Recommender systems (Part 1) – Knowledge-based recommenders

5 minute read

Updated: November 11, 2020

The CHIP Project - bridging the gap between virtual and physical museum experience

Rijksmuseum Amsterdam offers
- 7000 artworks in the museum
- 50000 artworks online
They have broad pool of users – schools, students, normal people have everything required for recommender system

Vision: Personal museum guide in your pocket

They decided to work together with the scientists and was looking at how people go through this massive
Need to come up with a vision when creating the project – what will you add to the user experience
- Came out with the idea of personal museum guide – When you go to a museum you can have a personal guide and the guide creates a narrower museum
  - - But doesn’t scale and not tailored to various types of users
- Take the metaphor of having a personal guide and develop a mobile guide
What will a guide be? What do they do? Features of guides
- Personal guide knows about the artworks and paintings – has the domain knowledge
- Ability to identify what user might be interested with, from limited interactions
- Can recommend paintings and explain based on what they know about the world and the user

CHIP tools

Knowledge Graph: Semantically Enriched Museum Data

Behind this recommender system – it is knowledge based.
First component before developing personalised app, is the knowledge
Snippet of the knowledge model – computer scientist together with curators worked together to identify the core knowledge required for painting rec
They used concepts – graph-based models – and concepts are linked with relationships and relationships are named
- Creator has place – birthplace and deathplace
Have implicit relationships - Once you have the knowledge base, you can infer
- Can infer style of the author, painting etc bc there’s a connection
Powerful knowledge model that tells about the world but this is NOT the recommender. Now need to bring this to the recommender

How can you bring this external knowledge to help you to explore the paintings
Based on the metadata mapped with the external knowledge, you can bring other paintings bc they share some similarities with the same metadata
- Having the knowledge model allows you to bring the content to filter what the most relevant is to the painting
Eg
- Painting has relation to other 2 painters (teacher, student etc)
- Painting showing portrait, military scenes etc
- Exploring all different aspects by looking at one painting, due to the relationships
- Not personalised, but it enriches the space
Can talk about much smaller pool to recommend to the user. Still need to evaluate what’s relevant to the user user model comes in.
Then you need to think about mapping the content and the user model

Problem with cold start – need to know about the user and map the user to the content
How are you going to create the user profile?
- Can look at the browsing history of the user and rate the things
- Sometimes we ask explicitly to rate the paintings.
- The paintings shown to the users are not just random paintings; it is selected in a way that different area of the knowledge graph is being identified
Gives user model and we think you’d be interested in these topics – all that is the top categories under the knowledge model
- End up with showing the user a list of keywords that they are interested in
- It can tell me why! Based on which rating and why has it inferred that I might be interested in x
Cold start is solved by explicit rating ‼
Use model now has = graph of user interests + graphs with painting information - overlap

Art recommender

They are looking at similarity metrics
They have to decide if this painting is going to interest the user – how similar this is to user parameter
- How you map user model to the painting model for each of the painting
And based on that it decides how its going to recommend – it shows the relevance of the painting
- All the descriptions of the painting and some weight, to what extent it is related to the user (which user already approved)
- You can map directly the user characteristics; but you can look further - How much you are propagating with thin the graph, given the concept about the painting within the graph
Also recommends other artwork that is related to the PAININTG not the user – gives diversity but still has relevance

Artwork description

Once you have the knowledge, can automatically generate description of the painting with the metadata
Then decide within the description, what to show as important and what to suppress another level of personalisation ‼
- Knowledge model allows to do that as it identifies parameters and it amplifies the important info more
Knowledge allows:
- To describe why something has been recommended
- Allows you to diversify and to recommends other paintings
- Allow you to automatically generate description that is tailored to the user (CBF won’t allow this)

Need to get a good knowledge graph to implement the recommender
They build on content-based filtering – user profile and information about the content
But to bring in the knowledge model of the world
Content, user AND knowledge – powerful
Main advantage - reasoning!!
- You can infer things – and user tend to trust the recommendations that way
- Can diversify user’s experience
Main limitation
- We still have the cold start problem
- Need to prepare the content and mapping
Without the knowledge, can’t do the recommender! Time to get some knowledge and map it to the model