Recommendation.jl: Modeling User-Item Interactions in Julia

Presentation at JuliaCon 2022 @ Online https://pretalx.com/juliacon-2022/talk/VWVY9S/

Repository: https://github.com/takuti/Recommendation.jl

Previous JuliaCon 2019 presentation: https://www.youtube.com/watch?v=kC8LKQ_YjyM

Transcript

1. Takuya Kitazawa
   Recommendation.jl
   Modeling User-Item Interactions in Julia
   

   View Slide

2. julia> using Pkg; Pkg.add(“Recommendation”
   )
   
   github.com/takuti/Recommendation.jl
   

   View Slide

3. Recommender Systems
   Input data undergoes a series of vector/matrix computations
   item
   user
   value
   Rating, purchase,
   click, watch, …
   Price, category, …
   Demographics, …
   Event Matrix
   Recommender
   Top-k recommendations
   …
   

   View Slide

4. Recommender Systems
   Lifecycle that Recommendation.jl makes it easier to follow
   Data Pre-process Build Evaluate Post-process
   How to build recommender
   Intro to recommender algorithms
   Why Julia
   • Vector/matrix-friendly syntax
   
   • High simplicity and modularity
   
   • E
   ff i
   cient numerical computing
   
   • Recommender is not only about ML
   
   • Ecosystem for math tools matters
   

   View Slide

5. Infinite = Union{AbstractFloat, Integer}
   mutable struct Even
   t
   
   user::Integer
   item::Integer
   value::Infinite # e.g. rating, 0/1
   end
   struct DataAccessor
   events::Array{Event,1}
   R::AbstractMatri
   x
   
   user_attributes::Dict{Integer,Any} # user => attributes
   item_attributes::Dict{Integer,Any} # item => attributes
   # ...
   end
   data = DataAccessor(...)
   recommender = MostPopular(data)
   fit!(recommender)
   recommend(...)
   

   View Slide

6. • Change in interface & variable type

   e.g., build!(…) → fit!(…)
   
   
   • New algorithms: Factorization machines, 

   Matrix factorization with BPR loss
   
   • Additional logic for data validation & 

   missing data handling
   • Data downloader:

   MovieLens, LastFM,

   Amazon review
   
   • Synthetic data
   generator
   2019 → 2022
   Improving usability toward “v1.0.0”
   Data Pre-process Build Evaluate Post-process
   • Support non-accuracy metrics:

   Serendipity, novelty, diversity, coverage
   

   View Slide

7. using Recommendatio
   n
   
   data = load_movielens_100k()
   all_items = collect(keys(data.item_attributes))
   fm = FactorizationMachines(data)
   fit!(fm, learning_rate=0.3, max_iter=100)
   recommendations = recommend(fm, user, topk, all_items)
   # => list of (item, score)
   measure(Coverage(), map(first, recommendations), catalog=all_items)
   Data downloader
   New algorithm & interface
   Non-accuracy evaluation
   

   View Slide

8. Accuracy vs. Non-Accuracy Metrics
   What is the de
   fi
   nition of “good” recommender?
   Coverag
   e
   
   AggregatedDiversit
   y
   
   GiniInde
   x
   
   IntraListSimilarit
   y
   
   Novelt
   y
   
   Serendipit
   y
   
   ShannonEntropy
   AU
   C
   
   MA
   E
   
   MA
   P
   
   MP
   R
   
   NDC
   G
   
   Precisio
   n
   
   RMS
   E
   
   Recal
   l
   
   ReciprocalRank
   split_dat
   a
   
   cross_validatio
   n
   
   measur
   e
   
   evaluate
   

   View Slide

9. Recommender
   Other Tools
   How OSS communities design recommender systems
   Language Data modeling and linear algebra Characteristics
   MyMediaLite C# Built-in arithmetic operators with
   fi
   le IOs
   Simplicity and transparency of
   basic recommendation techniques
   LibRec Java
   Custom interfaces (e.g., dense/sparse
   matrices) & built-in data structures
   Wide range of algorithms
   implemented from scratch
   LensKit / LightFM Python NumPy/SciPy & Cython
   Rapid development and wider use
   cases in the Python community
   Recommendation.jl Julia Built-in vector/matrix representations
   Take full advantage of built-in
   o
   ff
   ering for simplicity and e
   ff
   i
   ciency
   Data model
   Algorithm
   Interface
   Metrics
   Utils
   

   View Slide

10. More details to be available in proceedings
    julia> using Pkg; Pkg.add(“Recommendation”
    )
    
    github.com/takuti/Recommendation.jl
    

    View Slide