of Google’s Tensor Processing Units (Cloud TPU v4), which is particularly suited for training efficiency at scale. Our aim is to improve food security and sustainable agriculture by creating a tool that can analyze and predict plants’ agronomic traits from genomic sequences. This will help identify which genes make some crops more nutritious, more efficient to grow, and more resilient and resistant to pests, disease and drought.

Genomic language models for sustainable agriculture

Ever since farming began, we have been, directly or indirectly, trying to breed better crops with higher yields, better resilience and, if we’re lucky, better taste too. For thousands of years, this was done by trial and error, growing crops year-on-year while trying to identify and retain only the most beneficial traits as they naturally arise from evolutionary mutations. Now that we have access to the genomic sequences of plants, we hope to directly identify beneficial genes and predict the effect of novel mutations.

However, the complexity of plant genomes often makes it difficult to identify which variants are beneficial. Revolutionary advances in machine learning (ML) can help to understand the link between DNA sequences and molecular phenotypes. This means we now have precise and cost-effective prediction methods to help us close the gap between genetic information and observable traits. These predictions can help identify functional variants and accelerate our understanding of which genes link to which traits – so we can make better crop selections.

Moreover, thanks to the vast library of available crop genetic sequences, training large models on hundreds of plant genomes means we can transfer the knowledge from thoroughly-studied species to those that are less understood but important for food production – especially in developing countries. And by doing this digitally, AI can quickly map and annotate the genomes of both common and rare crop variants.

One of the major limitations of traditional ML methods for plant genomics has been they mostly rely on supervised learning techniques. They need labeled data. Such data is scarce and expensive to collect, severely limiting these methods. Recent advances in natural language processing (NLP), such as Transformer architectures and BERT-style training (Bidirectional Encoder Representations from Transformers), allow scientists to train massive language models on raw text data to learn meaningful representations. This unsupervised learning technique changes the game. Once learned, the representations can be leveraged to solve complex regression or classification tasks – even when there is a lack of labeled data.

InstaDeep partners with Google Cloud to train the new generation of AI models for genomics on TPUs

Researchers have demonstrated that large language models can be especially effective in proteomics. To understand how this works, imagine reading amino acids as words and proteins as sentences. The treasure trove of raw genomics data – in sequence form – inspired InstaDeep and Google Cloud to apply similar technologies on nucleotides, this time reading them as words and chunks of genomes as sentences.

Moreover, the representations that the system learned improved in line with the size of the models and datasets, NLP research studies showed. This finding led InstaDeep researchers to train a set of increasingly larger language models on genomics datasets ranging from 1 billion to 20 billion parameters.

• Models of 1 billion and 5 billion parameters were trained on a dataset comprising the reference genomes for several edible plants, including fruit, cereal and vegetables for a total of 75 billion nucleotides.
• The training dataset must increase in the same proportion as the model capacity, recent work has shown. Thus, we created a larger dataset gathering all reference genomes available on the National Center for Biotechnology Information (NCBI) database including human, animal, non-edible plant and bacteria genomes. This dataset, which we used to train a 20-billion-parameter Transformer model, comprised 700 billion tokens, exceeding the size of most datasets typically used for NLP applications, such as the Common Crawl or Wikipedia dataset.
• Both teams announced that the 1 billion-parameter model will be shared with the scientific community to further accelerate plant genomics research.

The compact and meaningful representations of nucleotide sequences learned by these models can be used to tackle molecular phenotype prediction problems. To showcase their ability, we trained a model to predict the gene function and gene ontology (i.e. a gene’s attribute) for different edible plant species.

Early results have demonstrated that this model can predict these characteristics with high accuracy – encouraging us to look deeper at what these models can tell us. Based on these results, we decided to annotate the genomes of three plant species with considerable importance for many developing countries: cassava, sweet potato, and yam. We are working on making these annotations freely available to the scientific community and hope that these will be used to further guide and accelerate new genomic research.

Overcoming scaling challenges with massive models and datasets with Cloud TPUs

The compute requirement for training our 20 billion-parameter model with billions of tokens is massive. While modern accelerators offer impressive peak performance per chip, to utilize this performance often requires tightly coupled hardware and software optimizations. Moreover, maintaining this efficiency when scaling to hundreds of chips presents additional system design challenges. The Cloud TPU’s tightly-coupled hardware and software stack is especially well suited to such challenges. The Cloud TPU software stack is based on the XLA Compiler which offers out-of-the-box optimizations (such as compute and communication overlap) and an easy programming model for expressing parallelism.

We successfully trained our large models for genomics by leveraging Google Tensor Processing Units (TPUv4). Our code is implemented with the JAX framework. JAX provides a functional programming-based approach to express computations as functions that can be easily parallelized using JAX APIs powered by XLA. This helped us to scale from a single host (four chips) configuration to a multi-host configuration without having to tackle any of the system design challenges. The TPU’s cost-effective inter- and intra-communication capabilities led to an almost linear scaling between the number of chips and training time. This allowed us to train the models quickly and efficiently on a grid of 1024 TPUv4 cores (512 chips).

Conclusion

Ultimately, our hope is that the functional characterization of genomic variants predicted by deep learning models will be critical to the next era in agriculture, which will largely depend on genome editing and analysis. We envisage that novel approaches, such as in-silico mutagenesis – the assessment of all possible changes in a genomic region by a computer model – will be invaluable in prioritizing mutations that improve plant fitness and guiding crop improvements. Attempting similar work in wet-lab experiments would be difficult to scale and nearly impossible in nature. By making our current and future annotations available to the research community, we also hope to help democratize breeding technologies so that they can benefit all of global agriculture.

Further Reading

To learn more about the unique features of Cloud TPU v4 hardware and software stack we encourage readers to explore Cloud TPU v4 announcement. To learn more about scaling characteristics, please see this benchmark and finally we recommend reading PJIT Introduction to get started with JAX and SPMD parallelism on Cloud TPU.

This research was made possible thanks to the support of Google’s TPU Research Cloud (TRC) Program which enabled us to use the Cloud TPUv4 chips that were critical to this work.