Building a Recommendation System for the Cooper Hewitt Design Museum

 

In 2014 the Cooper Hewitt Design Museum reopened its doors after a 3 year renovation that included not only a renovation of the building itself but a reconceptualization of the museum experience. The implementation of The Pen, issued to each visitor, enables the visitor to scan objects that they like or inspire them, allows the visitor to participate in the immersion room and process lab, both spaces in which a visitor can draw or sketch out their own designs as they are inspired by the collection around them. Additionally, the visitor can take their experience beyond the museum walls by accessing their visit online after they have left.  This interactive participation is in stark contrast to the typical passive experience; a one-way conversation of knowledge transfer from the museum to the visitor through signs or ‘tombstones’. In my opinion, the integration of technology and data into the visitor’s experience has repositioned the cultural institution into a space of conversation and active participation between the artwork and patrons.

Furthermore, data collected from The Pen provides some exciting opportunities for analysis and inquiry:

• What departments have the most objects?
• Where do most of the objects in the collection originate from?
• What objects are scanned the most?
• What types of objects are scanned the most?
• Does an objects location in the museum influence if it gets scanned?
• Can we predict areas of interest of a visitor or their profession? Would this information help inform curators for future exhibitions?

The API developed by the museum is a phenomenal opportunity for anyone interested in the museum to access their collection, exhibitions, cafe openings, graphic color schemes, or integrate museum information in apps.

Play with the recommendation app prototype.

Learn more about the Cooper Hewitt Museum API.

Explore the Cooper Hewitt Museum github repository. So much to play…

See code on github.

Reference Documents:

• recommenderlab documentation
• Developing and Testing Algorithms for 0-1 Data using recommenderlab…

The Data

The data was collected from the Cooper Hewitt’s github repository as a snapshot of the collection items that have been digitally archived. This did not include all of the items in their collection as they are constantly updating the repository. Two CSV files were downloaded and imported as data frames into R:

Objects data frame:

• 194,316 total objects in the collection
• 38 features: ID, country of origin, tags, type, medium, museum department, dimensions, description, primary image url, historic period, accession number, gallery text, on_loan, year acquired, etc.

Visitor Pen Data:

• 3,559,735 scanned items between August 16, 2014 to February 15, 2016
• 584 days / 1.5 years of pen data
• 111,467 unique visits
• 7,784 unique objects out of 194,316 total objects scanned

There were some challenges with the object data frame features; namely tag, medium, type. I had hoped to use these features to do basic EDA to understand the distribution of collection based on medium or type; i.e. to understand if the collection was predominantly architectural, graphic design or textiles, etc. Unfortunately these features were not useful in analysis because they were not distinct categories but rather the multitude of unique values too vast to be useful. For instance, the feature type contained 6 different variation for a category glass of which their distance values were differentiated by capital and lowercase letters.

Of the useful and distinct features Department and WOE Country (country of origin) basic EDA reveals the distribution of objects mostly in Drawing, Prints and Graphic Design with most of the objects originating from the US, France, Germany, and Italy. With a bit of knowledge of design history this makes sense with the influence of the Italian Renaissance, German Bauhaus movement, the French Beaux-Arts, and contemporary design of the the Americas.

The visitor pen data consisted of a data frame with two features: bundle ID (unique visit ID) and object ID.  If a visitor scanned more than one item each item (or scan) was recorded as an observation along with its corresponding bundle ID.

Basic EDA with pen data also reveals that on average visitors scan approximately three to four items per visit.

Collaborative filtering + Association rules

Collaborative filtering and association rules were researched in implementing the recommendation system with collaborative filtering being the method utilized in the shiny app prototype.  Recommenderlab, the package used to apply collaborative filtering, provides an infrastructure to test and develop recommender algorithms which includes association rule-based algorithms. There has been extensive research into recommendation systems that use large data sets of user elicited rating data (i.e. ratings of 1 to 5 stars) while limited research exists where no large amount of directly elicited rating data or sparse, binary data is available. The methods tested for creating a recommendation system for the Cooper Hewitt Museum were:

• Popular, recommendations made based on popular items
• Association Rules, recommendations made on relationships found amongst items in the collection
• User-Based Collaborative Filtering, recommendations made based on what other similar users liked
• Item-Based Collaborative Filtering, building off of association rules recommendations are made on relationships found between items.

User-Based Collaborative Filtering is a memory based algorithm that makes recommendations based on patterns of similar users, or neighbors, that are identified by employing k nearest neighbor algorithm. A sparse binary matrix is constructed with columns representing unique items, rows representing each user, and populated with 1 if the user ‘liked’ the object or 0 if otherwise. All columns are summed up, the k nearest neighbors are identified (highlighted in blue), items liked by the active user are removed from the recommendations (red ‘x’) and recommendations for the active user (highlighted in yellow) are made of the remaining items in which their sum is greater than 0.  A detraction of this method is its computational cost.  All data must be read into local memory and similarities calculated between all users. My attempts at training a User-based CF model were unsuccessful as all session were timed out.