As a practicing AI product manager, you and your product team have successfully kick-started ML algorithms , overcome cold start challenges , and established regular model re-training via data pipelines to keep your product's recommender system up-to-date and highly performant.
But how can you further improve the product to address customer feedback and reach your next goals? In this article, we'll explore one approach to enhance the feature set of your recommender system by incorporating more data for more accurate recommendations.
The Need for Data Doesn't Stop Let's revisit our running example: you're an AI product manager leading a machine learning-driven product recommender solution on an e-commerce platform. Based on user feedback and performance analytics, you've discovered that while your recommender system works well during normal trading periods and for evergreen items, it struggles to respond to predictable events, such as:
Seasonal weather changes (from winter to spring to summer and so on). National calendar holidays (winter and summer school breaks, or religious holidays). Significant but infrequent social or sports events (Super Bowl, World Cup, Olympics). So far, your algorithm is primarily trained on the purchase history of your shoppers, which includes some seasonality signals from the past few years. But how can you enhance your recommender system to better forecast for these predictable events?
More Features via More Data Sources You may have heard the adage, "the more data, the better" or perhaps that more parameters lead to better predictive performance. However, the relationship between data and parameters is not that simple.
Predictive modeling analyzes historical data to forecast future values, such as customer demand or behavior. Relying solely on internal data, like customer purchase history, may not sufficiently capture the nuances of seasonal events or weather-related changes in consumer preferences.
Increasing the number of model parameters can improve its capacity to learn complex patterns, but this risks overfitting if the dataset is not large enough. On the other hand, adding more relevant data allows the model to generalize better and reduces both bias and variance.
One common approach to extracting more features is to tap into additional data sources.
This is where a tool like Airbyte can play an important role. Airbyte is a data integration platform that enables you to easily connect to various data sources and consolidate the data into a single destination. By leveraging Airbyte's connectors , you can efficiently incorporate external data, such as weather APIs or event calendars, into your recommender system's training and inference pipelines.
For example, Airbyte offers connectors for popular data sources like:
Google Calendar for importing holiday and event data OpenWeatherMap for current weather conditions and forecasts Google Analytics for user behavior data from your website or app However, it's important to keep in mind that more data is not always the answer, especially if the data quality is low. While extending the data feature set with additional sources is generally beneficial, be cautious about the quality of the data you feed your algorithms. Ensure that the data is accurate, complete, and relevant to your use case to avoid the "Garbage In, Garbage Out" (GI-GO) situation.
Improving Demand Forecasting with External Data Sourcing additional relevant, quality third party data can not only improve customer-facing product recommendations but also benefit business stakeholders in the supply chain teams of e-commerce businesses who desire more accurate demand forecasting. Without accurate availability of items, an e-commerce site would not be helpful recommending out-of-stock products.
Let's take the example of a demand forecasting algorithm that predicts how many items food retailers should stock to fulfill the forecasted customer demand for a given period. Such an algorithm might be initially developed based solely on internal online purchase history data. To improve prediction accuracy, these teams could similarly incorporate additional features from external data sets, such as weather forecasts and seasonal event calendars for the various countries where the e-commerce site operates.
However, it's not sufficient to extend only the training of the prediction model with external datasets. The team also needs ongoing data pipelines to tap into such data sources for inferring timely and accurate predictions. For example:
Easter holiday dates vary each year Sporting or concert calendars might get updated quarterly Weather forecasts change daily, even hourly Systems need to handle these dynamic inputs on an ongoing basis, a use case where Airbyte's connectors can serve both purposes seamlessly.
In Summary As an AI product manager, continuously improving your recommender system is necessary for staying ahead of the competition and meeting evolving customer needs. Incorporating new data features from external sources can significantly enhance the performance of your recommender engine, especially in handling predictable events like seasonal changes or weather-related trends.
In our next article, we will delve into the steps needed to establish a single customer view in case of fragmented customer interactions. Stay tuned!
