Content Recommendation

This guide takes you through the detailed steps to set up Seldon to serve content recommendations.

A worked step-by-step example using the Movielens 100K dataset is provided.


The process of creating and serving content recommendation models is summarized in the diagram below:

Content Recommendation

Live activity is sent to Seldon via the Seldon API. This will include user activity (e.g. which pages they viewed on a website) as well as item data (e.g. details about new articles published on a website with their metadata). Offline content recommendation models will be created. Presently Seldon provides several Spark based models as well as some examples in python using libraries such as gensim and scikit-learn. For runtime content prediction we need a runtime scorer to take the models created and use them to provide recommendations in real time. Seldon provides several builtin runtime scorers for the various Spark based models. Optionally, a microservice can be deployed to do the runtime scoring. An example microservice deployment is shown here.

Create client and meta-data schema

To serve content recommendation you first need to create a client which will have an associated consumer key.

You can create this via with seldon-cli client.

Next you will need to define the item meta-data schema. Here is an example schema for an item representing a music album:

    "types": [{
            "type_id": 1,
            "type_name": "music",
            "type_attrs": [
                {"name":"sales_count", "value_type":"int"}

Let’s go though these fields one by one

  1. type_id: Distinguishes between different types of items for example movies and music.
  2. type_name: unique name for this type of item
  3. type_attrs: a list of attributes that can be associated with this item type
  4. type_attrs -> name: the name of the attribute in question
  5. type_attrs -> value_type: what type of data to expect for this attribute. Valid values are ‘string’, ‘text’, ‘double’, ‘datetime’, ‘boolean’, ‘int’ and a list. The list is a special case where the data can be one of a restricted list (an enum essentially)

One you have created the attributes JSON file you can associate it to your client using seldon-cli attr.

Import historical Data

If you have historical data you will to use with the built in Seldon Spark jobs then you need to convert it into JSON format with references to the internal IDs used by seldon for the users and items. This can be done using the Seldon CLI as described below..

Add historical items

Items must be provided as a CSV and conform to the schema with 1 or two extra fields. An example for the schema above is


Note that we have ‘id’ and ‘name’ that were not mentioned in the schema. ‘id’ is a required field for all items in all schemas. It can be any unique string and is your identifier for the item. ‘name’ is optional and should be a string that you might use to search for the item. A few other things to mention here are

  1. Boolean fields (is_compilation for example) can be 0 or 1, 0 meaning false and 1 meaning true.
  2. Enum fields (genre for example) must be one of the values you defined in the schema
  3. You must provide a header line
  4. There is an example in /your_data/items_data/example_items.csv

Load the CSV into Seldon using seldon-cli import.

Add historical users

Users are much easier as currently it is not presently possible to specify a schema. So we just need an id and optionally a username:


Load the CSV into Seldon using seldon-cli import.

Add historical actions

If you have existing historical activity data you can import these “actions” into seldon if they can be provided as a CSV file.

Actions again have no schema to contend with but we need a few extra fields:


The first two columns should be obvious. ‘value’ is a field that represents the magnitude of the action. If all actions are created equal, then you should just set this to one. ‘time’ is the unixtimestamp of the action.

The actions are not added to the DB, but they require transformation so that the Spark jobs can consume them.

Use seldon-cli import to ingest the actions data.

Ingest activity via API

In production (or if you have no historical data) you would send new user activity and item meta data to Seldon via its REST and JS API.

This activity (actions) data - is then transformed and re-located to a central location by Fluentd.
The actions data can be for multiple clients if necessary.

The Seldon CLI tool can then be used to process this data to separate it for each client via a spark job, see example seldon-cli client --action processactions.

Create a recommendation model

Recommendation models can be built using any available technology that can be Dockerized and run inside Kubernetes. However, we provide some pre-packaged Spark based models and associated runtime scorers for those models. We also provide a python library which allows you to build and create models and runtime scorers exposed as microservices.

We use luigi to package the offline model creation process. The luigi task can be run as a Kubernetes Job. In production you might run the modelling process each day to update your models with newly ingested data. We provide Make process to create Kubernetes Jobs in kubernetes/conf/models. Call the Makefile with client and start day variables to create a Kubernetes job. Som examples are shown below.

You can create a matrix factorization Kubernetes job for client “test” starting at unix-day 16907 (17th April 2016) as follows:

cd kubernetes/conf/models
make matrix-factorization DAY=16907 CLIENT=test

This will create a Kubernetes Job file in the jobs folder called matrix-factorization-test-16907.json. Looking inside this JSON you will find a definiton of an image which calls the luigi command to run the matrix factorization job:


"name": "matrix-factorization",
"image": "seldonio/seldon-control:1.3_v4",
"command": ["luigi","--module","seldon.luigi.spark","SeldonMatrixFactorization","--local-schedule","--client","test","--startDay","16907"],


The luigi Task definition can be found in our pyseldon library in seldon.luigi.spark.SeldonMatrixFactorization. In this case it simply calls the seldon-cli to run a matrix factorization job. Other cases might have more complex luigi jobs.

You can provide your own custom configuration either by changing the luigi.cfg or supplying further parameters to the call to luigi.

Built-in Models

We provide several recommendation algorithms as part of Seldon out of the box. See here for a full description.

Configure runtime recommendation scoring

Once a model is built the final step is to provide a runtime scorer for the model. You should choose an associated runtime scorer for your particular model as outlined, so for example if you built a Matrix Factorization model you should use an associated scorer, e.g. recentMfRecommender or mfRecommender.


If your runtime scrorer will be exposed as an internal microservice you need to package it as a Docker container that exposes the microservice recommendation API. Once done you can start it using the command line script start-microservice.

The script create a Kubernetes deployment for the microservice in kubernetes/conf/microservices. If the microserice is already running Kubernetes will roll-down the previous version and roll-up the new version.

Serve recommendations

Recommendations can be accessed via the Seldon API. For initial testing you can use seldon-cli api

Worked example

A worked step-by-step example using the Movielens 100K dataset is provided.

Advanced Settings

Run A/B Tests

When running multiple recommendation models in production you will want to A/B new models to check they perform better than existing models with live clients before you place them fully into production for all users.

The ability to run A/B and multivariant tests is available within Seldon. You can find the details here

Combine Multiple Algorithms

In some setting you may wish to combine multipe algorithms together to get a combined result. You can find the details here

API Controlled Variants

In some more complex content recommendation installation you may want to control for a particualar single client (single API key) various different algorithms for different settings, e.g. provide in-section Sport content recommendations and cross-site general content recommendations in different parts of a web page. The changes need to implement this are discussed here