Previous proposalApproximate Query Processing
Applying Lambda Architecture in Machine Learning realm
In mature information retrieval systems, predictions and scoring happen in multiple layers in cascaded fashion. In batch processing layer, update intervals are big and disperse. In the ingestion layer, it is done as and when the updates arrive,close to near real time. This layer is non user-path but still carries a reasonably wide feature set. Lastly, final scoring is done in user path using a much smaller yet important feature set. These layers can be seen as part of a spectrum covering a range of tradeoffs in computing predictions. To build non-overlapping layers, we need to introduce feature classification as,
- Product Features - slow changing
- User Features - available only in request path (Geo, various user affinities e.g. brand etc)
- Fast Changing business metric governing features - e.g. price, offers and availability. Needed for freshness and for an optimal ordering of products to users
On the spectrum from pure batch(left) to pure real-time(right), the cost of sourcing features and score computation involved varies immensely. The batch size reduces drastically from left to right. Feature fluctuations increase from left to right. Sensitivity to latency increases from left to right. The batch get normalized on aggregate data and is not as pure as real-time. While implementing these layers, we observed that all parts of this spectrum are equally important to counterbalance anomalies introduced by individual layers.
Putting succinctly, in this talk we’ll cover few use cases where we did a series of experiments at different layers with varying feature sets. We’ll go into how these patterns are applied at scale in Flipkart search and recommendation systems for scoring candidate result set, given the query or product context respectively.
In this talk, we’ll cover :
a) Overview of different feature types for information retrieval and ranking systems, and how important is the freshness aspect
b) Different processing layers : Batch, Indexing and Real-Time, characterized by the reaction time to feature updates
c) Tradeoffs of doing scoring computation in a cascaded fashion in the real world
d) Case Studies : Taking examples from Flipkart search and recommendation systems, we’ll cover how these different layers are employed for production use cases of retrieval and ranking
Akash is a software developer with Search Relevance team at Flipkart, working on improving Autosuggest. Previously, he has worked on building Flipkart Recommendation System. He designed real time and batch pipelines to power recommendations, including use cases such as product bundling, similar products and personalisation. He is interested in applying Machine Learning for pattern mining, and deploying data processing pipelines at scale. He graduated with a dual degree in Computer Science & Engineering from IIT Delhi.