Explainability and Understanding in e-commerce: The Challenge of XAI
This thought-provoking blog post tries to tackle the challenge of AI / XAI (Explainable AI) or: how to create effective ML/LLM glass box systems and general challenges and the specific case for recommendations and relevancy in the world of e-commerce.
Introduction: Is Explainability a well-developed field in software?
The landscape of e-commerce (and software in general) is undergoing a significant transformation, largely attributed to the growing adoption of Machine Learning (ML) and new Data Model strategies like Large Language Models (LLM) and the widely recognised term Artificial Intelligence (AI).
From compilers to ML models, software evolves to create better abstractions, making really complex operations automatic and easy to reproduce. Explainability and understanding are endless topics that have already been solved and ignored in previous decades.
Given the example of a compiler as one of the main revolutions we had in our industry and having and its generic definition: “a program that converts instructions into a machine-code or lower-level form so that they can be read and executed by a computer,” we can think about the following questions:
- Is the output of a compiler understandable?
- Is the process of the compiler something we can explain?
- Who needs to understand these concepts and the details of this process?
- Is understandability/explainability a requirement of the system?
Most of the time, we end up in a situation where the response falls into one of three categories: yes, no, or it depends. To provide a more accurate response, it becomes essential to tackle the inherent bias within the problem first.
- What is the purpose/objective of the system?
- How mature is the system? What are the potential risks and impacts of the system?
- Who is using the system? Who gets value from the system?
By using these questions, we can enhance our conclusions about explainability to inquiries about any system’s use case.
The purpose of the system and the potential risk and impact are really similar topics with pretty common and connected information. In the first stages, a system should have a component of explainability since this will help with scalability and extension; and it could be a much better foundation to create abstractions on top of and to massively increase the integration.
The acceptance of a system depends on a great percentage of the trust the users have in the output, and making the process understandable is one of the keystones.
When the system is mature enough, we can assume that some user personas won’t need to understand the whys since the trust is already there. The system will always need power users with the capacity to understand the solution, which will offer credibility to be accepted by users who don’t have the knowledge or skills to understand the system.
The final stage of a system is pure acceptance based on the trust and the joy (easy to use and control, explainability) that is provided to the final user.
So, to lay the framework for proposing some key points for tackling explainability in the “new” AI systems, I ask you (reader):
- Can you explain how relevancy works in a Search Engine (i.e. Google/Yahoo)? Why do you get these results, and why do you get them in this order?
- Can you explain the process behind the ready-made lists that a music app prepares for you based on your preferences?
- Can you clarify the steps from using your credit card in your favourite e-commerce store to seeing a reduction in your bank account balance?
Moreover, can these already-accepted systems be AI or partly AI?
We trust, we enjoy, and thus, we create acceptance of a black box model created on top of an explainable (or unexplainable) system that experts and users trust and validate.
Then, how can this be applied to AI? What are the answers to these questions in the case of new LLM/ML/AI technologies?
AI Explainability in e-commerce tools: relevancy, recommendations, and utilities
From the explosion of vector DBs to productise the expensive trained models or usages in content creation and information enrichment, to more specific use cases like product recommendations or relevancy utilities, everyone is trying to use the power of ML and AI techniques to provide value in their e-commerce brands.
In any use case, based on our proposed framework, we need to ask:
- What is the purpose/objective of the system?
- How mature is the system? What are the potential risks and impacts of the system?
- Who is using the system? Who gets value from the system?
In the actual landscape, we have been fighting against black box data systems over the last few years. Many companies provide real value and impact for big retailers around the world, just by gathering and connecting data points without AI/ML or complex data model systems. We can assume that the impact is high and the risk WAS very low, but it is steadily increasing due to data privacy issues, regulations, and “the personal data discussion.”
The “old” systems were largely accepted by brands and consumers because the purpose was almost fair for everyone: we liked to be advised and helped with intelligent and personalised capabilities. This, in turn, made a huge impact on brands, from sales numbers and KPIs to the possibility of better coverage and more segmentation in their offerings.
While privacy and data policies were always a concern, no one asked for explainability of the how-to on the data side. But we started to need explainability and understanding.
Brands and merchandisers needed to understand the impact of the solutions, and when they got involved, they required fine-tuning and curation.
Consumers started to become aware of the value of the information, so they asked for more precise output and the possibility of “non-personalised” experiences and privacy controls.
Current status and foundations
We have already begun to go down this path, and we’re still discovering.
SaaS and software providers for retailers create fully decentralised controls in their solutions, and the capacity to explain what’s happening in each e-commerce has become harder and harder due to monetisation, ad platforms, and highly fine-tuned ranking algorithms.
Some of our colleagues have started to provide real foundations for explainability of systems, from OpenSourceConnection’s Quepid to the tons of open-source big data and analytics community tools in the field of ML that have emerged during the last decade.
But it’s not only on the “less profitable” side of software… The concept of control and trust can be found in monetisation and marketing platforms, and it’s becoming a really important field to consider in all types of software and business.
Lastly, closer to pure AI, initiatives from HuggingFace to start experimenting with the visibility of the data training sets are laying the groundwork for the next advancements in the field of explainability for the big players.
All e-commerce sub-systems, not only AI systems, are lacking in explainability; thus why using this context, with AI systems in mind, aiming for acceptance, integration, and usage of these complex systems while increasing the transparency and explainability is key.
Now, let’s get into the proposed actions and steps to follow to enhance explainability in e-commerce tools.
Enhancing Trust and Credibility
AI integrations need to win their space. Your customer metrics are diffused, and it might be difficult to attribute fluctuations to causes, but always being able to check and understand what’s happening is the best boost for trust.
By limiting the use cases and integrating explainable solutions into current tools, we can iteratively include these “strange algorithms” while maintaining the product experience.
Having a central source of truth for the data is the main foundation for integrating AI solutions with high credibility and acceptance.
It’s known that AI/ML models are difficult to explain (we’ll talk about that later), but insights and tools to explore and understand the connection between the data will help to create trust in the outputs of the models.
Fostering Collaboration Between Humans and AI
AI can be extremely beneficial. It can also be confusing. That’s why sometimes the feedback goes like:
“Overall, it’s great, but this brand/product is not paying to be promoted even if data said that customers loved it.”
“The metrics from last week look great, but we’re releasing a new catalogue next week, new keywords need to be added now!”
We get it. Humans are awesome, and our own processes are still far from being codified. Once the AI solutions are in place, things can happen, new rules and elements can be added, and others deleted.
Trends and topics move fast and we need control. Once basic (but useful) explainability is in place, we need curation and control, observing two tendencies here for AI:
- Use AI as a helper, use humans as the source: suggestions based on LLM trained for the specific domain of the customer.
- Use AI as the source, use humans as the decision makers: bring models to production by giving your customers the power to enable them. Allowing power users to enable the solution and curate the impact.
Mitigating Bias and Discrimination
AI is inherently biased, which is why explainability and control are necessary to reduce the impact of this problem.
We need solutions that are highly transparent and multivariate testing that can be executed out-of-the-box.
Providing testing tools, offline evaluations, clear explainability, connection to insights, and the ability to enable, control, and curate is the best way to mitigate, understand, and act on AI bias.
Adapting to Regulatory Requirements
The baseline for all the actions is to ask: how is data being collected?
My colleague Gerardo had some interesting thoughts about how messy all of this data collection has become over the past few years and how things are changing.
The topic of AI is generating a great revolution in the field of data privacy and personal information usage. With international organisations implementing new regulations about data collection and usage, we’re a long way from establishing solid, common sense rules to follow, but we can start tackling the issue by providing tools to analyse it.
There are great things coming up to help with Privacy Notice Controllers and other basic needs to control data collection.
Conversations about AI regulations and data ethics are constantly evolving, so remember to follow trusted sources for the latest news.
You can’t explain it, and you know it
One of the main arguments against these points is that in mainstream literature, Neural Networks and LLM are close to being black box systems by definition, so acceptance and integration might be harder if we lose the potential capacity of explainability.
There is a ton of literature on the processes behind ML and AI. One of my favourites is about AI functionality in trendy use cases (ChatGPT, generative AI).
We don’t need to understand; we need to believe.
Big data algorithms and new models are difficult to explain, the operations are complex, and the increasing complexity of retrieval implies reductions, simplifications, and approximations that make the initial effort of explaining useless.
Explainability is difficult to define. It depends on the use case, on the model, on the user persona. But we need to bring these solutions to fruition.
To create sustainable solutions based on clean data, you first have to respect your customers and their customers.
Let’s avoid skeletons in the closet.
References
All visuals were created with Midjourney, using custom prompts.
- i.e: [droid connected to device, person leaking privacy data, personal information danger]::20 + [very intricate, hd, tim burton]::10 + [vice to devices, enjoyment, fun, buying, expensive articles]::5 + [screens, shopping articles, complex electronic system, screens, mobile, smartphones]::3 — ar 4:2 — v 5.2 — s 1000
https://www.the-future-of-commerce.com/2019/03/11/what-is-explainable-ai-xai/
https://edrone.me/blog/explainable-ai-explained
http://wwwhomes.uni-bielefeld.de/gibbon/Handbooks/gibbon_handbook_1997/node389.html