rfscorer 0.3.2

Recency-Frequency based recommendation scoring

rfscorer

rfscorer is a Python package for Recency-Frequency based recommendation scoring.

It estimates revisit probabilities — the preference score for each user-item pair, forming a matrix analogous to a rating matrix — from interaction histories, using two simple but powerful behavioral signals: recency, which captures how recently a user interacted with an item, and frequency, which captures how often the user has interacted with it.

The package is designed for product recommendation and revisit modeling, especially in settings where interpretable scoring based on interaction history is preferred over black-box recommendation models.

Note: In this package, RF stands for Recency-Frequency, not Random Forest.

Features

• scikit-learn-style API — familiar fit() / transform() interface makes it easy to integrate into existing data science workflows
• Minimal data requirements — works with any interaction log that has three columns: user, item, and datetime; no ratings or explicit feedback needed
• Explainable scoring — probabilities are derived through mathematical optimization under RF monotonicity constraints, making every score fully traceable and auditable; 3D surface visualization further supports intuitive understanding
• Probabilistic output — revisit probabilities serve as preference scores, enabling expected value calculations and probabilistic ranking of recommendations
• Extensible — the user–item probability matrix produced by transform() can be directly used as input to collaborative filtering or other downstream recommendation models

Installation

pip install rfscorer

Usage

import pandas as pd
from rfscorer import RecencyFrequencyScorer

Prepare an interaction log with three columns: user, item, and datetime. The same user-item pair may appear multiple times, representing repeat visits.

user	item	datetime
u_001	i_032	2026-07-01
u_001	i_017	2026-07-03
u_001	i_032	2026-07-05
u_002	i_011	2026-07-02
u_002	i_058	2026-07-04

Split users into training and test sets, then split each by target_date into an observation window and an evaluation window.

target_date = "2026-07-07"

df_train_obs  = df_train[df_train.datetime <= target_date]
df_train_eval = df_train[df_train.datetime >  target_date]

Call fit() to estimate empirical revisit probabilities. Recency and frequency are computed from the observation window; the evaluation window provides ground-truth revisit labels.

scorer = RecencyFrequencyScorer()
scorer.fit(df_train_obs, df_train_eval)

The empirical surface reflects raw revisit rates and may be irregular due to sparse data.

fig = scorer.plot_probability_surface(kind="emp")

Call optimize() to smooth the surface under RF monotonicity constraints using convex quadratic programming. kind="mono" enforces recency and frequency monotonicity.

scorer.optimize(kind="mono")
fig = scorer.plot_probability_surface(kind="mono")

kind="mcc" additionally adds convexity in recency and concavity in frequency, yielding a smoother surface.

scorer.optimize(kind="mcc")
fig = scorer.plot_probability_surface(kind="mcc")

Call transform() to score each user-item pair in the test observation window. It returns a DataFrame with columns user, item, recency, frequency, probability, and order (rank within each user, sorted by probability descending).

df_test_obs  = df_test[df_test.datetime <= target_date]
df_test_eval = df_test[df_test.datetime >  target_date]

df_rec = scorer.transform(df_test_obs, target_date, kind="mcc")

user	item	recency	frequency	probability	order
u_001	i_032	1	4	0.1167	1
u_001	i_017	2	3	0.0789	2
u_001	i_045	3	1	0.0248	3
u_002	i_011	1	2	0.0621	1
u_002	i_058	4	1	0.0182	2

Within each user, rows are sorted by probability descending; order represents the recommendation rank.

Call evaluate() to measure recommendation quality at each rank cutoff. It returns precision, recall, and F1 for each cutoff from 1 to order.

scorer.evaluate(df_rec, df_test_eval, order=5)

Examples

• examples/basic_usage.ipynb — end-to-end walkthrough: load data, fit, optimize, transform, and evaluate

References

• Jiro Iwanaga, Naoki Nishimura, Noriyoshi Sukegawa, and Yuichi Takano, “Estimating product-choice probabilities from recency and frequency of page views,” Knowledge-Based Systems, Volume 99, 2016, Pages 157–167.

• Jiro Iwanaga, Kyota Ishihara, Naoki Nishimura, and Ikki Tanaka, Pythonではじめる数理最適化 ―ケーススタディでモデリングのスキルを身につけよう―(in Japanese), Ohmsha, 2021.

  • Chapter 7: 商品推薦のための興味のスコアリング(in Japanese)

• Jiro Iwanaga, Naoki Nishimura, Noriyoshi Sukegawa, and Yuichi Takano, “Improving collaborative filtering recommendations by estimating user preferences from clickstream data,” Electronic Commerce Research and Applications, Volume 37, Article 100877, 2019.

Citation

If you use rfscorer in academic work, please cite the following paper:

• Jiro Iwanaga, Naoki Nishimura, Noriyoshi Sukegawa, and Yuichi Takano, “Estimating product-choice probabilities from recency and frequency of page views,” Knowledge-Based Systems, Volume 99, 2016, Pages 157–167.

@article{Iwanaga2016,
  author  = {Jiro Iwanaga and Naoki Nishimura and Noriyoshi Sukegawa and Yuichi Takano},
  title   = {Estimating product-choice probabilities from recency and frequency of page views},
  journal = {Knowledge-Based Systems},
  volume  = {99},
  pages   = {157--167},
  year    = {2016},
  url     = {https://www.sciencedirect.com/science/article/abs/pii/S0950705116000848}
}

If you additionally use the probability matrix as input to a collaborative filtering model, please also cite:

• Jiro Iwanaga, Naoki Nishimura, Noriyoshi Sukegawa, and Yuichi Takano, “Improving collaborative filtering recommendations by estimating user preferences from clickstream data,” Electronic Commerce Research and Applications, Volume 37, Article 100877, 2019.

@article{Iwanaga2019,
  author  = {Jiro Iwanaga and Naoki Nishimura and Noriyoshi Sukegawa and Yuichi Takano},
  title   = {Improving collaborative filtering recommendations by estimating user preferences from clickstream data},
  journal = {Electronic Commerce Research and Applications},
  volume  = {37},
  pages   = {100877},
  year    = {2019},
  url     = {https://www.sciencedirect.com/science/article/abs/pii/S1567422319300547}
}

License

MIT License