Recommendation systems workflows lectures

In brief

The Recommendation Systems (RS) lectures outlined below are to be recorded for Wolfram U in October-December 2020.

• Each tutorial lecture is between 25 and 40 minutes.

• Most of the workflows in the lectures have both WL-code and R-code versions.

• We are going to extensively use the Domain Specific Language (DSL) Raku programmed system DSL::English::RecommenderWorkflows, [AA1], that generates code for the software monads SMRMon-WL and SMRMon-R; see [AA2,AA3].

General description

In this tutorial cycle we consider workflows of typical recommendation systems. We use both item-item and item-user recommenders. The item-item recommenders are based on information retrieval principles; the item-user recommenders are based on Graph theory and collaborative filtering principles.

For clarity we use both programming code and sequences of natural language commands. We also introduce and exemplify the use of different recommender interactive interfaces.

The discussed recommender workflows are constructed with dedicated R packages. We mostly use (monadic) pipeline specifications.

All parts of the tutorial are illustrated with a diverse set of examples based on the following types of data: digital content (books, movies, music), financial time series, document collections, retail invoices, medical data.

The following plan is (roughly) followed:

1. General set-up: problem formulation, scope, areas of focus.

2. Description of the standard workflows followed by more advanced or specialized workflows.

3. Elements of the theory of different aspects of recommendation computation algorithms and related interpretations.

4. Demonstration of different extensions of recommendation systems for doing progressive learning and time series collections analysis.

Detailed tutorial plan

• General set-up

  • The recommendation problem
  • Approaches
  • Recommendation systems types
  • Introductory examples

• Core workflows

  • Creation
    • Creation by contingency matrices
    • Creation by specifications
  • Categorization of numerical data
  • LSI functions application
  • Recommendation by history
  • Recommendation by profile
  • Explanations and proofs
  • Recommendations re-ordering
  • Cut-off selection by outlier identifiers

• Further workflows

  • Using dimension reduction
  • Tuning recommendations
  • Recommendations breakdown
  • Classification
  • Anomalies identification

• Computations theory

  • Sparse Matrix Linear Algebra
  • Using scored-lists / hash-maps
  • Using graphs and centrality measures
  • Pre-computation of nearest neighbors scored-lists
  • LSA and LSI utilization
  • Designing more specialized and/or advanced algorithms
    • Based on quick recommender computations

• Recommender object transformations

  • Joining two or more recommenders
  • Reduction by predicates
  • Adding and removing sub-matrices / tag types

• Application to Progressive learning

  • Incremental recommender enhancement
  • Why it works?
  • Evaluation
  • Examples

• Time series search engine

  • The general approach
  • The special interactive interface
  • Finding trends
  • Finding nearest neighbours

• Making an item-user recommender

  • Bi-partite graph based
  • Pure collaborative filtering
  • Examples

• Composite recommenders

  • The Composite design pattern
  • Normalization and weighted voting

• Tags significance determination and Learning to rank

• Conversational agents for generating recommender workflows

• Summary and Q&A

Visual aids

Here is a diagram for the recommender workflows:

---

This call of ToRecommenderWorkflowCode over a sequence of natural commands:

ToRecommenderWorkflowCode( 
    "create from dfTitanic; 
     apply the LSI functions inverse document frequency, term frequency, and cosine;
     recommend by profile female->3, 30->0.1; 
     extend recommendations with dfTitanic; 
     show pipeline value" )

generates this WL code:

SMRMonUnit[] ⟹ 
SMRMonCreate[dfTitanic] ⟹
SMRMonApplyTermWeightFunctions["GlobalWeightFunction" -> "IDF", "LocalWeightFunction" -> "TermFrequency", "NormalizerFunction" -> "Cosine"] ⟹
SMRMonRecommendByProfile[<|"female"->3, "30"->0.1|>] ⟹
SMRMonJoinAcross[dfTitanic] ⟹
SMRMonEchoValue[]

References

[AA1] Anton Antonov, DSL::English::RecommenderWorkflows Raku package, (2020), GitHub/antononcube.

[AA2] Anton Antonov, Monadic Sparse Matrix Recommender Mathematica package, (2018), MathematicaForPrediction at GitHub.

[AA3] Anton Antonov, Sparse Matrix Recommender Monad in R, (2019), R-packages at GitHub.