Skip to Content
Product Information

How Covid-19 is changing everyone’s behaviour and how businesses can react

Part of the SAP AI Business Services introductory-, and product portfolio series (see all parts at the end of this post). Credit to the co-authors Steven Fu  and Leonard Dinu for drafting this blog post with me.

 

The current pandemic has accelerated the push towards digital transformation for many businesses as they deal with this unprecedented threat. While online shopping was already gaining traction prior to Covid-19, it is even more popular today. Figure 1 below shows that U.S. e-commerce has grown as much in the past 8 weeks, as it did in the previous 10 years. This trend is likely to be reflected globally and across several industries [1], as the pandemic accelerates the rate of customers and employees meeting, learning and purchasing virtually.

Retailers and solution providers will need to find more creative ways to showcase online inventory, display interesting contents and drive user engagement. In the world of experiential retail, it is imperative for merchants to create personalized, in-context offers and deliver an exceptional customer experience. Likewise, a shift in learning towards virtual classrooms will require online course providers to provide personalized learning recommendations.

Figure 1: US E-commerce Penetration (% of Retail Sales)

Source: Bank of America, U.S. Department of Commerce, ShawSpring Research 2020

 

How can SAP’s AI-based Recommendation Service help?

Recognizing the need for an intelligent recommendation service to better serve our customers, SAP has developed a novel Recommendation Service API based on a state-of-the-art neural network language model. The Machine Learning model draws insights on user patterns from historical user behaviour (e.g. clickstream) as well as available metadata (e.g. item catalogue, user profile). This enables the model to learn complex relations between consecutive item interactions, their respective attributes and the user features, resulting in highly personalized recommendations for each user. In a live environment, this means that multiple users interacting with the same item will receive different recommendations based on their unique history.

The below figure shows a sample input item sequence and the recommended outputs based on the similarity scores. In this example and all subsequent examples in this blogpost, we will use a public and well-known dataset (MovieLens 20M [2]) to demonstrate the capability of the Recommendation Service.

Figure 2: Personalized Recommendations generated from the input item sequence (based on the MovieLens dataset)

Let’s go through some of the key features built into the model to show how users can benefit from using our AI-based Recommendation Service.

Cold Start solution

A common issue with recommendation systems occurs when new users or new items are added to the catalogue, as they will not be recognized by the model during inference, making it difficult to get meaningful recommendations for them. In our approach, new items and new users are represented based on their attributes, which were learned by the model, ensuring that users receive personalized recommendations right from the first click. Figure 3 below illustrates how the top three videos are recommended for the cold start item “Toy Story 4″based on its attributes: title, categories and tags.

Figure 3: Cold Start recommendations (common attributes between input and recommended items are underlined)

Explainability

Explainability is a widely researched topic in the field of Artificial Intelligence. Customers increasingly expect a certain level of transparency in the predictions and are no longer content with the concept of the so-called “black box’. Teams specialized in AI increasingly need to share more information about the underlying AI models, as well as reasoning behind the results given by the algorithms in order to build trust in the automation.

Using machine learning techniques, the recommendation results can be explained by linking each recommended item and its attributes to the inference inputs. To provide the full picture of each recommendation, there are three levels of explainability available to the user.

To illustrate each type of explainability, let’s consider an inference call for a known user (with user profile and item interaction history) where an item (with catalogue entry) has been recommended. The explainability model can output:

  1. sequence attention – how much did each past item impact the current recommendation
  2. item attribute contribution – how much did each attribute of the recommended item contribute
  3. user attribute contribution – how much did each user feature matter when recommending a specific item

In addition to justifying each recommendation, we can also offer advanced insights into the entire dataset. Our “Feature Importance” functionality can use the trained recommendation model to determine an importance score to each item and user attribute (known from the item catalogue and user profile respectively). The importance is based on how much each attribute can impact the model evaluation metrics.

Figure 4: Feature Importance visualization showing the impact of each attribute (note: loss – model loss function, HR – hit rate, MRR – mean reciprocal rank)

With these insights, we can determine how much each item attribute drives the overall user behavior based on the user patterns learned by the model. Marketers can make use of this insight to customize or boost certain item attributes for example by enhancing certain brands or keywords, thereby providing complementary recommendations or promotions to increase sales and turnover. Figure 4 shows a sample calculation and visualization of each attribute for the MovieLens dataset, measured based on their individual impact on evaluation and business metrics.

Smart Search capability

Leveraging what we learned from the text attributes, the user can input free text or query and receive recommendations that are beyond a simple string match. The user could also take one or more attributes as an input (which would happen when a user selects one or more category filters) and recommend the closest items. We can also have combinations between one or more attributes and a free text query.

Scalability

The Recommendation Service is built on a scalable architecture that can handle and support increasing numbers of users or catalog items. The end-to-end process from data preparation, model training to deployment pipeline is simple to use and easily delivered in a short amount of time. The demo video below provides an overview of the whole process, and also presents how the API can be consumed in a website scenario.

 

Business Scenarios

Right from conceptualization SAP has developed the Recommendation Service as a reusable service that can be applicable to a wide range of business scenarios. Some examples are in industries such as e-commerce retail, B2B procurement and in human resources. Figure 5 shows some of the use cases that we have worked on.

Figure 5: Application of the Recommendation Service in multiple use cases

Scenario 1: E-commerce

Over the past decade several retailers have started and developed e-commerce and online shops as part of their overall channel strategy. The current pandemic has catalyzed and perhaps created a long-lasting shift to online retailing. With the integration of the Recommendation Service into SAP Commerce Cloud Context Driven Services (CDS), the solution will enable merchants to drive higher user engagement, boost item visibility and promote higher value products by customizing the item attributes. This would in turn help merchants drive key metrics such as cart revenue, customer retention and customer conversion.

Figure 6: Personalized Recommendations in e-commerce

Scenario 2: Procurement

Delivering personalized recommendations in the procurement process, with simple-to-use options for purchasing solutions online, can drive employee engagement and save time and effort. This would provide employees with the same “Amazon-like” user experience they would have in a normal B2C e-commerce platform. Employees can thus quickly make the best purchasing decision using their preferred vendors and suppliers. From our interactions with customers, we note that B2B buying typically has slightly more complex requirements (compared to B2C). Some common use cases are:

  • Recommendations for products that are commonly bought together (complementary recommendations): the recommendations act as a reminder to the user to purchase the related items that were bought together in the past.

 

  • Recommendations for a list of products to purchase based on past purchases (basket of products recommendations): this is an extension of the earlier use case on complementary item recommendation. In this use case the user has a list of items that he usually purchases and the list of recommendations he receives here can speed up his next set of purchases.

 

  • Recommendations for an alternative product should a product be unavailable (alternative recommendations): the recommendations can guide the user to purchase an alternative product should their initial option is not available.

 

Scenario 3: Human Resources – Learning Recommendations and Career Path Recommendations (a.k.a. People Like Me)

Over the years, we have seen how the Human Resources function in organizations has moved from a pure transactional role to a more strategic role focussing on employee engagement. From a traditional role in the past focussing on tasks such as creating standards for compensation and rewards, hiring and staffing in a legal and appropriate way, evaluating performance fairly etc., HR has evolved into a strategic function utilising AI and cloud technology to implement agile talent practices increasing employees’ productivity.  The rapid pace of change in today’s business world and the disruption caused by the Covid-19 pandemic has increased the need for HR leaders to adopt innovative ways to meet new business demands and new learning requirements.

We described two ways on how we incorporated AI-based recommendations in the SuccessFactors HR solution. Research has shown that employees struggle to stay on top of all the information available and select appropriate materials from the vast sea of choices. Two of the biggest challenges to developing new skills and knowledge are the overwhelming volume of information, and the lack of effective tools to find the most useful information [3]. In response to this, the team has built a Learning Recommendation service that outputs recommended learnings based on the profile, interests and browsing history of the user, in order to guide the user in his e-learning journey. An example of a Recommendation Tile can be found in example 7.

Figure 7: Recommendation Tile in the SuccessFactors Learning Management System (LMS)

 

Given the high demand for specialised skills in the market nowadays, coupled with the high cost of talent acquisition and attrition, it is imperative for organisations to build and retain talent, and to provide a framework for employees to identify a career path. Succession planning and career planning need to be well structured and part of a data driven process. Drawing on employee data, the “Career Path Recommender” solution provides career guidance for employees with an AI-informed view into what similar employees have set as goals, learned, taken on as roles, etc. Managers can use this recommendation service to identify employee competency and uncover hidden talent to enable strategic succession management.

 

Scenario 4: Dynamic Content Strategy and other use case

The Recommendation Service can be used to dynamically provide recommended webpages that the user will most likely access based on his profile and past browsing history. This helps the user to better navigate complicated websites and shorten the time needed to obtain the required information.

There are potentially many more use case that would be possible with an AI-based Recommendation Service. Readers are encouraged to reach out to the authors of this blog. You can also find a section below with further consumption options and information on how to engage with SAP.

Integration Architecture

The Recommendation Service runs on SAP Cloud Platform and is embedded in C4 Commerce and SuccessFactors Learning Management System (LMS), as well as Career Development Planning modules (see figure 8 below).

All functionalities are delivered via web services over the HTTPS protocol. The communication with the services is secured by the OAuth 2.0 protocol. The standard user authentication and authorization mechanisms provided by SAP Cloud Platform for Cloud Foundry is used. The service consumer can create an instance of the service and generate credentials to communicate with the service instance. For more information on this topic, see the documentation on Data Privacy and Security in the SAP Cloud Platform documentation.

As visible in the illustration below, the service consumer – which could be an SAP or non-SAP application – would call the service via the HTTPS-based API which is secured by the OAuth 2.0 protocol. The functionalities of the services (e.g. to classify a document or to extract the information contained in a PDF file) are available as RESTful APIs with respective endpoints and HTTP methods (especially GET, POST, DELETE). The data is provided back to the service consumer in the JSON format.

Figure 8: An example of a reference architecture for integrating the Recommendation Service

Consumption Options

There are two options for consuming the Recommendation Service.

Ready to use

Recommendation Service is natively embedded in SAP SuccessFactors Learning and will be rolled out in SAP Commerce Cloud and SAP SuccessFactors Succession & Development; the service can be consumed as part of the respective license in each solution.

Ready to test

Recommendation Service is also available for test and evaluation. The user can easily activate the service, test it on your own item catalogue and build a proof-of-concept around it. Do contact the authors of this blog for further information.

How to Engage with SAP

Influence SAP

Especially, the SAP Customer Engagement Initiative enables you to get early insights into SAP’s product developments and directly work with the developers to define and shape future product directions. Three times a year, a list of new projects is offered. In a one-month-registration phase, the user can register for a first informative virtual session and decide afterwards, if the user wishes to participate in the project and influence the developments.

Co-Innovate with SAP

You can work together with us on enhancing the functional scope of the offerings such as developing new recommendation features or to support new business scenarios. You can join beta testing programs of new features or engage with other SAP products in testing the integrated service within these products. Do reach out to the authors of this blog.

Personalization is the future

As many parts of the world continue to battle with Covid-19, the pace of digital transformation will accelerate, and the adoption of automation technology and AI will increase. In many industries the shift to the online world is gaining more traction. Personalization and personalized recommendations are no longer an option but a necessity in today’s business world, where customers expect a simplified customer journey with recommendations tailored to their needs.

Authors

Product Manager: Lai Kwok Peng

Product Owner: Steven Fu

Machine Learning Developer: Dinu, Leonard

 

More Information:

If you have more questions feel free to reach out to us or ask your questions over SAP Community using the hashtag #Artificial Intelligence.

Read all blog posts of the SAP AI Business Services introductory–, and product portfolio series:

 

1 Comment
You must be Logged on to comment or reply to a post.