Over a million developers have joined DZone.
{{announcement.body}}
{{announcement.title}}

Scaling a Sales Recommendation Engine With Apache Spark and MongoDB

DZone's Guide to

Scaling a Sales Recommendation Engine With Apache Spark and MongoDB

Last time, we covered building a sales recommendation engine with Apache Spark. To go one step further, I have replaced the file system layer with MongoDB.

· AI Zone ·
Free Resource

Did you know that 50- 80% of your enterprise business processes can be automated with AssistEdge?  Identify processes, deploy bots and scale effortlessly with AssistEdge.

In my last post on building a sales recommendation engine with Apache Spark, we built a standalone version of spark-mllib's ALS demonstration based using local CSV files.

ml-bigdata-starter/README.md at master · ERS-HCL/ml-bigdata-starter

To go one step further, I have replaced the file system layer with MongoDB. MongoDB provides a spark-mongo connector that wraps the standard Java/Scala connector with Spark's interpolatable data format/APIs.

Getting Started

Apart from the spark core APIs, you need the following dependency to connect to MongoDB server.

ml-bigdata-starter/README.md at master · ERS-HCL/ml-bigdata-starter

<dependency>
<groupId>org.mongodb.spark</groupId>
<artifactId>mongo-spark-connector_2.11</artifactId>
<version>2.2.1</version>
</dependency>

Preparing the Datasets

In the current scenario, i.e MongoDB, instead of creating files, we need same data in JSON format in collections.

Sales orders:

UserId UserName ProductId ProductName Rate Quantity Amount
1 User 1 1 Product 1 10 5 50
User 1 2 Product 2 20 10 200
1 User 1 3 Product 3 10 15 150
2 User 2 1 Product 1 10 5 50
2 User 2 2 Product 2 20 20 400
2 User 2  4 Product 4 10 15 150

Sales leads:

UserId UserName ProductId ProductName
1 User 1 4 Product 4
1 User 1 5 Product 5
2 User 2 3 Product 3
2 User 2 6 Product 6

We need to predict/recommend the most relevant product for both the users based on their past order history. Here, we can see that both User 1 and User 2 ordered Product 1 and Product 2; also, they ordered one item separately.

Now, we predict their rating for alternate products and one new product.

Implementation

Step 1

Our first step is making a database connection using MongoDB specific properties.

SparkConf conf = new SparkConf().//
setAppName("rnd").//
setMaster("local").//
set("spark.mongodb.input.uri", "mongodb://127.0.0.1:27017/sparkdb.myCollection").//
set("spark.mongodb.output.uri", "mongodb://127.0.0.1:27017/sparkdb.myCollection");

Step 2

Now you can read training model via JavaMongoRDD API and convert it to Rating format using JavaRDD API.

private static JavaMongoRDD < Document > getJavaMongoRDD(JavaSparkContext jsc, String collName) {

 Map < String, String > readOverrides = new HashMap < String, String > ();
 readOverrides.put("collection", collName);
 readOverrides.put("readPreference.name", "secondaryPreferred");
 ReadConfig readConfig = ReadConfig.create(jsc).withOptions(readOverrides);

 JavaMongoRDD < Document > mongoRDD = MongoSpark.load(jsc, readConfig);
 return mongoRDD;
}

JavaMongoRDD < Document > salesOrdersRDD = getJavaMongoRDD(jsc, "SalesOrders");


// Map file to Ratings(user,item,rating) tuples
JavaRDD < Rating > ratings = salesOrdersRDD.map(new Function < Document, Rating > () {
 public Rating call(Document d) {
  return new Rating(d.getInteger("userCode"), d.getInteger("productCode"), ((Number) d.get("amount")).doubleValue());
 }
});

Step 3

The next step is to train the matrix factorization model using the ALS algorithm.

ml-bigdata-starter/README.md at master · ERS-HCL/ml-bigdata-starter

MatrixFactorizationModel model = ALS.train(JavaRDD.toRDD(ratings), rank, numIterations); 

Step 4

Now, we load the sales lead file and convert it to tuple format.

ml-bigdata-starter/README.md at master · ERS-HCL/ml-bigdata-starter

JavaMongoRDD < Document > salesLeadsRDD = getJavaMongoRDD(jsc, "SalesLeads");

// Create user-item tuples from ratings
JavaRDD < Tuple2 < Object, Object >> userProducts = salesLeadsRDD.map(new Function < Document, Tuple2 < Object, Object >> () {
 public Tuple2 < Object, Object > call(Document d) {
  return new Tuple2 < Object, Object > (d.getInteger("userCode"), d.getInteger("productCode"));
 }
});

Step 5

Finally, we can predict the future rating using a simple API.

ml-bigdata-starter/README.md at master · ERS-HCL/ml-bigdata-starter

// Predict the ratings of the products not rated by user 
JavaRDD<Rating> recomondations = model.predict(userProducts.rdd()).toJavaRDD().distinct();

Step 6

Optionally, you can sort the output using a basic pipeline operation:

ml-bigdata-starter/README.md at master · ERS-HCL/ml-bigdata-starter

// Sort the recommendations by rating in descending order 
recomondations = recomondations.sortBy(new Function<Rating, Double>() {  
  @Override  
  public Double call(Rating v1) throws Exception {   
    return v1.rating();  
  }
 }, false, 1);

Step 7

Now, you can display your result using the basic JavaRDD API.

ml-bigdata-starter/README.md at master · ERS-HCL/ml-bigdata-starter

// Print the recommendations . 
recomondations.foreach(new VoidFunction<Rating>() {  
  @Override  
  public void call(Rating rating) throws Exception {   
    String str = "User : " + rating.user() + //   " Product : " + rating.product() + //   " Rating : " + rating.rating();   
    System.out.println(str);  
    } 
  });

Output

User : 2 Product : 3 Rating : 54.54927015541634
User : 1 Product : 4 Rating : 49.93948224984236

Conclusion

The above output recommends that User 2 would like to buy Product 3 and User 1 would go for Product 4.

This also recommends that there is no recommendation for new products, as they do not match any similarity criteria in past.

Consuming AI in byte sized applications is the best way to transform digitally. #BuiltOnAI, EdgeVerve’s business application, provides you with everything you need to plug & play AI into your enterprise.  Learn more.

Topics:
spark ,mllib ,mongodb ,recommendation engine ,ai ,tutorial

Opinions expressed by DZone contributors are their own.

{{ parent.title || parent.header.title}}

{{ parent.tldr }}

{{ parent.urlSource.name }}