What is Pattern X?
Infinite Revamp of Algorithms
“There is only one boss. The customer. And he can fire everybody in the company from the chairman on down, simply by spending his money somewhere else.” — Sam Walton.
Sam was always right in understanding the customer from the early days of Walmart and so that is how it has become today in delivering the needs of customers at a low price every day. Amid various businesses striving towards understanding the consumer, there is a rise in technology-enabled ways to do this. One way is Machine Learning, a subset of Artificial Intelligence.
Walmart stores are one of the great places that many people love to visit, not only for shopping the lowest prices every day but also for the interaction with our store associates. We can find many inspiring and comforting stories that customers share daily on how associates are helping them by making their shopping easier. Traditionally, our customers’ shopping inspiration, discovery, and transactions primarily happened across our stores. Our store associates interact with customers on a regular basis (patterns) to understand their favorites (personalization), receive their feedback (ratings), and many other factors that have helped us create a great experience for them.
Being a human-led and tech empowered innovator, Walmart Global Tech is enabling our associates to go above and beyond to positively impact the lives of customers by leading the next retail disruption. While our store associates continue to interact with customers daily, our technologists and engineers are building tech-enabled products that can understand, personalize, redefine, and help customers shop anytime and anywhere with a tailored “wow” experience.
Top 3 Areas of ML Enablement in Retail
- Inventory & Stocking
2. Dynamic Pricing
3. Understanding & Predicting Customer Behavior
From these aspects, understanding customer behavior is the most significant segment for any kind of business and more specifically to retailers. Providing hyper personalization with unique customer experience is possible given that the algorithm is ingested with users’ feed of accurate data from all kinds such as images, selections, clicks, purchase history etc. The new oil is in many forms, and it must be cleaned and restructured to identify patterns, understand the user behavior and their preferences.
The Grey Area of Prediction
Let’s say John is searching for a blue shirt:
Scenario A: Based on his previous patterns in searches or orders, the action of algorithm determines all sorts of navy-blue colored shirts and suggests a few of them instantly. In this scenario there are three possibilities.
A1: If John is really looking for a navy-blue color, then the ML-enabled pattern is positive, and improves with the action of the consumer to provide more unique personalization next time.
A2: On the other hand, if he is looking for aqua blue, it fails to suggest them and considers it as a negative. The following action from the consumer can search again with exact words and the action from the algorithm can be to understand the error difference and analyze by taking the result as input to enhance/improve.
A3: One of the suggestions can buy John’s attention which is clearly changing his thought process, and he buys it.
Scenario B: Based on his previous patterns in searches/orders, John also purchased aqua blue colored shirts, so the algorithm scans for the sets of combination and recommends the results.
B1: If there is available stock and John finds a good aqua blue colored shirt from the recommendations, then the prediction keeps improving.
B2: If there are no aqua blue colors in the scan and it’s able to fetch only navy blue and sky-blue colors, then there are two possibilities again.
→ B2.1: John changes his search words to mention exact words as per his needs and finds out that the required unit does not exist which is negative because it’s out of stock.
→ B2.2: One of the recommendations pulled his attention, eventually that overrides John’s original thought/need and he buys a navy-blue shirt.
In addition to the above-mentioned scenarios, there can be n number of possibilities that could have happened. The different perspective here is the grey area I am trying to represent at the scenarios of A3 and B2.2. The customer wanted to buy an aqua blue colored shirt but because the suggestions that came up are navy blue, and one of them pulled his attention he proceeded to place the order and buy it.
Two things happened here from two different views: change of customer’s opinion/thought at that point of time, and the other view is the suggestions override the customer’s opinion/thought/wish. In simple words, it clearly altered the behavior of the customer instead of understanding and predicting. Of course, there are people who can search again with specific words but some of them will deviate from their initial thought. Ever happened to you? This is the most common behavior you might have experienced while surfing on different platforms like music, movies, e-commerce, etc.
Popular Menace to ML
A biased algorithm is a nightmare of the ML and AI, and it keeps haunting every pattern and algo (algorithm). Avoiding bias is the huge challenge for the developers since it is untraceable as either it can be from the designers/group of people who are involved, or it comes from the data that is ingested for training set. The data is again from humans who are inevitable to bias. History shows us how big a threat it is to the data, and here are a few examples of danger bells.
→ A popular organization’s auto recruiting process as an AI project used data from the last 10 years to train their AI model. This data from history contained biases against women since there was a male dominance across the tech industry. Therefore, the recruiting system incorrectly learned that male candidates were preferable, and the organization later stopped using that algorithm.
→ In 2016, there was the first beauty contest judged by AI which showcased some controversial results. Out of 44 winners, the majority were white, with a few who had light-brown skin, and only one winner with darker skin. This occurred because the data set used while training the model included mostly white people.
Some experts believe that technically it is possible that AI will be completely unbiased. However, in the real world, it is not anticipated that AI will be completely unbiased in near future. There are numerous human biases and ongoing identification of new biases is rapidly growing. Therefore, it is unfeasible to have a 100 percent unbiased human mind, therefore the AI and the ML algo are as well. The goal of human-centric tech is that the predictions/recommendations evolving from different algorithms and patterns should never override the consumers’ thoughts/decisions.
Finding the x with Feedback Loops
We have a lot of variables/factors to be derived in understanding the consumers’ behavior but one of the greatest assets is the x. It’s erroneous to think that once an ML model is trained and put into action, it always needs good supervision to continuously update it accordingly to changes in the world. Pattern x is all about the quest to understand the customer right and predicting the choice of thoughts/decisions even before they engage with the screens.
Given the moment, every door of ML algo leads to the feedback loops at the end of their pipeline in the system which are the only way to understand and compare with the prediction of the algorithm so that it can always improve next time. It is important to think and choose carefully about:
- Where feedback can be captured?
- How to incorporate it back into your model?
These two decisions create a prominent role when bringing the human to make the algo better by using a human-centered design. When a user searches for blue shirt, after the results are displayed, it’s always a better practice to throw a popup for collecting their opinion if your algo did good or should be improved. Now this feedback component can carry different variables to collect a precise opinion from user. Often, the process of designing the feedback loop and capturing feedback effectively is the hard work that separates successful ML projects from failed ones.
The 3 Faced Feedback Loops
1. Customers provide feedback when the model requests
2. Customers skip the feedback
3. Collect feedback on a purpose within the boundary
1. Customers provide feedback as when the model requests:
Customer’s View: This is the sunny day of the flow where customers give appropriate feedback in accordance with their search/experience throughout the product usage. This set of people expect to get better results next time they do the action and so contribute their valuable feedback.
Algorithm’s View: The model gets what it has expected and by considering the feedback, it will improve the prediction next time to understand and provide better results.
2. Customers skip the feedback:
Customer’s View: Some people would not understand the value of feedback completion but still they are correct as it’s not their duty to give feedback every time they utilize the product in their hands, so users might skip this at times.
Algorithm’s View: These are tricky to handle and of course a significant part of the faces, the quality of decision engine should be maintained the same and need to find other forms like a periodic feedback scheduler, changing the experience of requesting feedback, quick surveys to understand why they skip feedback etc., or maintaining a benchmark score so that the model should always maintain it for the quality of prediction.
3. Collect feedback on a purpose within the boundary:
With the increase of large social media platforms, a set of people started boycotting these feedback loops as these are improvising to become smarter and understand more about the users than was originally required. For example, once a customer ordered a product for one of their friends/family, but he/she never intends to buy such items in the future. Since purchase history is one of the factors in understanding the customers, the algo will start recommending related products. I hope you are seeing one more blind spot where the model should enable an option of customizing the feedback loops within itself and assure the customer that their online behavior is not being monitored. The need of the algorithm is to continue finding only the pattern x but not all variables because if the consumer gets any hint of the thirst from the model, the pattern x becomes more robust to find and solve.
In the world of digitalizing every possible asset, the ever-growing data is the key to opening unimaginable doors by creating opportunities to identify, understand, and redefine the markets. As a data enthusiast, I strongly believe in continuous revamping from the smallest line of code to a complex model built on a purpose. Data is the core of the ML with numerous forms, yet it is not unleashed with full power.
Mankind has evolved in solving many challenges around technology — we build machines, we code algorithms, we design the patterns, and we are always capable of finding the x in each problem to make the impossible, the possible.
References:
https://research.aimultiple.com/ai-bias/
https://www.impira.com/blog/feedback-loop
https://mitsloan.mit.edu/ideas-made-to-matter/machine-learning-explained
https://www.dataversity.net/making-machine-learning-datasets-unbiased/
