Claprec: Machine Learning in Practice (4/6)
Series Roadmap
- Product Overview (1/6)
- Architecture: N-Tier Design, Decoupling, and Microservices (2/6)
- Frontend: Angular, NgRx, and OpenAPI Integration (3/6)
- Machine Learning in Practice (4/6)
- Database Design and Data Modeling (5/6)
- Engineering Tradeoffs: Limited Time vs. Ideal Architecture (6/6)
Building the Recommendation Engine: A Hybrid Approach to Personalization in Claprec
When building Claprec, the primary challenge wasn’t just storing reviews; it was surfacing them. The landing page is the gateway to the platform, and a generic “latest posts” feed simply doesn’t cut it when you’re trying to build user retention. A system was needed that could handle the cold start problem, respect geographical relevance, and leverage user behavior – specifically, how much time a user spends consuming content.
In this post, I’m going to break down the architecture of the recommendation algorithm that powers the main feed. It’s a hybrid system combining collaborative filtering via Matrix Factorization, geo-spatial heuristics, and other factors.
The Architecture: Strategies for Different Contexts
One of the first decisions I made was to avoid a “one size fits all” approach. The backend logic needs to pivot based on the client’s state. Specifically, are we dealing with an authenticated user with or without location data, or an anonymous user?
The entry point handles these three distinct scenarios:
- Anonymous & No Location: The system defaults to a popularity contest, using recency as a fallback metric.
- Authenticated User: The architecture deploys a personalized ML-driven feed that prioritizes interests, followed by popularity and recency.
- Authenticated User & Location: The logic applies a personalized ML-driven feed that prioritizes geospatial factors, followed by interests, popularity, and recency.
// HigherCrudServiceReview.cs
public async Task<PaginatedResponse<DtoRelatedReview>> Recommended(BaseSearchDto search, UriBuilder uriBuilder, double? lat, double? lng, User? user)
{
// ...
return await PaginatedResultWithAttachedProperties(...);
}
Handling Geo-Spatial Complexity
Before we even get to the Machine Learning, we have to solve the location problem. In Claprec, a review isn’t just a point on a map. It can be attached to a specific business address, but it can also be attached to a product that is sold across multiple locations.
This required a robust resolution strategy. We don’t just check a single column; we traverse a hierarchy. We aggregate coordinates from the specific BusinessAddress, but if the review targets a product, we fetch all business addresses stocking that product. Furthermore, we resolve user locations not just from GPS pings (lat/lng), but also from Zipcode and City-level data provided in their profile.
// HigherCrudServiceReview.cs
// ...
var reviewsLatitudesLongitudes = await PairReviewsWithTheirUniqueLatsAndLngs(reviews);
var latitudesAndLongitudesOfUser = await PairLatitudesAndLongitudesOfUser(user!.Id);
// ...
// HigherCrudServiceReview.cs
private double? CalculateClosestProximity(double lat, double lng, HashSet<(double, double)> latitudesLongitudes)
{
if (latitudesLongitudes.Count == 0)
{
return null;
}
var absLat = Math.Abs(lat);
var absLng = Math.Abs(lng);
var proximity = double.MaxValue;
foreach (var latitudeLongitude in latitudesLongitudes)
{
var latDiffAbs = Math.Abs(absLat - Math.Abs(latitudeLongitude.Item1));
var lngDiffAbs = Math.Abs(absLng - Math.Abs(latitudeLongitude.Item2));
var proximityOfTuple = Math.Abs(latDiffAbs - lngDiffAbs);
if (proximityOfTuple < proximity)
{
proximity = proximityOfTuple;
}
}
return proximity;
}
This logic ensures that when we calculate proximity, we are matching a user’s “area of influence” against a review’s “area of availability“.
The ML Core: Matrix Factorization on Implicit Feedback
For authenticated users, simple sorting isn’t enough. We need personalization. I opted for a collaborative filtering approach using Matrix Factorization, but with a twist: we don’t have explicit ratings (1-5 stars) for the act of viewing a review. Instead, we rely on Implicit Feedback – specifically, “Time Spent“.
When a user stares at a review for 10, 30, or 60 seconds, that is a signal. We treat this duration as the label for the interaction.
Model Training Pipeline
The model is built using Microsoft.ML. The pipeline is relatively standard but effective. We map the UserId and ReviewId to keys and feed the Label (time spent in seconds) into a MatrixFactorizationTrainer.
// ReviewModel.cs
var estimator = mlContext.Transforms.Conversion.MapValueToKey("UserIdEncoded", "UserId")
.Append(mlContext.Transforms.Conversion.MapValueToKey("ReviewIdEncoded", "ReviewId"));
var options = new MatrixFactorizationTrainer.Options
{
MatrixColumnIndexColumnName = "UserIdEncoded",
MatrixRowIndexColumnName = "ReviewIdEncoded",
LabelColumnName = "Label",
NumberOfIterations = 20,
ApproximationRank = 100
};
We set the ApproximationRank to 100 and run for 20 iterations. This balances latency and the ability to capture complex user-item interactions. Although implicit feedback data can be noisy, we train the model to predict how long a user would spend on a review they haven’t seen yet, relying on the algorithm to find signal amidst the noise.
Operationalizing the Model: Retraining Strategy
A model is only as good as its data. Retraining a model on every single interaction is a resource drain. To solve this, I implemented a threshold-based retraining trigger.
We maintain a counter for specific activity types (staring at reviews). The model retrains only when the count of new activities hits a specific threshold (e.g., every x new interaction signals from a user). This ensures the model adapts to new trends without crippling the database with constant training jobs.
// HigherCrudServiceActivity.cs
public async Task<DtoPlainActivity> Insert(InsertDtoActivity insert, User user)
{
// ...
if (insert.ActivityTypeId is
(short) EnumActivityType.STARING_AT_REVIEW_FOR_10_SECONDS or
(short) EnumActivityType.STARING_AT_REVIEW_FOR_30_SECONDS or
(short) EnumActivityType.STARING_AT_REVIEW_FOR_60_SECONDS
)
{
await TrainReviewModelIfThresholdAchieved(user);
}
}
private async Task TrainReviewModelIfThresholdAchieved(User user)
{
var reviewActivitiesOfUserCount = await GetQueryable()
.Where(a => a.UserId == user.Id &&
(a.ActivityTypeId == (short) EnumActivityType.STARING_AT_REVIEW_FOR_10_SECONDS ||
a.ActivityTypeId == (short) EnumActivityType.STARING_AT_REVIEW_FOR_30_SECONDS ||
a.ActivityTypeId == (short) EnumActivityType.STARING_AT_REVIEW_FOR_60_SECONDS)
)
.LongCountAsync();
if (reviewActivitiesOfUserCount > 1 && reviewActivitiesOfUserCount % THRESHOLD_FOR_REVIEW_MODEL_TRAINING != 0)
{
return;
}
var reviewActivities = await GetQueryable()
.Where(a => a.ActivityTypeId == (short) EnumActivityType.STARING_AT_REVIEW_FOR_10_SECONDS ||
a.ActivityTypeId == (short) EnumActivityType.STARING_AT_REVIEW_FOR_30_SECONDS ||
a.ActivityTypeId == (short) EnumActivityType.STARING_AT_REVIEW_FOR_60_SECONDS)
.ToListAsync();
ReviewModel.OrchestrateReviewRecommenderModel(reviewActivities, _webHostEnvironment, reviewActivities.LongCount());
}
The “Buffer” Optimization & Popularity Dynamics
One of the more interesting engineering challenges was pagination. Standard pagination (Skip(x).Take(y)) breaks down when you have a volatile sorting order influenced by ML predictions and volatile proximity.
To solve this, I implemented a Buffer Strategy. Instead of querying the exact page size from the database, we fetch a larger “buffer” (e.g., 100 records) based on a stable initial sort (creation date). We then apply the heavy lifting – ML predictions, proximity calculations, and category matching – in-memory on this buffer.
// HigherCrudServiceReview.cs
// ...
var paginationCreator = new PaginationCreator<BaseSearchDto>(search, uriBuilder, totalRecords);
var bufferPage = paginationCreator.CalcSkippingRecordsCount() / REVIEW_BUFFER_SIZE;
query = query.Skip(bufferPage * REVIEW_BUFFER_SIZE).Take(REVIEW_BUFFER_SIZE);
var reviews = await query.ToListAsync();
// ...
Crucially, this buffer approach mitigates a common pitfall of popularity-based ranking: the “rich get richer” feedback loop. If we ranked the entire database solely by total time spent, legacy reviews would dominate the feed indefinitely, burying new content.
By anchoring the buffer to CreatedAt, we enforce a sliding window of relevance. As new reviews are created, they enter the buffer, and older reviews are naturally pushed out. This ensures that “popularity” is only a deciding factor among recent content, guaranteeing freshness while still surfacing the best of what’s new. It keeps the feed from stagnating and ensures that popularity serves as a quality signal for new content, rather than a permanent high ground for the old.
The Ranking Hierarchy
Finally, how do we merge all these signals? For a logged-in user, the sorting priority is strict:
- Prediction Score: A personalized prediction of how long the user will spend on the review.
- Proximity: How close is the user to the review’s location?
- Category Matching: Does the review match the user’s explicit interests?
- Global Popularity: Total time spent by all users on this review.
- Recency: Fallback to creation date.
Conclusion
Building the recommendation engine for Claprec was an exercise in balancing complexity with performance. By offloading geo-spatial resolution to helper methods, utilizing a buffer for pagination, and implementing an implicit-feedback Matrix Factorization model, I’ve created a system is created that feels instant yet deeply personalized.
It’s a constant iteration – tuning the weights, adjusting the retraining threshold, and refining how we define “time spent“. But for now, it provides a solid technical foundation for the user experience at claprec.dev.
What’s Next?
While the recommendation engine serves as the “brain” of the platform, it relies on a robust data skeleton to function. In the next post, Claprec: Database Design – Modeling the Data (5/6), I’ll peel back the layers on the storage architecture. I will share the full ER diagram for the project and detail the Entity Framework Core configurations that bring it to life. We’ll look at how I modeled the relational database to empower the ML components – designing a schema that ensures data integrity and provides the flexibility needed to feed our algorithms effectively.
I’d love to hear your thoughts on this approach to recommendation systems or answer any questions you might have in the comments below.
