|
Listen to this story
|
Search ranking is the backbone that any e-commerce website stands on. It is directly tied to the visibility and accessibility of products and, consequently, to sales as well. But like we learn, there can be more finesse.
Analytics India Magazine caught up with Kavitha Krishnan, Senior Manager, Data Science at Tredence, to understand how new AI algorithms can build a better search experience and the effectiveness of a good search ranking.
AIM: What are some of the challenges that e-commerce websites face when it comes to search and how can they overcome such challenges?
Kavitha: There are 2 main challenges in a search cycle.
(a) To generate the most relevant search results with respect to a query
(b) To decide the ranking/sorting for these results.
While a lot of attention is typically paid to the first challenge, the importance of addressing the second challenge should not be overlooked.
Effective search ranking algorithms are crucial for maximizing the visibility and accessibility of products to potential customers, which in turn can increase the likelihood of a purchase and enhance user satisfaction. When it comes to e-commerce websites, the key question is how to rank products on a search results page in a way that satisfies multiple objectives, including maximizing relevance and purchase likelihood. Achieving a balance between these objectives can be a complex task that requires leveraging AI/ML techniques.
For example, when a customer searches for snacks on an e-commerce website, the website may display hundreds of snack products. However, if the most appropriate products are not displayed prominently, the website may miss out on potential sales. Balancing multiple competing objectives to ensure an optimal customer experience and higher sales/margin for the business is critical. If a website consistently displays only higher-priced products, for instance, it may deter customers and negatively impact sales. Finding the right balance is key.
AIM: What are some of the different approaches to search ranking and how do they work?
Kavitha: Develop a system to scientifically sort products on the search results page to optimize “Relevance”, “Add to cart rates”, “Gross Margin Dollars” etc. Balancing multiple objectives is the main goal of the ranking exercise.
The most appropriate algorithms for this use case are ‘Learning to Rank’ and ‘Multi-Objective ranking optimization’ systems.
Learning to Rank systems use machine learning to understand behavioural patterns and generate ranks based on a combination of features, such as product descriptions, reviews, ratings, and inventory. The weights of these features can be adjusted based on business goals and priorities or left to the model to tune and learn. This algorithm considers the multidimensional nature of relevance and business constraints, making it suitable for building relevance ranking models in production. Ranking models typically work by predicting a relevance score for an input feature set containing the query term and the list of products and their features. And we sort these products according to the generated relevance score to determine the ranking.
Multi-Objective Ranking Optimization is the task of learning a ranking model from training examples while optimizing multiple objectives simultaneously. A ranked list conforming to all business requirements can be generated by this system, also known as MORO. Both Learning to Rank and Multi-Objective systems can be easily incorporated into search engines for any retailer to improve overall user experience and generate maximum profit.
AIM: How do you think these search ranking algorithms take into account factors like the relevance of the product or its popularity or its profitability?
Kavitha: When websites rank their products, they need to factor in various aspects such as relevance, popularity, and profitability.
Out of these, relevance holds utmost importance as it ensures that the displayed product matches the user’s search query. For example, if a user searches for “organic carrots,” the website should show relevant results such as organic carrots before showing normal carrots or baby carrots. The algorithm looks at several aspects such as keyword density, title tags, meta descriptions, and content quality to generate this relevance aspect. Using word embeddings of the search term and the product terms helps in generating the right relevance.
Popularity is determined by the number of views or purchases made earlier, along with customer reviews and ratings. For instance, when a user searches for hotels on a booking site, the website displays a list of popular hotels based on customer reviews and ratings. Similarly, the products with higher engagement and customer reviews should get boosted.
The profitability factor should be taken into account as another consideration. Focusing only on the relevance and popularity of products may yield a better customer experience, but may result in inadequate profits and failure to meet business goals. It’s essential to factor in pricing and profit margins when determining the ranking of products.
Websites need to incorporate all these factors into the ranking algorithm that generates the search results list, especially emphasizing the top few rows of products, to achieve an optimized search grid.
AIM: What are the indicators for e-commerce companies to assess the performance or effectiveness of search ranking algorithms?
Kavitha: The most important way to assess the performance and effectiveness of search rankings is to use A/B testing. A/B compares two versions of a webpage to determine which one performs better. Users are randomly split into two groups and shown different versions. Deliver the new search ranking to a subset of the customer, while others still see the existing search grid. Identify the uplift in business and click stream metrics like Add to Cart rate, Engagement rate, Abandonment rate, and Gross Margin between these two groups. With this A/B testing, we will be able to access how good our search ranking process is.
In addition to A/B testing, there are other ways to assess the performance and effectiveness of search rankings. One key metric is the average time spent on the search results page. A well-optimized search grid should reduce the time users spend searching for the right product. By analyzing these metrics, retailers can identify areas for improvement and make data-driven decisions to optimize their search ranking algorithms. Additionally, it’s important to regularly monitor and evaluate search metrics to ensure that the search grid continues to perform effectively and meet business goals.
AIM: How do you think AI/ML can help with search rankings? How is this tech developing?
Kavitha: Our focus is on addressing the challenge of “Converting searches into sales”, and I firmly believe that leveraging AI/ML is the optimal approach for accomplishing this goal. By identifying the barriers that prevent online shoppers from discovering the products they seek and making purchases, AI/ML algorithms help retailers to enhance the search experience and effectively meet business goals.
The key characteristic of an exceptional search experience is the ability to promptly provide every user with the most pertinent results. However, the challenge lies in efficiently and gracefully executing this task behind the scenes. So choosing the right algorithm and right implementation strategy is the key. Making even minor enhancements to the relevance ranking system can have a dual effect of improving the shopping experience for millions of customers and significantly boosting revenue.
Although the application of learning to rank (LTR) in web search has been extensively researched, its utilization in E-Com search remains unexplored to a very large extent. Using the right feature representation and obtaining dependable relevant features and utilizing user feedback as features are still areas that need a lot of Business +AI acumen.
Our goal is to provide the correct perspective for solving the problem in the E-commerce use case, despite the availability of various algorithms that can tackle it. Unfortunately, only a small number of retailers have taken full advantage of these systems so far. Google and Amazon have even introduced new approaches to address the ranking problem in recent times, indicating how significant it is for retailers to act. Therefore, retailers must seize this abundant opportunity at hand. The most effective way to advance is to leverage the potential of AI/ML, employing the latest and most effective algorithms and integrating them into your E-commerce search engine.
