Items are recommended to users based on collaborative filtering. The implicit feedbacks, e.g. views of details pages of items, are collected and used to recommend similar items to users. The recommendations are computed using algorithms from Apache Mahout. The data used by Mahout for the collaborative filtering must contain for each record the userID (long), the itemID (long) and a value (double). Nonetheless the itemIDs are assumed to be of type string. Mahout provides a special hash map with the mappings between the string Id and the numeric Id. The data model is built from the user navigation log files using the code in the producer repository. The file containing the aggregated records must be set in the config.properties file. The recommendations are available as a web service. The user sends the userid and the number of items to be recommended and the service returns the items and the score. The service can use different algorithms with different data models. A servlet is started and is initialized with a data model in the web.xml file. The signals are read from a base file, e.g. signals.csv, and from other files in the same folder whose name begins in the same way as the base file, say signals-20180424.csv. The signals data are kept in a Mahout FileDataModel. If a new signal file has been created and the last refresh was before the minimum reload interval time, the data can be reloaded and the recommendations updated sending a refresh request to the recommender.
You need Java 8 and Maven to build the code. Jetty is used as a servlet container.
$ git clone https://github.com/luigiselmi/mahoutrecommender.git
The recommender is a web service and it must be packaged as a war file and installed in a servlet container such as Jetty.
$ mvn install
You can run the recommender as a Java Web application in a servlet container such as Jetty or in a docker container.
In case Jetty is used as a servlet container you can create a base for Jetty, as explained in the Jetty documentation, and copy the war file in the webapps folder. You can rename the war file as ROOT.war so you will not need to use a context. From the Jetty base folder execute the command
$ java -jar $JETTY_HOME/start.jar jetty.http.port=8100
You can build the recommender's docker image using the Dockerfile provided in this repository executing the command
$ docker build -t doeeet/recommender:v0.1.0 .
In order to run the docker container you need to specify the docker volume containing the signals files with the mapping from the folder in the host machine (e.g. /storage/mahout ) to the folder in the container where the recommender will look for the data (e.g. the same as in the host machine).
$ docker run -d -v /storage/mahout:/storage/mahout -p 8100:8100 --network=doeeet-net --name recommender doeeet/recommender:v0.1.0
The running container can be inspected using the command
$ docker exec -it recommender /bin/bash
You can send a request to the web service using the HTML page at http://localhost:8100/ with the id of the user for which you want the recommendation and the number of items to recommend, ore use curl
$ curl 'http://localhost:8100/recommend?userID=21585&howMany=2'
The response is a JSON stream with a ranked list of recommended items
{
"userID":21585,
"recommendedItems": [
{
"itemID": "40526f54-3c60-4d87-9628-368524cff90c",
"value": 3.945368528366089
},
{
"itemID": "b352ac86-28ca-4c21-b586-c0edfa19ac0b",
"value": 3.4908947944641113
}
]
}
In order to update the recommendations when a new signal file is available in the signal folder you can send the following request to the recommender
$ curl 'http://localhost:8100/refresh'
The recommender will reply with a message containing the updated number of users and items.
0.1.0
None
Apache 2.0