There are product combinations that aren't strongly supported by the model; in these cases, no recommendation is emitted. If you need a recommendation for every single input, you could choose to emit a random recommendation when there is no strongly supported recommendation, or you could choose to improve your model through better and generally larger training datasets.

• Start by creating a Maven project called arules-topology and add the following dependencies:
  复制代码
  
  
• Next, create a main topology class called RecommendationTopology using the idiomatic Storm main method. For this recipe, we will be receiving the product list as a JSON array on a Kafka topic. We will therefore need to coerce the byte array input into a tuple containing two separate values, one being the transaction ID and the other being the list of products, as shown in the following lines of code:
  复制代码
  
  
• We will also need to publish the output message using the Kafka partition persist. The recommendation and transaction ID need to be coerced into a single value consisting of a JSON array as follows:
  复制代码
  
  
• We then need to define the topology as described here:
  复制代码
  
  
• The Storm-R project's standard function supports only a known input array size. This works for most use cases; however, for the association case, the input size will vary for each tuple. It is therefore necessary to override the execute function to cater for this particular case as shown here:
  复制代码
  
  
• These elements are all that is required to create the recommendation engine. You can now start your topology in local mode from Eclipse. In order to test it, a test script is provided with the chapter code bundle named sendSelection.py. This takes a single parameter, which is the number of transactions, to publish onto the queue as follows:
  复制代码
  
  
• You can view the output recommendations by issuing the following command from the Kafka command line:
  复制代码