Sequential Models

Sequential Models#

Models that use ordered interaction history for next-item prediction.

Production models (CI-tested)#

Model

Import

Tutorial

SASRec

from corerec.engines.sasrec import SASRec

SASRec

HSTU (generative)

from corerec.engines import HSTU

below

Not sequential

SAR (Smart Adaptive Recommendations) is item-similarity collaborative filtering, not a sequential model. See Matrix Factorization.

SASRec / BERT4Rec (interaction matrix)#

Sequential production models require a user×item interaction matrix (not raw rating triplets alone):

from corerec.engines.sasrec import SASRec
import numpy as np

user_list = sorted(train_df["user_id"].unique())
item_list = sorted(train_df["movie_id"].unique())
# build train_mat[user_idx, item_idx] = 1.0 for observed interactions

model = SASRec(
    hidden_units=64,
    num_blocks=2,
    num_heads=2,
    max_seq_length=50,
    num_epochs=10,
    batch_size=256,
    verbose=True,
)
model.fit(user_list, item_list, train_mat)
recs = model.recommend(user_id=1, top_k=10)

Sandbox models (experimental)#

Model

Import

Tutorial

RBM

corerec.sandbox.collaborative_full.rbm

RBM

RLRMC

sandbox sequential

RLRMC

SLi-Rec

corerec.sandbox.collaborative_full.sli

SLiRec

SUM

corerec.sandbox.collaborative_full.sum

SUM

NextItNet

sandbox sequential_model_base

NextItNet

Caser

sandbox nn_base

Caser

When to use#

  • Session-based or next-item prediction

  • User behavior is strongly order-dependent

  • E-commerce / streaming click sequences

See also#

HSTU: generative next-item recommendation#

HSTU is the encoder from Zhai et al., Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024). It reads each user’s history in time order and is trained like a language model, predicting the next item at every position. Unlike SASRec it gates attention with SiLU instead of softmax, multiplies the attention output by a learned gate, and adds learned biases for position distance and, when timestamps are given, for the time between two interactions.

It takes one row per interaction, like every other model, plus optional timestamps:

from corerec.engines import HSTU

model = HSTU(epochs=50)                      # paper recipe: 50-dim, 2 layers, 200-item history
model.fit(user_ids, item_ids, timestamps=timestamps)
model.recommend(user_id, top_k=10)           # history items are excluded by default
model.save("artifacts/hstu")

Training uses the paper’s public ML-1M recipe: sampled softmax over 128 random negatives, temperature 0.05, L2-normalised embeddings. One difference: the negatives are shared by all positions of a sequence instead of drawn per position. Every position still sees 128 uniform random negatives, only correlated across positions, and it is about six times cheaper on a CPU. HSTU(encoder="sasrec") trains a SASRec encoder with the identical recipe, for comparisons where only the architecture changes.

corerec serve events.csv --model HSTU passes the file’s timestamp column to the model automatically. The ML-1M comparison against SASRec is in BENCHMARKS.md and reproduces with Findings/bench/generative_bench.py.

Contents