Learning to rank and ML pipelines

Ranking is a core task in information retrieval (IR) and recommender systems. In previous posts, I talked about search engines and machine learning (ML), and now we will see how to apply ML to rank (or re-rank) search results. Additionally, I will discuss why ML workflows should be automated and look into a few alternatives to build ML pipelines.

Search relevance

Even when using existing technologies like Elasticsearch, engineers still have to tune their index/queries to optimize results for relevancy. For example, when searching news, the title is more relevant than the body and recent news are more relevant than older ones. Tuning a search engine for optimal results can be quite challenging and requires proper instrumentation and and validation. You can find a presentation about using Logistic regression to select boost parameters here.

Although machine learning is computationally expensive, it can be applied for ranking problems and return highly relevant results using plugins like Elasticsearch Learning to Rank (LTR), which supports linear models and xgboost, a tree-based model. Numerical features are required to train ML models and significantly impact performance metrics. The MSLR dataset is often used to benchmark LTR models and contains query/relevance judgments, with features like tf-idf statistics, pagerank, bm25 and so on.

LTR workflow

As you can’t improve what you can’t measure, let’s assume you already have some infrastructure in place, like an analytics solution to measure search relevancy (manual/crowdsourced relevance judgments or click/session data). Afterwards, you can gather the data, build features, train different models/hyperparameters and deploy the best one. CI/CD for ML projects is harder to get right as additional artifacts should be tracked and the models need retraining as performance drifts over time.

Python, Scala and R are among the most popular languages for ML and they’re all suitable for NLP/Ranking as they have a healthy ecosystem of numerical packages with efficient sparse matrix representations, text processing techniques and ML algorithms. Python has the edge, especially with transformers, while Scala is popular for data streaming (Flink, Kafka streams, Spark streaming) and R is the best for data visualization while more limited for production workloads. In this repo you will find a Python notebook to train different LTR models on the first fold of MSLR-WEB10K, using pontwise/pairwise/listwise approaches, along with a comparison of Python and Scala for prediction.

These are the metrics I got on the test set with ~240k samples:

1
2
3
4
5
6
7
8
9
10
11
12
Linear model with standard scaler:
- [email protected]: 0.4
- training: 7s
- python predict: 80ms
- scala predict: 50ms

XGBoost model:
- [email protected]: 0.5
- cpu training: 52s
- gpu training: 16s
- python predict: 800ms
- scala predict: 400ms

The batch inference times in Python are very fast, as most code executed is actually native. I haven’t made tests in a real-time inference scenario (e.g.: REST API or queue/stream), but even if Python is slightly slower serializing/deserializing or transforming data, it is trivial to increase the number of instances and setup load balancing. The development effort of deploying workloads on more performant languages, such as Scala, does not seem worth it for the small latency gains. LTR is also usually applied as only a re-ranking strategy on the top K results to reduce loading times.

For more info on LTR:

ML Pipelines

CI/CD systems like Jenkins or Github Actions allow workflows to be triggered via API/Cron, but they are not suitable for ML workflows, since they lack some features/integrations and agents may not have the required resources. A possible alternative is Airflow, which is commonly used for scheduled ETL workflows and has operators to run tasks in Kubernetes (k8s) and cloud services like AWS Batch/ECS. I didn’t find the development experience to be the fastest (build docker image, upload image, update dag, trigger dag) and would require maintaining more glue code for ML tasks. Prefect Orion is a mature successor to Airflow, with a simpler architecture and better development experience, but since it is a generic workflow orchestration platform, it requires additional work to integrate with the ML ecosystem.

A viable alternative, which has a low maintenance cost, though higher lock-in, are the could offerings like AWS Sagemaker and Databricks with MLflow. However, my favorite alternative so far is Metaflow, a modular framework to build ML pipelines. Tasks can be run locally, in AWS Batch or in k8s, allowing data scientists to develop and test the pipelines from their machine, handles dependency management with Conda, has caching, integrates with existing schedulers (AWS Step Functions, Argo Workflows), provides a web UI, and the Python client can be used to retrieve models or experiment metrics. Metaflow does not handle model deployment or monitoring, but it can be done with a specialized platform like Seldon or by reusing the current organization’s infrastructure. I recommend reading the excellent book Data Science Infrastructure if you are interested in learning more about the topic.

These articles cover ML pipelines in more detail: