FAQ
How much data do I need to get started?
Shaped requires, user, item and event data to power the underlying machine-learning algorithms that understand your data. The more unique interactions and user and item metadata you have, the better Shaped will perform. However, Shaped works in very low data regimes because we pre-train our models on external datasets. In some cases you can get started with less than 1000 events. Likely you have more data than you think!
How many items do I need to get started with Shaped?
The more items you have the harder it is for your users to find what they want and the more beneficial personalized discovery is for your application. We've seen that 100 items is enough to get reasonable uplift by using Shaped.
How do I know what personalization use-case is going to be best to start with?
We've found that the best way to get started is to start with the most obvious use-case for your application. For example, if you're a news site, you might want to start with personalized news recommendations. If you're a social network, you might want to start with personalized feed recommendations. If you're an e-commerce site, you might want to start with personalized product recommendations. When you work with us, we'll help you identify the use-case that's going to generate the most uplift.
My personalization use-case isn't listed, can Shaped help with it?
Yes, please reach out, we're continuing to add support for any discovery or personalization use-case and would love to hear what you're trying to do.
Does Shaped work for anonymous and/or semi-persistent users?
Yes, we can personalize to anonymous traffic by using contextual information like location, device, time of day, etc. We can use semi-persistent user information from cookies and we use real-time session based ranking to personalize to the users first few clicks.
I want recent items to be shown more often than older items, how can I make sure Shaped respects this?
Yes, if you provide a created_at or updated_at field within your item metadata, we can use it as signal to rank more recent items higher. You can also put in hard filter limits using metadata filtering or the Create Model filter queries.
I want social posts to be shown more than trending posts, how can I make sure Shaped respects this?
Please contact us about this. We have some private beta features around multi-objective optimization that can help you control your final ranking based on factors like social affinity vs trending signal.
How do I optimize for both the consumers/purchasers and producers/sellers for my use-case?
Please contact us about this. We have some private beta features around multi-objective optimization that can help you balance all sides of your use-case.
I'm not sure if i'm logging the correct data for ranking, how do I know if Shaped will work?
Please contact us about this. We can help you identify the best events to use for your use-case and help you get started.
My data is spread-out across many data sources, how can I integrate Shaped?
Yes, take a look at our guide on using multiple connectors to integrate Shaped with your heterogeneous data sources.
How can I perform an A/B test with Shaped?
Please contact us about this. We can provide best practices and tooling for running A/B tests to compare recommendations from Shaped to your existing recommendation solution.
Can I weight different events/interaction more than others (e.g. value model)?
Yes! Event labels must be a numerical value, where <= 0 is negative and > 0 is positive. The magnitude of these values convey the confidence of the label either in the positive or negative direction. You can experiment with different values for each label type to work out what works best for your use-case and goals.
How often do you retrain my model?
By default we retrain your model every 24 hours but it can be adjusted up to every few hours depending on your use-case.
What performance uplift should I expect from Shaped?
We've seen up to a 15% uplift in conversion rates and a 25% uplift in engagement rates when using Shaped.
How can I keep my API key secure for client side applications?
For now, we recommend that you use a server side proxy to make requests to Shaped. We're working on a solution to make this easier.
My data isn't 'tagged' or is very messy, will Shaped still work?
Yes, we can work with messy and unstructured data. We use unstructured data encoders to extract the things that matter from your data.
Does Shaped work for my unstructured image, video and language data?
Yes, we use unstructured data encoders for image, video and language data, which can be helpful if you have few interactions.
What are the best events/interactions to optimize for?
The events that correlate with your business goals are the best events to optimize for. For example, if you're optimizing for revenue, you might want to optimize for purchases or add-to-cart. If you're optimizing for engagement, you might want to optimize for clicks.
What are your scale limits and request queries per second?
Our scale limits can be found in the 'Why Shaped' page. Our Rank API endpoints handle 1000s of requests per second.
How do you handle timestamp data?
We have timestamp encoders that bin your timestamps into different time periods and use cyclic encodings to capture seasonal patterns.
How do you handle location data?
We have location encoders that bin your locations into different regions.
How do I paginate ranked results?
We re-rank and filter after every interaction so Shaped's ranking API can be called
each time if you need a new page of results. Just set the limit
flags to your
page/slate size.
How should I evaluate the performance of Shaped?
There's two primary ways customers quantitatively evaluate recommendation systems:
- Offline evaluation, involves computing the accuracy of predicting historic interactions from a held out set.
- Online evaluation, involves evaluating your engagement or revenue uplift from a live experiment.
We do 1. for you, and we can help you do 2., just let us know!
Do you share my data with other customers?
No we use external datasets to pre-train our models, your data is never shared with other customers.
I have some promotional content that I want interleaved with my rankings, how can I do this with Shaped?
Please contact us about this. We've worked with a few customers to do this.
I already have search but want to make the search results personalized, how can I use Shaped for this?
Yes, you can just use Shaped's re-ranker to score and rank your search results. Take a
look at the item_ids
argument in our Rank API.