Recotem

English

日本語

English

日本語

Appearance

AWS Personalize Alternative (Open-Source, Self-Hosted) ​

Amazon Personalize is a fully managed recommendation service on AWS. It is a capable product, and for many teams the managed model is exactly right. But some teams go looking for an AWS Personalize alternative — usually for one of these reasons:

  • Cost and cost predictability. Personalize bills on consumption (data ingestion, training, inference requests, and per-hour recommender capacity), and real-time campaigns reserve throughput that bills even when idle. Teams with steady, self-owned infrastructure sometimes prefer a fixed compute cost they already pay for.
  • Vendor lock-in. The trained model lives inside AWS. You cannot export the solution version and run it elsewhere, and the recommendation API is AWS-specific — so the recommender is tied to the platform.
  • Data residency. Personalize requires interaction data to be ingested into AWS (S3 and Personalize datasets, in an AWS region). Some organizations must keep training data on their own infrastructure or in a specific jurisdiction.
  • Self-hosting. Some teams simply want to run the whole pipeline themselves — on their own cloud account, on-prem, or in an air-gapped environment.

Recotem is an open-source (Apache-2.0), self-hosted recommender that covers this ground. One YAML recipe defines the data source, the training, and the served endpoint. You run recotem train to produce a signed model artifact, and recotem serve to expose it over HTTP. This page compares the two fairly and shows how to move Personalize concepts across.

Why teams evaluate an alternative ​

None of the reasons above make Personalize a poor product — they are trade-offs. If you want a managed service and are already all-in on AWS, Personalize removes a lot of operational work. The rest of this page is for teams whose constraints (cost model, data residency, portability, or a preference for self-hosting) point toward an open-source option they run themselves.

Recotem vs AWS Personalize ​

The table below compares the two on the dimensions that usually drive the decision. Pricing and feature specifics for AWS are as of 2026-07 — always confirm against the current AWS Personalize pricing page.

DimensionAWS PersonalizeRecotem
HostingFully managed AWS service. No servers to run.Self-hosted. You run it on any cloud, on-prem, or in Docker/Kubernetes.
Data locationInteraction data is ingested into AWS (S3 + Personalize datasets, in an AWS region).Data stays in your environment. Recotem reads CSV/Parquet, BigQuery, or SQL; artifacts land in your own storage.
Pricing modelConsumption-based: data ingestion (~$0.05/GB), training, inference (per-1,000 requests), and per-hour recommender capacity by user count. Real-time campaigns reserve a minimum throughput that bills even at zero traffic. (as of 2026-07)Free and open-source (Apache-2.0). You pay only for the compute and storage you already run. No per-request or per-hour license fee.
AlgorithmsAWS proprietary "recipes": User-Personalization / -v2 (transformer/HRNN), Similar-Items, SIMS, Personalized-Ranking, Popularity-Count, and more.Open-source irspack algorithms: IALS, CosineKNN, RP3beta, DenseSLIM, TruncatedSVD, BPRFM, TopPop — with automatic Optuna hyperparameter search over the ones you list.
APIAWS SDK / GetRecommendations and GetPersonalizedRanking, authenticated with AWS SigV4.Your own FastAPI service: POST /v1/recipes/{name}:recommend (plus :recommend-related and :batch-recommend), authenticated with an X-API-Key header. See the Serving API.
Ops burdenLow. AWS handles training infrastructure, serving, scaling, and retraining orchestration.You own it. You operate the train host and the serve host, manage signing keys, and scale the service — in exchange for full control and portability.

That last row is the honest core of the trade-off: Personalize trades money and lock-in for operational simplicity; Recotem trades operational ownership for control, portability, and a fixed cost you already pay.

How Recotem maps to Personalize concepts ​

If you know Personalize, the mental model transfers directly. Recotem collapses several Personalize resources into one recipe file plus one artifact.

AWS PersonalizeRecotem equivalent
Interactions dataset + dataset import jobRecipe source block (CSV/Parquet on S3/GCS, BigQuery, or SQL query)
Schema (Avro)Recipe schema block (user_column, item_column, time_column)
Recipe + solution (algorithm + parameters)training.algorithms list + Optuna search that picks the best model
Solution version (trained model)Signed .recotem artifact written by recotem train
Campaign (deployed solution version)recotem serve endpoint POST /v1/recipes/{name}:recommend
Similar-Items / SIMSPOST /v1/recipes/{name}:recommend-related
Batch inference jobPOST /v1/recipes/{name}:batch-recommend

The key difference: in Personalize you choose a single recipe per solution; in Recotem you list several algorithms and let the built-in Optuna search train and evaluate each, then keep the best by your chosen metric (ndcg, map, recall, or hit). One Recotem recipe therefore plays the role of a Personalize dataset, solution, and campaign at once.

A migration sketch ​

Suppose you have a Personalize User-Personalization solution trained on an Interactions dataset of USER_ID, ITEM_ID, TIMESTAMP. Here is the shape of the equivalent Recotem recipe. You already have the interaction data — point the source at wherever it lives (an S3 CSV export, your warehouse, or a database):

yaml
name: user_personalization

source:
  type: csv
  path: s3://my-bucket/interactions.csv.gz
  dtype:
    user_id: str
    item_id: str

schema:
  user_column: user_id     # ← Personalize USER_ID
  item_column: item_id     # ← Personalize ITEM_ID
  time_column: ts          # ← Personalize TIMESTAMP

cleansing:
  dedup: keep_last
  min_rows: 5000

training:
  algorithms: [IALS, RP3beta, TopPop]   # implicit-feedback CF + a popularity fallback
  metric: ndcg
  cutoff: 20
  n_trials: 40
  split:
    scheme: time_user
    heldout_ratio: 0.1

output:
  path: s3://my-bucket/artifacts/user_personalization.recotem

Then the two commands that replace "create solution version" and "create campaign":

bash
recotem train recipe.yaml          # → signs an artifact (the "solution version")
recotem serve --recipes ./recipes/ # → serves it (the "campaign")

Finally, swap your client call. Where you called GetRecommendations, call the self-hosted endpoint instead:

bash
curl -s -X POST http://localhost:8080/v1/recipes/user_personalization:recommend \
  -H "X-API-Key: <plaintext>" \
  -H "Content-Type: application/json" \
  -d '{"user_id": "u1", "limit": 10}' | jq .

The response is an ordered list of item_id / score objects, optionally enriched with item metadata — the same shape of result you consumed from Personalize. For "customers who bought this" style item-to-item recommendations (the Personalize Similar-Items / SIMS use case), call :recommend-related with seed_items instead.

Your IDs stay your IDs

Recotem keys recommendations on the raw user_id and item_id strings from your source data, so you do not need to maintain a separate ID mapping the way you might around Personalize datasets.

When AWS Personalize may fit better ​

An honest comparison has to say where the managed service wins. Choose AWS Personalize over Recotem when:

  • You want zero operational ownership. Personalize runs training, serving, scaling, and retraining for you. Recotem is a service you operate — if you have no capacity to run one, the managed model is worth the money.
  • You are deeply integrated with AWS and value native ties to S3, EventBridge, and IAM, plus elastic scaling without capacity planning.
  • You need real-time / streaming updates. Personalize can ingest live events and update recommendations continuously. Recotem follows a train-then-hot-swap model: you retrain on a schedule and the server atomically reloads the new artifact. It is not an online-learning system.
  • You need managed features Recotem does not offer, such as Next-Best-Action, user segmentation, contextual recommendations, item exploration for cold-start, or automatic incremental retraining.

Recotem's sweet spot is the opposite set of constraints: you have interaction data you can already query, you want the model and data to stay on your own infrastructure, and you would rather run a small, well-defined service than pay per request. If that is you, an open-source self-hosted recommender is a strong fit.

Next steps ​