Approach for Content Based Recommendation Engine in AEM

  1. RecommendationsGeneratorScheduler runs on regular interval & starts the process for generating recommentations
  2. Data extraction begins by running the query passed by users via "AEM Recommendation Engine" OSGi Config
  3. Using DataCleaningUtil.java bogOfWords is generated
    1. Properties to be read are passed in the DataCleaningUtil constructor.
    2. generateBagOfWords() method returns a map of ProductId mapped to bagOfWords associated to that product. bagOfWords is just space separated String with different Tags, or other values of properties like jcr:title, cq:tags etc.
  4. showCountVectorizer() method from GenerateRecommendations is called. Note: This class is just calling all the required methods from Tokenizer, Dictionary & CountVectorizer classes to convert each product with words associated to them into numeric values which can be passed to an Algorithm for getting recommendations.

  5. Generating recommendations:

  6. Tokenizer object is created using custom SimpleTokenizer implementation
    1. SimpleTokenizer provides a method to generate Array of tokens from BagOfWords liked to each product
  7. Dictionary object is initialized using custom SimpleTermDictionary implementation
    1. SimpleTermDictionary maintains a Vocabulary
    2. SimpleTermDictionary provides methods:
      1. getTermIndex(): returns the Index (Number given to each term in Dictionary) of given term.
      2. getTotalTerms(): returns the total number of terms/items in Vocabulary
  8. CountVectorizer Object is initialized with dictionary & tokenizer objects initialized in previous steps. CountVectorizer provides 2 methods:
    1. getCountVector() : which returns the SparseVector of the current product's bagOfWords.
      1. returns RealVector object
      2. SparseVector of size of Dictionary's Vocabulary size
      3. SparseVector which has represents a word as 1 in the array while 0 if that word from vocabulary is not part of Products BagOfWords
    2. getCountMatrix() : returns RealMatrix
      1. returns a matrix of Vectors for each Product's bagOfWords
  9. In SimilarityMatrixGenerator class's generateSimilarityMatrix() method RealMatrix generated using getCountMatrix() method is looped through and:
    1. For Each row i.e (each product's vector) cosine angle is calculated with all the other products
    2. And cosine is stored in a dotProductMatrix
    3. Finally SimilarityMatrix Object is created which contains dotProductMatrix and NodeIdIndexMap (mapping nodeId to Index, which is used for reading recommendations from dotProductMatrix)
  10. A separate Map is maintained which maps each ProductId (String) to the Index of product in RealMatrix from step 9.
  11. SimilarityMatrix object is serialized & stored in JCR under /var, which can be later read while getting the similarities. And Recommendation generation is an heavy operation, RecommendationsGeneratorScheduler can generate & update the SimilarityMatrix after fixed interval.
  12. RecommendationsReaderService can be used for reading the recommendations from SimilarityMatrix serialized & stored within JCR.

Note: There can be several RecommendationEngines configured in any AEM instance for generating different recommendations for different pages, products or items.