The concept of personalization is abstract. The most satisfying shopping experiences are those that are relevant and feel good. By personalizing a website, you can achieve more than the level of personalization that can be achieved in brick-and-mortar settings.

Technologies such as geolocation, face recognition, conversational commerce, augmented reality, AI, machine learning, and even marketing are transforming personalized shopping experiences in a dynamic, contextual, predictive, and proactive manner. We’re looking at an era of hyper-personalization.

According to Forbes and Mckinsey, the non-verbal demand has seen a major spike, with 80% of shoppers preferring buying from brands offering continued personalized experiences. 71% of customers just expect companies to deliver personalized interactions, and 76% of buyers get frustrated when this does not happen, which means increased page bounce rates and website drop-offs leading to a dip in sales.

Shopper and website interaction

Shoppers come to your website in various stages. They could be visiting for the first time or have a rich history of shopping on your website. Data from these shoppers helps build deep learning-based personalization models to create distinctive shopping experiences. Amongst the many ways that patterns can be broken down, here are three:

1. Repeat shoppers (user interaction history + product catalog)

The personalization model contains a sufficient history of such customers and can derive the affinity of the shopper towards brands, price, and style.

For instance, if a repeat shopper has shown a strong preference for a particular brand (Nike) or color of shoes (red), then the personalization model will promote the products of that particular brand and color – in this case, red Nike shoes.

2. First-time shoppers (user attributes + seasonality)

Personalization models do not have enough data history to detect the affinity of the shopper visiting the website. In this scenario, personalization models may apply user-related attributes or seasonality to personalize the results. For example, first-time users visiting from Canada searching for t-shirts see results from the fabric type “woolen,” while first-time users visiting from Africa see results from the fabric type “cotton,” depending on the time of the year or the geographical location.

3. Unspecified shoppers (user interaction)

The personalization model learns from the most recent interactions of the shopper on the site to promote products.

A shopper visits your site and starts interacting with the products on a “clearance sale.” This triggers the AI model to boost items on sale or relevant to “last searched for” based on recent interactions.

The mastery behind Unbxd personalization

Personalization is not as easy as it sounds. Unbxd AI algorithms analyze demographics, seasonality, browsing behavior, recent purchases, and affinities for each shopper to understand shopper intent better and dynamically curate results to show products that the shoppers are most likely to buy.

The AI model learns from automated feature detection, automated user profiling, segmentation & user context understanding, resulting in session personalization via real-time events integration.

Unbxd A/B tests the trained models to identify the high-performing products from each segment. We use up to ten models to achieve personalization. Some of the models are:

• YouTube DNN: Deep Neural Networks for YouTube Recommendations

• FM & NCF: Factorization Machines & Neural Collaborative Filtering

• DSSM: Deep Structured Semantic Models using Clickthrough Data

• SDM: Sequential Deep Matching Model

• MIND: Multi-Interest Network with Dynamic routing

No model can be considered a one-size-fits-all solution.

Unbxd can juggle between AI models and seamlessly select the best one for every scenario. Unbxd uses a Multi-Armed Bandit approach to choose the best model or version for each segment.

Segmentation defines an audience based on location, device type, or user type to show the shopper dynamic category pages that personalize on-site experiences. This segmentation helps run marketing campaigns or promotions for a particular segment. Explicit segmentation is done with the data that the user provides directly.

Let us explain one of the models: the YDNN model

The YouTube Deep Neural Network Structure (YDNN) model consists of the following two networks and can be considered a funneling approach to recommend user-specific recommendations:

1. The candidate generation network takes the user’s interactions history, search query, and demographics to display a few hundred products that could broadly apply to the user. This network mainly tries to optimize for precision.

2. Ranking networks use richer features (explained below) to score each product and show it to the user. This data is pulled up from the Candidate Generation Network.

The network is given a user’s interaction history until some time ‘t’ and asked what they would like to see at a time ‘t+1’.

On average, the YDNN model has brought about a 6% to 10% uplift in 12 weeks for new users. Imagine what the combination of models can do to boost your revenue!

Compared to traditional user personalization approaches like hrnn/hrnn-metadata, this model also considers search queries, user metadata, item metadata, and user interactions.

Overall, the YDNN model recognizes metrics, categorizes segments (through ranking data features), identifies the outcome for best conversions, and ranks products based on best-selling, most popular, etc.

How does data get collected: Learning to Rank (LTR) Model

If you have your own ecommerce website, you would want your top, popular or best-selling products to rank on top. However, if you are a retailer listing your products on another ecommerce site, you want your product to rank on top of the page amongst the other vendors. The LTR (Learning to Rank) model uses various features to learn to rank your products on a page. These features generate a score based on which you can personalize the shopping experience.

Data ranking models are a subset of personalization strategies. Dynamic ranking is looking at individual products and performance and seeing which product should be in which position. For example, Amazon has its own ranking logic – what brand comes first, products that need to appear on top, products facing starvation, or personalizing best-seller products based on intent. There are many ways to determine this; some determine the popularity, engagements, clicks, and carts. A combination of these creates a behavior score. The ranking score is different for each search and category. For example, Unbxd does ranking at a search, category, and website performance level.

Winding up

Delivering the right message to each consumer at the right time and place is challenging. Personalized experiences cannot be provided by traditional segmentation and only persona-based targeting.

Customer-centricity is essential to stay relevant in today’s hyper-competitive market. When choosing technology for your ecommerce website, think about your immediate goals and what will be needed in the future. After all, customers now hold power, and personalization is top-of-mind for modern consumers.

With Unbxd, retailers can adopt cutting-edge technology, keeping the shopper satisfied at every touchpoint, providing personalized product discovery, and increasing conversions and recall.

Do you want to see how Unbxd can benefit your ecommerce business? Request a demo