Recommendation Systems at Trunk Club: Overview (Part 1)
by Anshul Agarwal
This post, structured in two parts, describes how various kinds of machine learning driven recommendation systems developed by the Data Science team at Trunk Club enhance personalization for the customers as well as add value for the business. In Part 1, we describe a customer’s typical journey with Trunk Club and their interaction with recommendation systems along that journey. In Part 2, we dive a little deeper into the technical details of the algorithms that drive these systems. We also outline challenges and future outlook for recommendation systems and personalization at Trunk Club.
Business of “Trunks” with technology
As customers are getting inundated with choice in today’s economy, personalized services are increasingly becoming mainstay, especially in retail and e-commerce. Trunk Club has been providing personalized shopping experience to its customers since as early as 2009. It was the first one to offer personal stylists working closely with customers online to help choose wardrobe pieces for their lifestyle. Here, customers do not simply buy clothes but they buy “styles” carefully and deftly curated by their personal stylists. Customers receive coherent blends of outfits in boxes, called “Trunks”; they can keep what they love and return what they don’t. Stylists help customers find the clothes that make them feel great about the way they look, while making the process effortless and fun. With Trunk Club, customers get Nordstrom’s world-class selection of top-quality brands to choose from. Moreover, they experience the iconic Nordstrom standard of care and customer service in the world.
Being Customer Obsessed is one of Trunk Club’s core values. For Trunk Club, getting to know their customers is at the heart of everything. In order to achieve that, Trunk Club relies on one of its core pillars — Technology with human ingenuity. At Trunk Club, data science is a core discipline that is used to augment stylist expertise with recommendations generated from the state-of-the-art models and tools. Moreover, Trunk Club’s homegrown technology stack and platform provides vital foundation for the outstanding customer service it strives for. While maintaining the existing infrastructure, the data science and technology team is constantly innovating using new ideas in order to continuously improve customer satisfaction.
Throughout various touch points during their journey with Trunk Club, customers, directly or indirectly, interact with various recommendation system models. Driven by a combination of multiple advanced machine learning algorithms that learn patterns from the product data, customer attributes, and their past purchase history, these systems help stylists and customers sift through a plethora of choices and find clothing items that not only satisfy their style needs but also help them discover something new. The algorithms that power these recommendation systems are a home-grown blend of AI techniques that optimize for objectives tailored (pun intended) towards business needs.
Customer journey with Trunk Club
The figure above demonstrates a typical customer journey with Trunk Club. It also lists different data science models leveraged across different stages.
Onboarding
As customers sign up for the service, they go through the Trunk Club Onboarding experience in which they are asked several questions around their style and lifestyle preferences. These questions and customers’ responses provide vital data for the recommendation algorithms. At the completion of onboarding they are paired with a personal stylist.
Trunk Request
Once paired with their stylist, customers can start interacting with them and make a Trunk Request. The trunk request can range from a high level ask for general categories (such as shoes, jeans, etc.) to a highly granular and specific style request. A customer can also begin a blank Trunk Request, thus relying on the stylist’s knowledge of their style preferences to surprise them. The Trunk Request (either blank or not) provides vital information to stylists (and data science models) before shopping.
Stylist Shopping
The next step in the process is curating a digital trunk. The stylists, based on customers’ onboarding, personal communication with them, and the knowledge of their past keeps, returns, and feedback (except for new customers), prepare a digital assortment of 10 clothing items. This digital trunk is then shared with the customers for their preview through Trunk Club app and webpage. To prepare the digital trunk, stylists utilize Trunk Club’s internal catalog that lists all merchandise and inventory available at the time of shopping. The catalog, to imagine, is similar to any e-commerce website, such as Nordstrom.com. In the current setup customers do not have direct access to the catalog, only stylists have it. To enable stylists be more efficient and find most relevant personalized items, data science has developed various kinds of recommenders for the catalog, which we describe in the next section.
Member Preview
Trunk Club provides an opportunity for customers to review and provide feedback on the digital trunk assembled by their stylist in Member Preview. Trunk Club is one of the very few among its competitors that provide this option to preview the trunk (and replace disliked items) before it’s shipped to the customers. The digital trunk is shared with customers via web or Trunk Club app. For each item, customers can indicate if they agree with the selection (accepted items) or if they don’t like it (rejected items). Customers can also provide detailed item level feedback in the form of either free form text or by selecting from any of the pre-designed categorical options (such as too expensive, not my style, I own similar, etc.). Data science models leverage customer response during this step as additional data to train the recommendation systems.
Stylist Review
The customer feedback from Member Preview is used to update and finalize the trunk in the next stage called Stylist Review. This stage is primarily used to replace any rejected items from Member Preview. If the customer has provided any feedback on the rejected items (either open text or categorical), the stylist can use that to make a guided replacement. It is noteworthy to mention that there is no additional Member Preview step after this stage. The trunk is finalized, packed, and shipped to the customer for Home try on with some or all of the rejected items replaced.
Home try on
This is the final, and exciting, stage in which the physical trunk gets packed and shipped, and the customer gets to try on items at home. The customer can buy the items they like; the remaining items are returned at no cost to the customer. The customers can also exchange an item for a different color and/or size. Similar to the Member Preview stage, customers can provide detailed item level feedback in either free form text or by selecting from any of the pre-designed categorical options. The customer feedback is leveraged by data science models, along with the information on whether an item was kept or not. From a data science perspective this stage is exciting because we get to test the effectiveness of the models and recommendation systems.
… and now we present — Recommender Systems
Below we illustrate Data Science’s recommendation systems that enhance both stylist and customer experience during the aforementioned journey.
Outfit Recommendation
At the completion of onboarding customers go through the Express Preview experience. Here they get the first taste of Trunk Club in terms of a few example outfits curated based on their onboarding responses. Customers have the option to order one or more of these outfits in their first trunk. These curated outfits are generated from a data science model called Outfit Recommender. The model recommends personalized complete outfits based on customer’s onboarding questionnaire as well as historical data on the outfits curated by stylists for Trunk Club’s customers.
Brand Recommendation
In general, people associate their style preferences with a certain set of brands they like. Consequently, stylists rely on a customer’s brand preferences to identify potential items for the digital trunk. Data science has developed a brand recommender for the catalog with two key functionalities. Firstly, it utilizes brand attributes together with brand purchase history to identify topmost brands for a customer. Secondly, it uses similarity scores to show most similar brands. The model helps stylists in three potential ways:
- For a new customer, with no purchase history, the stylist can determine this customer’s likely brand preferences (other than the ones indicated during onboarding and communication with the stylist).
- For an existing/returning customer, stylists can not only ascertain if customer’s purchases match top brand recommendations, but also explore brands that haven’t been included in the trunks yet as well as observe customer’s changing brand preference over time.
- The brand similarity snippet helps discover novel and diverse brands thus alleviating popularity bias in stylist choices.
Personalized Sort
The personalized sort framework is an innovative Trunk Club technology that enables hundreds of thousands of products in the catalog to be sorted in real-time in a personalized order. Personalization is defined by how likely an item will be kept by a customer while ensuring it’s novel and relevant. While the data science model identifies how to calculate personalization score for every item in the catalog for a customer, the underlying architecture ingeniously designed by the data engineering team on top of Elastic Search computes the score and sorts catalog in real time. The architecture ensures the system is highly available with constant inventory and availability changes as well as is lightning fast with low latency for online consumption by all stylists. Personalized sort has several advantages for stylists:
- It can potentially save time as they can “pack” the trunk from topmost items without spending much time on navigating the catalog.
- It can help them avoid applying many catalog filters to find relevant items thus making them more efficient.
- Because the sort is already optimized for keep likelihood, by choosing from topmost items, the stylists can position the trunk for a high sales likelihood.
Fit recommendation
A customer’s likelihood of keeping an item depends not only on how well it reflects the customer’s taste, but also how well it fits. Data science has developed a comprehensive fit prediction framework to not only increase the likelihood of a good fit, but also improve stylists shopping efficiency. A stylist’s choice for the digital trunk is not only dictated by brand and style, but also if the customer’s fit for the item is currently available in the inventory. The fit prediction framework is a collection of numerous machine learning/AI models that stack their scores to predict the most likely fit for the customer for each item in the catalog. The modeling framework not only utilizes customer purchases, but also leverages fit related feedback from the previous trunks. Moreover, because the sizing format and range can vary substantially across numerous brands carried by Trunk Club, the modeling framework marries vendor sizing templates, customer indicated sizes, as well as data driven methods with business rules to determine which models to stack in order to generate the most accurate size prediction.
The figure below demonstrates how the fit prediction is deployed in the catalog. The model predicted size is indicated with a checkmark. The unavailable sizes are greyed out. The predicted size, even if unavailable, is indicated on the greyed out boxes. The stylists are not bound to choose only the predicted size. Based on their knowledge about the customer fit and preferences, they can choose other fits as well. The predicted size can also be different from the customer indicated size. The predicted size, if unavailable, can help the stylist pick another item for the digital trunk instead of packing the item in another size which, though available, could be a misfit. The framework thus helps reduce the negative fit feedback, improve the sales likelihood, as well as increase customer satisfaction.
Trunk Curation
A more scalable model for Trunk Club to grow its unique personalized styling service is to move the stylist from being responsible for everything to being responsible for those steps that require that unique human touch and expertise that only a stylist can give. To realize this vision, Product, Engineering, and Data Science are collaborating on an endeavor to develop a framework that provides the stylists a complete “model curated” trunk from the very beginning of shopping.
Particularly, the framework utilizes advanced ML/AI algorithms to not only suggest what to pack in the digital trunk, but also determine an assortment that can coherently constitute a trunk. The data science flow comprises two sets of algorithms. Depending on whether packing for a returning or a new customer, the first set of algorithms determine a coherent/connected set of trunk composition slots based on customer’s Trunk Request and past purchase history. The next set of algorithms populate the slots with personalized recommendations that are relevant for the current trunk. Such a framework allows a controlled, personalized scalable solution to packing trunks, while still utilizing stylists as the overall expert.
Trunk Completion
For the Stylist Review stage, Data science has developed a Trunk completion recommendation system to help determine what to replace the rejected items with. This algorithm particularly relies more on the accepted items and the rejected items from Member Preview to determine item recommendations. Item level feedback from Member Preview is also weighed heavily in the algorithm. It marries item level feedback with several custom business rules in order to determine inclusions and exclusions for the replacement items. The algorithm also incorporates explicit price sensitivity models and custom price constraints to ensure the recommended items remain within customer’s budget and price specifications. (Price limits and sensitivity models exist in other recommenders as well, though they are implicit with other features)
Continuous improvement mindset
All models, before complete roll-out, go through several rigorous A/B tests that measure business impact. More on A/B tests is discussed in Part 2. The Home try on stage provides information on the business impact of models through customer purchase and item level feedback. The more the data collected during this stage, the more the models learn about the customers and improve the quality of the recommendations, thus fueling a powerful flywheel of personalization and customer satisfaction.
Part 2 of the article discusses algorithmic details, validation and A/B testing, and challenges and future outlook.
