Suggestions that delight

Really fast. Most recommendation requests will take from 1 to 20 milliseconds to process.

Under the hood recommendations rely on supervised machine learning models and the Algolia foundation.

For both models, the data corresponding to the past 30 days is collected. This results in a matrix where columns are userTokens and rows are objectIDs. Each cell represents the number of interactions (click and/or conversions) between a userToken and an objectID. Then, Algolia Recommend applies a collaborative filtering algorithm: for each item, it finds other items that share similar buying patterns across customers. Items would be considered similar if the same users set interacted with them. Items would be considered bought together if the same set of users bought them.

Getting recommendations is a four-step process:

• Capture your users’ conversion events
• Send your data to Algolia
• Train the models with the push of a button
• Add recommendations to your UI

Our recommendation engine is language-agnostic: it supports alphabet-based and symbol-based languages (such as Chinese, Japanese or Korean).

Essentially a recommendation engine will analyse interactions of users with different items to draw links between those items. Deep dive here.

An example of a recommendation engine is a product recommendation engine for ecommerce. It will analyse what products shoppers buy together or what products shoppers interact with in a short amount of time, to generate “Frequently Bought Together” or “Related Products” recommendations. Learn more here!

The key components of a high-performance recommender system are: Data Sources, Feature Store, Machine Learning Models, Predictions & Actions, Results & Metrics. More details in this dedicated series.

The best way to improve a recommendation engine is to make sure you’re feeding it qualitative data: user interactions and items. Additionally there are filters that you can apply to the recommendations that are being generated. Ultimately, key performance indicators must be accurately tracked in order to identify areas of improvement.

The most obvious operational goal of using a personalized recommender system is to recommend items that are relevant to the user, as people are more likely to buy items they find attractive. Learn more about personalized recommendations and their benefits here!

Content-based recommendations are based solely on items’ descriptions. Personalized recommendations are also based on user’s interactions and each user will see a different set of recommendations, depending on their individual preferences. Learn more here!