Nick Emmons


At Upshot, we believe a robust interpretation layer for NFT data is key to the success and growth of the NFT industry. Keeping this in mind, we built collection recommendations: A feature that enables users to better discover NFT collections based on their current interests and holdings.

This article will cover why collection recommendations are important for the NFT space as a whole, how we used AI to put personalized curation front and center for our users, and of course, how you can get started using them today.  

[Try it: Collection Recommendations is live on our aggregator](https://app.upshot.xyz "Upshot App")
![Upshot Web App with Collection Recommendations feature](//images.ctfassets.net/lhkcghnr8ql4/0ZW23Boca6kirS9hdq9W4/030b3ebbf9f4fee19b55506cbfbe6aa7/Screen_Shot_2023-02-28_at_6.53.57_AM.png)

## Why Recommend

The NFT space is exploding. Amongst many things, this has led to an ever-increasing number of collections in various categories such as PFPs, gaming, music, generative art, etc.

With more than 160K collections (at the time of writing) available through various marketplaces, it has become increasingly more challenging for individuals to decide which collection to focus on. 

Analogous to the evolution of the e-commerce industry, the NFT space will soon realize the importance of content curation and personalization for individual wallets. To create this will need to develop data-driven techniques that reduce the search space for wallets and help individuals navigate the sheer volume of content produced in the NFT space.

### For the User
A recommendation system uses algorithms that take a large set of items as input to determine what to display to a user next based on those inputs. Recommendations are part of our everyday life - think of Facebook News Feed, YouTube's homepage, and Netflix's "Recommended for you" navigation page.  At Upshot, our recommendation algorithm can infer “what the user may like to own” based on the collections they currently hold in their wallet. 

### For Developers
Using our collection recommendations algorithm (which is bundled into the [Upshot API](https://docs.upshot.xyz/docs)) developers can create a better, more personalized experience for their end users. 

This personalization can lead to new products, enhanced user stickiness and transaction volume, and an overall better experience in the NFT space - one that cuts through the noise and presents the relevant information on a specific per-user basis.

## Our Methodology

We based our initial approach on inferring the wallet's preferences for collections based on what the wallet already owns. 

For baseline recommendations, collaborative filtering is the first step. This step doesn't require too much information about collections except for which collections specific wallets have traded, owned, viewed, etc. 

Collaborative filtering analyzes the relationships between wallets and interdependencies among collections to identify new wallet-collection associations. The technique recommends collections to wallets based on the preferences of other wallets by identifying similar wallets or collections. 

In our case, we use NFT ownership as a wallet's preference. A critical step in the recommendation process for a given wallet is the identification of similar wallets (known as user-based) or similar collections (known as item-based).

At Upshot, our recommendation technique is item-based where similarities between collections are leveraged to determine the associations between wallets and collections. This item-based technique is explained using Figure 1 which shows wallets ownerships of collections.

![collaborative filtering figure](//images.ctfassets.net/lhkcghnr8ql4/7zDQO0delNgh4CI49G11Qi/c79f69078e745358f228d44b0b5b2bdb/collaborative-filtering.png)
Figure 1: Item-based Collaborative Filtering

To recommend more collections to wallet 1, we first find the collections owned by wallet 1 which are collection 1 and collection 2. Now, the next step is to find similar collections to collection 1 and collection 2. As collection 2 is also owned by wallet 3 who also owns collection 3 and it is not owned by wallet 1, the process will present collection 3 to wallet 1 as one of the recommendations.

From a technical perspective, the similarity between collections is determined using Cosine Similarity, which calculates the similarity between two vectors by calculating the cosine angle between them. The advantage of using this method is that it is invariant to vector size as in our dataset, the number of collections each wallet interacts with varies a lot. This difference in magnitude does not affect the similarity measure since it only considers the angle between vectors instead of their sizes. 

Mathematically, cosine similarity is calculated as the dot-product of collection pairs divided by the product of normalized vectors. i.e.

$$
cos(\theta) = \frac{\sum_iA_iB_i}{\sqrt{\sum_iA_i^2} \sqrt{\sum_iB_i^2}}
$$

Once a collection-collection similarity matrix is available, the next step is to calculate a score between the wallet and every collection not currently owned by the wallet. The association between a wallet and a collection is the weighted average of a set of collections that are similar to collections owned by the wallet. Notice we use weighted average where weights are the similarity score between collections. The weighted average reduces the bias in situations where similar collections are not equally similar to the owned collection.

The final step in the recommendation process is ranking the wallet-collection associations in the descending order and determining the top 5 recommendations per wallet. Here, we apply a recency factor based on seven day volume in order to generate recommendations that are not just personalized to the wallet but also are relevant in the current period.
Our complete process of generating recommendations can be visually represented as a sequential pipeline of multiple stages as:
![Recommendation build flow](//images.ctfassets.net/lhkcghnr8ql4/5EGiB0BcHIcPmROPmiEysA/cf0f039bab9c87dc46e4a878190ffce6/image2.png)

## Conclusion
In conclusion, a reliable recommendation engine is important for the growth of the NFT industry and the marketplaces that operate within it. As more content producers enter the NFT space, it is becoming increasingly challenging for users to identify relevant collections that match their interests. This makes it hard for marketplaces to retain the attention and transactions of their users. 

Finally, at Upshot, we’re constantly iterating. This recommendation system will continue to improve with time. For example, we plan to add auxiliary information about wallets and collections, such as collection category, number of assets in the collection, asset traits, trait rarity, etc., to greatly enhance the recommendations. Furthermore, from the wallet’s perspective, we’re looking into ways to handle the  “cold-start problem” where we can still make accurate recommendations to a wallet that has not owned anything yet. 

As the NFT space continues to evolve, it is essential to have a tool that can keep up with the industry's changing landscape. Upshot's collection recommendations are an excellent step towards achieving this goal.


A reliable recommendation engine is important for the growth of the NFT industry and the marketplaces that operate within it. So, we built one. 

Introducing Collection Recommendations

About Upshot

Related Posts