Recency-Frequency based recommendation scoring
Project description
RFscorer
rfscorer is a Python package for Recency-Frequency based recommendation scoring.
Note: In this package, RF stands for Recency-Frequency, not Random Forest.
It estimates a recommendation score (product-choice probability) — how likely each previously interacted item is to be chosen — from two behavioral signals:
- Recency: time since the last interaction. More recently viewed items tend to attract more interest.
- Frequency: number of interactions. More frequently viewed items tend to attract more interest.
You can set any event as the prediction target (revisits, purchases, conversions, etc.).
Product recommendation comes down to deciding which previously interacted items to prioritize. For example:
- [By frequency] Item A viewed once 🆚 Item B viewed twice ▶ recommend B, viewed more often
- [By recency] Item A viewed 1 day ago 🆚 Item C viewed 2 days ago ▶ recommend A, viewed more recently
- [Trade-off] Item A viewed once 1 day ago 🆚 Item D viewed twice 2 days ago ▶ hard to judge by intuition
For such non-trivial cases, rfscorer uses mathematical optimization to estimate scores that capture the recency–frequency interaction while satisfying their natural monotonicity, giving a data-driven, natural recommendation order over previously interacted items.
📄 Based on the paper: Jiro Iwanaga, Naoki Nishimura, Noriyoshi Sukegawa, and Yuichi Takano, “Estimating product-choice probabilities from recency and frequency of page views,” Knowledge-Based Systems, Vol. 99, 2016, pp. 157–167. [paper]
rfscorer's scores also serve as input to downstream models — a rating matrix for collaborative filtering or features for ML models. These features capture the recency–frequency interaction and can help improve downstream accuracy (paper).
Recommendation system architecture using RF scoring
Slides
Overview slides (PDF) (CC BY 4.0)
Features
| Feature | Description |
|---|---|
| scikit-learn style | fit() / transform() interface |
| Minimal data | works with any behavior history with three columns: user, item, datetime |
| Explainable | scores are estimated by mathematical optimization under RF monotonicity, making the reasoning behind each recommendation easy to explain |
| Stable probability estimation | product-choice probabilities are estimated directly from recency and frequency, avoiding the instability of converting ML model outputs to a probability scale |
| Downstream use | usable not only as a standalone recommendation score but also as a rating matrix for collaborative filtering or as features for ML models |
| Rich diagnostics & visualization | extensive statistical outputs and visualization features let practitioners explain results in their work and researchers report them in papers |
Installation
pip install rfscorer
Usage
Below is a minimal example of building a model and scoring recommendations from a behavior history (for complete, runnable code, see Examples).
Minimal Example
import pandas as pd
from rfscorer import RecencyFrequencyScorer, split_by_date
# Load your behavior history
df = ... # columns: user, item, datetime
# Split by target date
target_date = "2026-07-07"
df_obs, df_gt = split_by_date(df, target_date, 7, 1) # observation: 7 days, ground truth: 1 day
# Fit and optimize
scorer = RecencyFrequencyScorer()
scorer.fit(df_obs, df_gt)
scorer.optimize(kind="mono")
# Score recommendations (on test data)
df_test = ... # test data (columns: user, item, datetime)
df_test_obs, _ = split_by_date(df_test, target_date, 7, 1)
df_scores = scorer.transform(df_test_obs, target_date, kind="mono")
| 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 |
Recommend items to each user from highest to lowest probability. Since scores are probabilities, expected value calculations are straightforward (e.g., expected revenue per recommendation). Use the order column to apply business rules (e.g., recommend the top 2 items per user).
Visualization: Comparing Optimization Approaches
The package supports many optimization approaches. Here we visualize three representative methods:
scorer.plot_probability_surface(kind="emp") # empirical (raw rates)
scorer.optimize(kind="mono") # RF monotonicity
scorer.plot_probability_surface(kind="mono")
scorer.optimize(kind="mcc") # convex in R, concave in F
scorer.plot_probability_surface(kind="mcc")
| Empirical | Monotonicity | Monotonicity-Convex-Concave |
Each surface clearly captures how the product-choice probability behaves with respect to recency (time since last interaction) and frequency (number of interactions):
- Empirical: Raw probabilities without constraints. Noisy and may violate monotonicity, sometimes recommending items in unnatural order.
- Monotonicity: Probabilities with RF monotonicity constraints. Guarantees items are recommended in natural order.
- Monotonicity-Convex-Concave: Probabilities with RF monotonicity and convexity-concavity constraints. Produces the smoothest surface.
Examples
- examples/tutorial_beginner_en.ipynb — end-to-end walkthrough: load data, fit, optimize, visualize, transform, and evaluate
- examples/tutorial_practical_en.ipynb — practical workflow: chronological train/test split, build the various models, compare accuracy, and save/load the model
- examples/tutorial_advanced_fit_rolling_en.ipynb — advanced workflow: time-series rolling training with
fit_rolling()to stabilize empirical probabilities across multiple reference dates
For the complete list of tutorials, see examples/.
References
Show references
Citation
Show citation & BibTeX
If you use rfscorer in academic work, you can cite it as follows in the body of your paper:
We used
rfscorer(Iwanaga et al., 2016), a Python library for Recency-Frequency-based recommendation scoring.¹
The full reference is:
@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 also use the probability matrix as input to a collaborative filtering model or as ML features, please also cite:
@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
RFscorer (日本語README)
rfscorer は、最新度(Recency)と頻度(Frequency)に基づいて、ユーザーが過去に接触した商品の推薦スコア(商品選択確率)を推定する Python パッケージです。
本パッケージでは RF は Recency-Frequency(最新度-頻度) を表します
ユーザーの行動履歴から、各商品の 選択されやすさ を表す推薦スコア(商品選択確率)を推定します。スコアは、次の2つの行動シグナルに基づいて計算されます。
- 最新度(recency):最後に接触してからの経過時間。最近接触した商品ほど関心が高い傾向。
- 頻度(frequency):商品への接触回数。何度も接触した商品ほど関心が高い傾向。
予測対象は、再閲覧、購買、コンバージョンなど用途に応じて自由に設定できます。
商品推薦は、過去の行動履歴を用いて、ユーザーが過去に接触した商品の中でどれを優先して推薦するかを決める問題と捉えられます。 たとえば、次のような比較です。
- 【頻度で判断】 1回閲覧した商品A 🆚 2回閲覧した商品B ▶ たくさん閲覧した 商品B を推薦
- 【最新度で判断】 1日前に閲覧した商品A 🆚 2日前に閲覧した商品C ▶ 最近閲覧した 商品A を推薦
- 【トレードオフ】 1日前に1回閲覧した商品A 🆚 2日前に2回閲覧した商品D ▶ 直感では判断できない
rfscorer は、このような非自明な比較に対して、数理最適化により、最新度と頻度の交互作用を考慮しつつ、自然な単調性を満たす推薦スコアを推定します。
これにより、ユーザーが過去に接触した商品に対して、データに基づく自然な推薦順位を与えることができます。
📄 Based on the paper: Jiro Iwanaga, Naoki Nishimura, Noriyoshi Sukegawa, and Yuichi Takano, “Estimating product-choice probabilities from recency and frequency of page views,” Knowledge-Based Systems, Vol. 99, 2016, pp. 157–167. [論文]
また、rfscorer が出力する商品選択確率は、下流のモデルの入力(協調フィルタリングの評価値行列・機械学習モデルの特徴量)としても有効です。最新度と頻度の交互作用が反映された有用な特徴量によるモデル構築ができます(論文)。
RFスコアリングを用いた推薦システム構成
紹介スライド
rfscorer 紹介スライド(PDF・日本語)(CC BY 4.0)
パッケージの特徴
| 特徴 | 説明 |
|---|---|
| scikit‑learn ライク | fit() / transform() によるインターフェースを提供 |
| 最小限のデータ要件 | 入力データは、user、item、datetime の3カラムをもつ行動履歴 |
| 説明可能性 | 数理最適化によりRF単調性を満たすスコアを推定するため、推薦理由を説明しやすい |
| 安定した確率推定 | 最新度と頻度から商品選択確率を直接推定するため、機械学習モデルの出力を確率スケールへ変換する際の不安定さを回避できる |
| 下流モデルへの活用 | 単独の推薦スコアとしてだけでなく、協調フィルタリングの評価値行列や機械学習モデルの特徴量としても利用可能 |
| 豊富な診断と可視化 | 各種統計量の出力や可視化機能が充実。実務家は業務で説明しやすく、研究者は分析結果を論文に記載しやすい |
インストール
pip install rfscorer
使い方
以下は、行動履歴からモデル構築と推薦スコア(商品選択確率)算出までを行う最小限の例です(実行可能な完全版はサンプルを参照)。
最小限の例
import pandas as pd
from rfscorer import RecencyFrequencyScorer, split_by_date
# 行動履歴の読み込み
df = ... # カラム: user, item, datetime
# 基準日で観測データ・正解データに分割
target_date = "2026-07-07"
df_obs, df_gt = split_by_date(df, target_date, 7, 1) # 観測データ7日間・正解データ1日間
# モデル構築と最適化
scorer = RecencyFrequencyScorer()
scorer.fit(df_obs, df_gt)
scorer.optimize(kind="mono")
# 推薦スコアを算出(テストデータ)
df_test = ... # テストデータ(カラム: user, item, datetime)
df_test_obs, _ = split_by_date(df_test, target_date, 7, 1)
df_scores = scorer.transform(df_test_obs, target_date, kind="mono")
| 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 |
各ユーザーに対して、商品選択確率(probability )の高い順に商品を推薦します。推薦スコアが確率値であるため、期待値計算(例:推薦結果に対する期待収益の計算)が容易です。order 列を使えば、業務ルール(例:「各ユーザーに上位2個の商品を推薦する」)を簡単に実装できます。
可視化:最適化手法の比較
本パッケージは多くの最適化アプローチをサポートしています。ここでは代表的な3つの手法を可視化します。
scorer.plot_probability_surface(kind="emp") # empirical (raw rates)
scorer.optimize(kind="mono") # RF monotonicity
scorer.plot_probability_surface(kind="mono")
scorer.optimize(kind="mcc") # convex in R, concave in F
scorer.plot_probability_surface(kind="mcc")
| Empirical | Monotonicity | Monotonicity-Convex-Concave |
各グラフは、最新度(recency)(ユーザーが商品に接触してからの経過時間)と 頻度(frequency)(接触回数)に基づく商品選択確率の特徴をよく表しています:
- Empirical(経験確率): 制約を課していない商品選択確率。ノイズにより単調性を満たさないため不自然な順序で商品を推薦する場合がある。
- Monotonicity(単調性): RF単調性制約を課した商品選択確率。商品を自然な順序で推薦することを保証する。
- Monotonicity-Convex-Concave(単調性+凸凹): RF単調性制約と凹凸性制約を課した商品選択確率。最も滑らかなグラフを生成する
サンプル
- examples/tutorial_beginner_ja.ipynb — 初級編では、最小限の利用方法を紹介します。データロード、モデル構築・最適化・可視化、推薦スコア算出、精度評価までのコードを紹介します。
- examples/tutorial_practical_ja.ipynb — 実践編では、主要機能を紹介します。具体的には、時系列での訓練・テスト分割、各種モデル構築と精度比較、モデルの保存・ロードを紹介します。
- examples/tutorial_advanced_fit_rolling_ja.ipynb — 応用編では、
fit_rolling()を用いたローリング集計を扱います。複数の基準日にわたって集計することで経験的商品選択確率を安定させる方法を紹介します。
全チュートリアルの一覧は examples/ を参照してください。
参考文献
引用について
引用方法と BibTeX を表示
学術論文等で rfscorer を利用する場合は、論文の引用と本Githubへのリンクを脚注を加え、本文中で以下のように引用できます:
We used
rfscorer(Iwanaga et al., 2016), a Python library for Recency-Frequency-based recommendation scoring.¹
参考文献とBibTexは以下のとおりです:
@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}
}
さらに、商品選択確率行列を協調フィルタリングモデルの入力として利用する場合や機械学習の特徴量として利用する場合には、以下の文献も併せて引用してください:
@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}
}
ライセンス
MIT License
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file rfscorer-0.5.2.tar.gz.
File metadata
- Download URL: rfscorer-0.5.2.tar.gz
- Upload date:
- Size: 6.5 MB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: uv/0.8.4
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
361c04c2d89ad54417d45ec3542b3cac8b318a9c431d66a22b2ba071f9c10f3e
|
|
| MD5 |
889f10948ef91685a3e1085765de2006
|
|
| BLAKE2b-256 |
7aecc9b491d0718d48217a44bc1a764da7314e0078b3f7fe6ba409f4a5e7b7d9
|
File details
Details for the file rfscorer-0.5.2-py3-none-any.whl.
File metadata
- Download URL: rfscorer-0.5.2-py3-none-any.whl
- Upload date:
- Size: 36.9 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: uv/0.8.4
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
bfe02786847dab1a77c7956e16d41969506ef0286d8935ab5da058a783f6fc0c
|
|
| MD5 |
90cef9ddf8237132b49bc37a9a9234a0
|
|
| BLAKE2b-256 |
2cdbf53c2767abd073559f305aae944b0928ee041a9155627d40ddfbd7971b86
|