From afd046379517fb38153d326f74e2c8bc6df59691 Mon Sep 17 00:00:00 2001
From: spirinamayya <90619187+spirinamayya@users.noreply.github.com>
Date: Fri, 13 Sep 2024 15:09:55 +0300
Subject: [PATCH] SasRec tutorial (#186)
Added sasrec tutorial
---
examples/tutorials/sasrec_tutorial.ipynb | 1203 ++++++++++++++++++++++
1 file changed, 1203 insertions(+)
create mode 100644 examples/tutorials/sasrec_tutorial.ipynb
diff --git a/examples/tutorials/sasrec_tutorial.ipynb b/examples/tutorials/sasrec_tutorial.ipynb
new file mode 100644
index 00000000..8f8dc929
--- /dev/null
+++ b/examples/tutorials/sasrec_tutorial.ipynb
@@ -0,0 +1,1203 @@
+{
+ "cells": [
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "# SASRec model"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "**Table of Contents**\n",
+ "\n",
+ "* Prepare data\n",
+ "* Model description\n",
+ "* Recommendations\n",
+ "* RecTools implementation\n",
+ " * Additional details\n",
+ "* Model application\n",
+ " * Additional details\n",
+ "* Under the hood: Dataset processing\n",
+ "* Under the hood: Transformer layers\n",
+ "\n",
+ "\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 1,
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "import numpy as np\n",
+ "import os\n",
+ "import pandas as pd\n",
+ "import torch\n",
+ "\n",
+ "from lightning_fabric import seed_everything\n",
+ "from pathlib import Path\n",
+ "\n",
+ "from rectools import Columns\n",
+ "from rectools.dataset import Dataset\n",
+ "from rectools.models.sasrec import SasRecModel\n",
+ "\n",
+ "# Enable deterministic behaviour with CUDA >= 10.2\n",
+ "os.environ[\"CUBLAS_WORKSPACE_CONFIG\"] = \":4096:8\""
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "# Prepare data"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "We are using KION dataset for this tutorial. The data was gathered from the users of MTS KION video streaming platform. To make recommendations only user-item interactions are required, as SASRec implementation does not support user and item features."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 2,
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Archive: data_en.zip\n",
+ " inflating: data_en/items_en.csv \n",
+ " inflating: __MACOSX/data_en/._items_en.csv \n",
+ " inflating: data_en/interactions.csv \n",
+ " inflating: __MACOSX/data_en/._interactions.csv \n",
+ " inflating: data_en/users_en.csv \n",
+ " inflating: __MACOSX/data_en/._users_en.csv \n",
+ "CPU times: user 83.8 ms, sys: 44.4 ms, total: 128 ms\n",
+ "Wall time: 5.51 s\n"
+ ]
+ }
+ ],
+ "source": [
+ "%%time\n",
+ "!wget -q https://github.com/irsafilo/KION_DATASET/raw/f69775be31fa5779907cf0a92ddedb70037fb5ae/data_en.zip -O data_en.zip\n",
+ "!unzip -o data_en.zip\n",
+ "!rm data_en.zip"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 3,
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "(5476251, 5)\n"
+ ]
+ },
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " user_id | \n",
+ " item_id | \n",
+ " datetime | \n",
+ " total_dur | \n",
+ " watched_pct | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 0 | \n",
+ " 176549 | \n",
+ " 9506 | \n",
+ " 2021-05-11 | \n",
+ " 4250 | \n",
+ " 72.0 | \n",
+ "
\n",
+ " \n",
+ " | 1 | \n",
+ " 699317 | \n",
+ " 1659 | \n",
+ " 2021-05-29 | \n",
+ " 8317 | \n",
+ " 100.0 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " user_id | \n",
+ " item_id | \n",
+ " datetime | \n",
+ " weight | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 0 | \n",
+ " 176549 | \n",
+ " 9506 | \n",
+ " 2021-05-11 | \n",
+ " 3 | \n",
+ "
\n",
+ " \n",
+ " | 1 | \n",
+ " 699317 | \n",
+ " 1659 | \n",
+ " 2021-05-29 | \n",
+ " 3 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " user_id | \n",
+ " item_id | \n",
+ " datetime | \n",
+ " weight | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 0 | \n",
+ " 176549 | \n",
+ " 9506 | \n",
+ " 2021-05-11 | \n",
+ " 3 | \n",
+ "
\n",
+ " \n",
+ " | 3815 | \n",
+ " 176549 | \n",
+ " 15469 | \n",
+ " 2021-05-25 | \n",
+ " 3 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " user_id | \n",
+ " item_id | \n",
+ " score | \n",
+ " rank | \n",
+ " title_orig | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 0 | \n",
+ " 176549 | \n",
+ " 11749 | \n",
+ " 2.834625 | \n",
+ " 1 | \n",
+ " Incredibles 2 | \n",
+ "
\n",
+ " \n",
+ " | 1 | \n",
+ " 176549 | \n",
+ " 7310 | \n",
+ " 2.768771 | \n",
+ " 2 | \n",
+ " Despicable Me 2 | \n",
+ "
\n",
+ " \n",
+ " | 2 | \n",
+ " 176549 | \n",
+ " 15266 | \n",
+ " 2.686491 | \n",
+ " 3 | \n",
+ " Monsters University | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " user_id | \n",
+ " item_id | \n",
+ " datetime | \n",
+ " weight | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 2493287 | \n",
+ " 324373 | \n",
+ " 10440 | \n",
+ " 2021-06-24 | \n",
+ " 3 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " user_id | \n",
+ " item_id | \n",
+ " score | \n",
+ " rank | \n",
+ " title_orig | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 0 | \n",
+ " 324373 | \n",
+ " 15297 | \n",
+ " 5.894632 | \n",
+ " 1 | \n",
+ " Klinika schast'ya | \n",
+ "
\n",
+ " \n",
+ " | 1 | \n",
+ " 324373 | \n",
+ " 9728 | \n",
+ " 4.081971 | \n",
+ " 2 | \n",
+ " Wrath of Man | \n",
+ "
\n",
+ " \n",
+ " | 2 | \n",
+ " 324373 | \n",
+ " 13865 | \n",
+ " 4.080128 | \n",
+ " 3 | \n",
+ " V2. Escape from Hell | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " user_id | \n",
+ " item_id | \n",
+ " datetime | \n",
+ " weight | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 2393877 | \n",
+ " 14630 | \n",
+ " 8871 | \n",
+ " 2021-03-28 | \n",
+ " 3 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " user_id | \n",
+ " score | \n",
+ " rank | \n",
+ " item_id | \n",
+ " title_orig | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ "
\n",
+ "
\n",
+ " \n",
+ " | user_id | \n",
+ " item_id | \n",
+ " weight | \n",
+ " datetime | \n",
+ "
\n",
+ " \n",
+ " | u1 | \n",
+ " i1 | \n",
+ " 0.1 | \n",
+ " 2021-09-09 | \n",
+ "
\n",
+ " \n",
+ " | u2 | \n",
+ " i1 | \n",
+ " 0.3 | \n",
+ " 2021-09-09 | \n",
+ "
\n",
+ " \n",
+ " | u2 | \n",
+ " i3 | \n",
+ " 0.2 | \n",
+ " 2021-09-05 | \n",
+ "
\n",
+ " \n",
+ " | u1 | \n",
+ " i2 | \n",
+ " 0.3 | \n",
+ " 2021-09-07 | \n",
+ "
\n",
+ " \n",
+ " | u3 | \n",
+ " i2 | \n",
+ " 0.4 | \n",
+ " 2021-09-05 | \n",
+ "
\n",
+ " \n",
+ " | u1 | \n",
+ " i3 | \n",
+ " 0.5 | \n",
+ " 2021-09-08 | \n",
+ "
\n",
+ "
"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "1. Filter out users with less than 2 interactions in train dataset. The model uses shifted user interactions to make next item prediction, thus at least 2 items should be in the history. \n",
+ "\n",
+ "2. Leave `session_maxlen` most recent interactions for each user.\n",
+ "\n",
+ "After first 2 steps, some users and/or items may be filtered out from train dataset. However, as it will be shown further, it is still possible to make recommendations for a previously unmet user, if interaction is known.\n",
+ "\n",
+ "3. Create user sessions: for each user specify items with which there was an interaction in the order from earliest to most recent. Sessions for example dataset are the following:\n",
+ "$$S^1 = (i2, i3, i1)$$\n",
+ "$$S^2 = (i3, i1)$$\n",
+ "\n",
+ "4. Before train stage each session is divided into train and target. As the task is to predict next item, shifted sequence is considered as target.\n",
+ "$$S^1_{train} = (i2, i3), S^1_{target} = (i3, i1)$$\n",
+ "$$S^2_{train} = (i3), S^2_{target} = (i1)$$\n",
+ "5. Both train and target sequences are adjusted to have user-defined `session_maxlen`:\n",
+ " * If session is longer than `session_maxlen`, cut earliest items\n",
+ " * If session is shorter than `session_maxlen`, pad earliest items with PAD element\n",
+ "$$S^1_{train} = (PAD, PAD, PAD, i2, i3), S^1_{target} = (PAD, PAD, PAD, i3, i1)$$\n",
+ "$$S^2_{train} = (PAD, PAD, PAD, PAD, i3), S^2_{target} = (PAD, PAD, PAD, PAD, i1)$$"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "# Under the hood: Transformer layers"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "* Multi-head attention layer receives queries after layer normalisarion, keys and values without it. Masked attention is required to forbid model consider future interactions: cannot access element t+2 when predicting element t+1. Following notation from original article: \n",
+ "$$ \\text{Attention}(Q, K, V) = \\text{softmax} (\\frac {QK^T}{\\sqrt{d}})V $$\n",
+ "$$S = SA(\\hat{E}) = \\text{Attention} (\\hat{E}W^Q, \\hat{E}W^K, \\hat{E}W^V)$$\n",
+ "\n",
+ "where $\\hat{E}$ - input embedding\n",
+ "* Point-wise feed-forward network has the following structure: $F_i = \\text{FFN}(S_i) = \\text{ReLU}(S_i \\cdot W^{(1)} + b^{(1)}) \\cdot W^{(2)} + b^{(2)}$,\n",
+ "\n",
+ "where $S_i, S_j$ - items of user sequence\n",
+ "\n",
+ "$W_1, W_2$ - weights\n",
+ "\n",
+ "$b_1, b_2$ - biases\n",
+ "* To avoid overfitting and stabelize training process, 2 residual connections are applied adding data after layer normalization.\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "metadata": {},
+ "outputs": [],
+ "source": []
+ }
+ ],
+ "metadata": {
+ "kernelspec": {
+ "display_name": "Python 3 (ipykernel)",
+ "language": "python",
+ "name": "python3"
+ },
+ "language_info": {
+ "codemirror_mode": {
+ "name": "ipython",
+ "version": 3
+ },
+ "file_extension": ".py",
+ "mimetype": "text/x-python",
+ "name": "python",
+ "nbconvert_exporter": "python",
+ "pygments_lexer": "ipython3",
+ "version": "3.8.2"
+ }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}