Please enable JavaScript.

Coggle requires JavaScript to display documents.

Recommendation Systems (Properties of RS (User preference, Prediction…

- - - - Long tail
- - - - offline preprocesing or model learning
      - at runtime learned model used for predictions
      - model updated/ retarined periodicaly
      - Types
        
        Association rule minnig
        
        Create model from rules
        
        aimd at detection of rules such as "if customer purchases beer then he also buys diapers in 70 percent of th cases
        
        Probabilistic methods
        
        RF-Rec
        
        RF
        
        Rating frequencies
        
        Graph based methods
        
        Spreading activation
      - Algorithms
        
        Item based KNN
    - - Does not scale for most real world scenarios
        
        Large e commerce sites / social networks have tens of millions of customers and millions of items
      - Everything is calculated at the time of the request
      - Algoritms
        
        user based KNN
  - - - CB
      - Wisdom of the crowd
      - Algorithms
        
        KNN
      - Pros
        
        Well understood
        
        Work well in some domain
        
        No knowledge engineering required
      - Cons
        
        requires user community
        
        No explanations
        
        sparsity problems
        
        No integration of other knowledge sources
    - - Types
        
        Prefiltering
        
        Post filtering
        
        Contextual modeling
        
        Multiverse recommendation
        
        Tensors for contexxt
      - Context
        
        Exmaples
        
        Winter vs summer
        
        gift vs yourself
        
        girlfriend in movie theater vs home with group of friend
        
        Alone vs kids
        
        Happy vs sad
        
        Problem
        
        I like ferrari but i will not buy any because of my money context
        
        I will not listen hard core music on Christamm even if Ilike it
        
        After 2 hours of listening pop music I like to change
        
        When I am alone I listen different type of music than with my family
        
        any infromation or conditions that can influence the perception of the usefulness of an item for user
        
        Types
        
        Physical
        
        time
        
        position
        
        weather
        
        light
        
        temperature
        
        Social
        
        preference and role people around
        
        Interaction media
        
        device used to access
        
        type of media that are browsed
        
        Modal
        
        State of mind
    - - User requirements
        
        example
        
        maximum accepted price
        
        Value ranges
      - Dependencies
        
        User requiremnets and Product properties
      - Content based + Metrics
      - Types
        
        Constraint based
        
        Cased based
        
        Both
    - - What we need?
        
        Content
        
        Text content
        
        Vector space model
        
        Representations
        
        Term-Frequency -Inverse Document Frequency
        
        2 more items...
        
        Inverse document frequency
        
        1 more item...
        
        Term frequency
        
        2 more items...
        
        Document frequency
        
        2 more items...
        
        Improving methods
        
        Use steaming
        
        1 more item...
        
        Size cut-offs
        
        1 more item...
        
        Remove stop words
        
        1 more item...
        
        Use lexical knowledge
        
        1 more item...
        
        Detection of phrases as terms
        
        1 more item...
        
        Limitations
        
        Semantic meaning remains uknowkn
        
        1 more item...
        
        Item properties
        
        Genre
        
        Author
        
        Price
        
        Preferences
        
        What user likes
      - Domena
        
        For recommending text documents
        
        Multimedia content
        
        Fashion
        
        Music
        
        E commerce
      - Algoritms
        
        item based KNN
        
        Rocchio s method
        
        Documents are weighted by relevant or irelevant
        
        Cons
        
        Ceratin number of ratings needed
        
        User interaction needed during retriveal phase
        
        Explicit decision models
        
        Pros
        
        Work best for small numbers of features
        
        Decision trees can be build automatickly
      - Cons
        
        Keywords alone may not be suifficient to judge quality/relavance of a document or web page
        
        Ramp-up phase required
        
        Some training data is still required
        
        Overspecialization
        
        More of the same
        
        Too similans news item
        
        Mutlicrtierial optimization
        
        Diversity
        
        Novelty
      - Pros
        
        Do not require user comunity
- - - - Examples
        
        New users
        
        New items
      - Problem solving
        
        Straightforward approchaes
        
        Use another method
        
        Content based
        
        Demographic
        
        Non personalized
        
        Default voting
        
        Assign default values to items that only one of two users to be compared has rated
        
        Ask / force users to rate a set of items
      - Alternatives
        
        Use better algorithms (beyond nearest neybor approaches)
        
        Examles
        
        Assume "transitivity" of neighbor hoods
- - - - Item based
      - User based
        
        Memory based algortihm
  - - - BPR
      - Používáme dvojice
        
        Výsledkem je větší menší
  - - - CNN
      - Mutlimedia content
    - - Word2Vec
- - - - Properties
        
        need to simulate user behaviors
        
        low cost
        
        short time
        
        answer only a few questins
        
        Snadno se opakuje
        
        no interaction with real users
      - Evaluation
        
        Simulation on existing dataset
        
        Train / Validation / Test split
        
        Prediction of "correct" object
        
        According to some metric
    - - test subjects
      - expensive
        
        time
        
        money
      - small scale
      - Těžko se opakuje
      - Subjekty nelze recyklovat
      - Lze testovat různě změny např GUI
    - - Vlastnosti
        
        risky - can loose some customers
        
        good to do after an offline testing showes good result
        
        Těžko se opakuje
        
        redirect small part of trafic to an alternative recommendation engine
      - Evaluation
        
        Proxy variables
        
        Examples
        
        Click through rate
        
        Basket size
        
        Conversions rate
        
        Loyalty of user
        
        Share
        
        Follows
        
        Target variables
        
        direct effect can be to small
        
        hard to measure directly
        
        Cascade of evaluation metrics
        
        From hight to low
        
        Detectability changes
        
        From low to high
        
        Impact on your true target
        
        Evaluation metrics
        
        Response time
        
        Train/retrain model time
        
        Recall on object
        
        Is portion of your objects ignored?
        
        Are there too many low profit bestseellers?
        
        Ability to predict
        
        Can calculate recomendation for all users
        
        Memory / CPU consumption
- - - - Clicks
      - Page views
      - Time spent on some page
    - - Can be interpreted incorectly
        
        Example
        
        User dont like this item he buy it for someone else
    - - Collected constantly
      - Dont reqiure any additional effort from user
  - - - Ratings
        
        Scales
        
        Likes or dislikes
        
        Different domains addopted differennt scales
        
        1 to 7
        
        1 to 5
      - Multidimensional ratings
    - - Users not always willing to rate many items
- - - - Testování oproti existujícím datům
    - - A/B testování
      - Testování na testovacích uživatelích
  - - - Explicit
        
        Examples
        
        Like/Dislike
        
        1 to 5
        
        1 to 7
        
        1 to 10
        
        Rating upon mutliple dimension
        
        Movie domen
        
        Sound
        
        Action
        
        Cons
        
        Users dont want to make to many ratings
      - Implicit
        
        Examples
        
        Clicks
        
        Page views
        
        item buy
        
        Pros
        
        Required constantly
        
        Cons
        
        Not sure if correctly interpreted
        
        Buy item for someone else
- - - - User model
      - Items
    - - Relevance score
  - - - Personalized + Comunity data
      - Algorithms
        
        User based nearest neighbor
        
        Item based neares neigbor
    - - User profile + Product features
    - - User profile + Product features + Knowledge models
    - - User profile + Community + Product features + Knowledge models
    - - I know your age and your hobbies
    - - location
      - Opened page
- - - - Nový uživatelé
      - Nové objekty
  - - - Hlavně pro memory based algoritmy