Self made SVD

import helpers
import pandas as pd
import numpy as np
import scipy.sparse as sparse
from collections import defaultdict
from itertools import chain
import random
import matplotlib.pyplot as plt

train_read = pd.read_csv("./Datasets/ml-100k/train.csv", sep="\t", header=None)
test_read = pd.read_csv("./Datasets/ml-100k/test.csv", sep="\t", header=None)
(
    train_ui,
    test_ui,
    user_code_id,
    user_id_code,
    item_code_id,
    item_id_code,
) = helpers.data_to_csr(train_read, test_read)

# Done similarly to https://github.com/albertauyeung/matrix-factorization-in-python
from tqdm import tqdm


class SVD:
    def __init__(self, train_ui, learning_rate, regularization, nb_factors, iterations):
        self.train_ui = train_ui
        self.uir = list(
            zip(*[train_ui.nonzero()[0], train_ui.nonzero()[1], train_ui.data])
        )

        self.learning_rate = learning_rate
        self.regularization = regularization
        self.iterations = iterations
        self.nb_users, self.nb_items = train_ui.shape
        self.nb_ratings = train_ui.nnz
        self.nb_factors = nb_factors

        self.Pu = np.random.normal(
            loc=0, scale=1.0 / self.nb_factors, size=(self.nb_users, self.nb_factors)
        )
        self.Qi = np.random.normal(
            loc=0, scale=1.0 / self.nb_factors, size=(self.nb_items, self.nb_factors)
        )

    def train(self, test_ui=None):
        if test_ui != None:
            self.test_uir = list(
                zip(*[test_ui.nonzero()[0], test_ui.nonzero()[1], test_ui.data])
            )

        self.learning_process = []
        pbar = tqdm(range(self.iterations))
        for i in pbar:
            pbar.set_description(
                f"Epoch {i} RMSE: {self.learning_process[-1][1] if i>0 else 0}. Training epoch {i+1}..."
            )
            np.random.shuffle(self.uir)
            self.sgd(self.uir)
            if test_ui == None:
                self.learning_process.append([i + 1, self.RMSE_total(self.uir)])
            else:
                self.learning_process.append(
                    [i + 1, self.RMSE_total(self.uir), self.RMSE_total(self.test_uir)]
                )

    def sgd(self, uir):

        for u, i, score in uir:
            # Computer prediction and error
            prediction = self.get_rating(u, i)
            e = score - prediction

            # Update user and item latent feature matrices
            Pu_update = self.learning_rate * (
                e * self.Qi[i] - self.regularization * self.Pu[u]
            )
            Qi_update = self.learning_rate * (
                e * self.Pu[u] - self.regularization * self.Qi[i]
            )

            self.Pu[u] += Pu_update
            self.Qi[i] += Qi_update

    def get_rating(self, u, i):
        prediction = self.Pu[u].dot(self.Qi[i].T)
        return prediction

    def RMSE_total(self, uir):
        RMSE = 0
        for u, i, score in uir:
            prediction = self.get_rating(u, i)
            RMSE += (score - prediction) ** 2
        return np.sqrt(RMSE / len(uir))

    def estimations(self):
        self.estimations = np.dot(self.Pu, self.Qi.T)

    def recommend(self, user_code_id, item_code_id, topK=10):

        top_k = defaultdict(list)
        for nb_user, user_scores in enumerate(self.estimations):

            user_rated = self.train_ui.indices[
                self.train_ui.indptr[nb_user] : self.train_ui.indptr[nb_user + 1]
            ]
            for item, score in enumerate(user_scores):
                if item not in user_rated and not np.isnan(score):
                    top_k[user_code_id[nb_user]].append((item_code_id[item], score))
        result = []
        # Let's choose k best items in the format: (user, item1, score1, item2, score2, ...)
        for uid, item_scores in top_k.items():
            item_scores.sort(key=lambda x: x[1], reverse=True)
            result.append([uid] + list(chain(*item_scores[:topK])))
        return result

    def estimate(self, user_code_id, item_code_id, test_ui):
        result = []
        for user, item in zip(*test_ui.nonzero()):
            result.append(
                [
                    user_code_id[user],
                    item_code_id[item],
                    self.estimations[user, item]
                    if not np.isnan(self.estimations[user, item])
                    else 1,
                ]
            )
        return result

model = SVD(
    train_ui, learning_rate=0.005, regularization=0.02, nb_factors=100, iterations=40
)
model.train(test_ui)

Epoch 39 RMSE: 0.7489999966900885. Training epoch 40...: 100%|██████████| 40/40 [01:02<00:00,  1.57s/it]

df = pd.DataFrame(model.learning_process).iloc[:, :2]
df.columns = ["epoch", "train_RMSE"]
plt.plot("epoch", "train_RMSE", data=df, color="blue")
plt.legend()

<matplotlib.legend.Legend at 0x7ff52c7e5100>

df = pd.DataFrame(
    model.learning_process[10:], columns=["epoch", "train_RMSE", "test_RMSE"]
)
plt.plot("epoch", "train_RMSE", data=df, color="blue")
plt.plot("epoch", "test_RMSE", data=df, color="green", linestyle="dashed")
plt.legend()

<matplotlib.legend.Legend at 0x7ff52f336dc0>

Saving and evaluating recommendations

model.estimations()

top_n = pd.DataFrame(model.recommend(user_code_id, item_code_id, topK=10))

top_n.to_csv(
    "Recommendations generated/ml-100k/Self_SVD_reco.csv", index=False, header=False
)

estimations = pd.DataFrame(model.estimate(user_code_id, item_code_id, test_ui))
estimations.to_csv(
    "Recommendations generated/ml-100k/Self_SVD_estimations.csv",
    index=False,
    header=False,
)

import evaluation_measures as ev

estimations_df = pd.read_csv(
    "Recommendations generated/ml-100k/Self_SVD_estimations.csv", header=None
)
reco = np.loadtxt("Recommendations generated/ml-100k/Self_SVD_reco.csv", delimiter=",")

ev.evaluate(
    test=pd.read_csv("./Datasets/ml-100k/test.csv", sep="\t", header=None),
    estimations_df=estimations_df,
    reco=reco,
    super_reactions=[4, 5],
)

943it [00:00, 8683.10it/s]

	RMSE	MAE	precision	recall	F_1	F_05	precision_super	recall_super	NDCG	mAP	MRR	LAUC	HR	Reco in test	Test coverage	Shannon	Gini
0	0.914393	0.717199	0.101697	0.042334	0.051787	0.068811	0.092489	0.07236	0.104839	0.04897	0.196117	0.517889	0.480382	0.867338	0.147186	3.852545	0.972694

dir_path = "Recommendations generated/ml-100k/"
super_reactions = [4, 5]
test = pd.read_csv("./Datasets/ml-100k/test.csv", sep="\t", header=None)

ev.evaluate_all(test, dir_path, super_reactions)

943it [00:00, 8505.85it/s]
943it [00:00, 9544.72it/s]
943it [00:00, 9154.80it/s]
943it [00:00, 8282.66it/s]
943it [00:00, 8432.23it/s]
943it [00:00, 9601.30it/s]
943it [00:00, 9158.89it/s]
943it [00:00, 12283.59it/s]
943it [00:00, 9500.43it/s]
943it [00:00, 10085.91it/s]
943it [00:00, 10260.90it/s]
943it [00:00, 9691.20it/s]

Model	RMSE	MAE	precision	recall	F_1	F_05	precision_super	recall_super	NDCG	mAP	MRR	LAUC	HR	Reco in test	Test coverage	Shannon	Gini
Self_TopPop	2.508258	2.217909	0.188865	0.116919	0.118732	0.141584	0.130472	0.137473	0.214651	0.111707	0.400939	0.555546	0.765642	1.000000	0.038961	3.159079	0.987317
Ready_SVD	0.950347	0.749312	0.100636	0.050514	0.055794	0.070753	0.091202	0.082734	0.114054	0.053200	0.248803	0.521983	0.517497	0.992153	0.210678	4.418683	0.952848
Self_SVD	0.914393	0.717199	0.101697	0.042334	0.051787	0.068811	0.092489	0.072360	0.104839	0.048970	0.196117	0.517889	0.480382	0.867338	0.147186	3.852545	0.972694
Ready_Baseline	0.949459	0.752487	0.091410	0.037652	0.046030	0.061286	0.079614	0.056463	0.095957	0.043178	0.198193	0.515501	0.437964	1.000000	0.033911	2.836513	0.991139
Ready_SVDBiased	0.939472	0.739816	0.085896	0.036073	0.043528	0.057643	0.077039	0.057463	0.097753	0.045546	0.219839	0.514709	0.431601	0.997455	0.168831	4.217578	0.962577
Ready_Random	1.521845	1.225949	0.047190	0.020753	0.024810	0.032269	0.029506	0.023707	0.050075	0.018728	0.121957	0.506893	0.329799	0.986532	0.184704	5.099706	0.907217
Ready_I-KNN	1.030386	0.813067	0.026087	0.006908	0.010593	0.016046	0.021137	0.009522	0.024214	0.008958	0.048068	0.499885	0.154825	0.402333	0.434343	5.133650	0.877999
Ready_I-KNNBaseline	0.935327	0.737424	0.002545	0.000755	0.001105	0.001602	0.002253	0.000930	0.003444	0.001362	0.011760	0.496724	0.021209	0.482821	0.059885	2.232578	0.994487
Ready_U-KNN	1.023495	0.807913	0.000742	0.000205	0.000305	0.000449	0.000536	0.000198	0.000845	0.000274	0.002744	0.496441	0.007423	0.602121	0.010823	2.089186	0.995706
Self_TopRated	1.030712	0.820904	0.000954	0.000188	0.000298	0.000481	0.000644	0.000223	0.001043	0.000335	0.003348	0.496433	0.009544	0.699046	0.005051	1.945910	0.995669
Self_BaselineUI	0.967585	0.762740	0.000954	0.000170	0.000278	0.000463	0.000644	0.000189	0.000752	0.000168	0.001677	0.496424	0.009544	0.600530	0.005051	1.803126	0.996380
Self_IKNN	1.018363	0.808793	0.000318	0.000108	0.000140	0.000189	0.000000	0.000000	0.000214	0.000037	0.000368	0.496391	0.003181	0.392153	0.115440	4.174741	0.965327

Embeddings

item = random.choice(list(set(train_ui.indices)))

embeddings_norm = (
    model.Qi / np.linalg.norm(model.Qi, axis=1)[:, None]
)  # we do not mean-center here
# omitting normalization also makes sense, but items with a greater magnitude will be recommended more often

similarity_scores = np.dot(embeddings_norm, embeddings_norm[item].T)
top_similar_items = pd.DataFrame(
    enumerate(similarity_scores), columns=["code", "score"]
).sort_values(by=["score"], ascending=[False])[:10]

top_similar_items["item_id"] = top_similar_items["code"].apply(
    lambda x: item_code_id[x]
)

items = pd.read_csv("./Datasets/ml-100k/movies.csv")

result = pd.merge(top_similar_items, items, left_on="item_id", right_on="id")

result

	code	score	item_id	id	title	genres
0	321	1.000000	322	322	Murder at 1600 (1997)	Mystery, Thriller
1	983	0.902748	984	984	Shadow Conspiracy (1997)	Thriller
2	985	0.894696	986	986	Turbulence (1997)	Thriller
3	778	0.890524	779	779	Drop Zone (1994)	Action
4	686	0.889220	687	687	McHale's Navy (1997)	Comedy, War
5	331	0.887596	332	332	Kiss the Girls (1997)	Crime, Drama, Thriller
6	987	0.886547	988	988	Beautician and the Beast, The (1997)	Comedy, Romance
7	1039	0.882845	1040	1040	Two if by Sea (1996)	Comedy, Romance
8	1022	0.882782	1023	1023	Fathers' Day (1997)	Comedy
9	929	0.877662	930	930	Chain Reaction (1996)	Action, Adventure, Thriller

project task 5: implement SVD on top baseline (as it is in Surprise library)

# making changes to our implementation by considering additional parameters in the gradient descent procedure
# seems to be the fastest option
# please save the output in 'Recommendations generated/ml-100k/Self_SVDBaseline_reco.csv' and
# 'Recommendations generated/ml-100k/Self_SVDBaseline_estimations.csv'

Ready-made SVD - Surprise implementation

SVD

import helpers
import surprise as sp

algo = sp.SVD(biased=False)  # to use unbiased version

helpers.ready_made(
    algo,
    reco_path="Recommendations generated/ml-100k/Ready_SVD_reco.csv",
    estimations_path="Recommendations generated/ml-100k/Ready_SVD_estimations.csv",
)

Generating predictions...
Generating top N recommendations...
Generating predictions...

SVD biased - on top baseline

algo = sp.SVD()  # default is biased=True

helpers.ready_made(
    algo,
    reco_path="Recommendations generated/ml-100k/Ready_SVDBiased_reco.csv",
    estimations_path="Recommendations generated/ml-100k/Ready_SVDBiased_estimations.csv",
)

Generating predictions...
Generating top N recommendations...
Generating predictions...

dir_path = "Recommendations generated/ml-100k/"
super_reactions = [4, 5]
test = pd.read_csv("./Datasets/ml-100k/test.csv", sep="\t", header=None)

ev.evaluate_all(test, dir_path, super_reactions)

943it [00:00, 11456.53it/s]
943it [00:00, 11932.50it/s]
943it [00:00, 10853.07it/s]
943it [00:00, 9426.44it/s]
943it [00:00, 8757.09it/s]
943it [00:00, 9999.67it/s]
943it [00:00, 11323.49it/s]
943it [00:00, 9764.72it/s]
943it [00:00, 9692.41it/s]
943it [00:00, 9052.77it/s]
943it [00:00, 8645.18it/s]
943it [00:00, 10594.54it/s]

Model	RMSE	MAE	precision	recall	F_1	F_05	precision_super	recall_super	NDCG	mAP	MRR	LAUC	HR	Reco in test	Test coverage	Shannon	Gini
Self_TopPop	2.508258	2.217909	0.188865	0.116919	0.118732	0.141584	0.130472	0.137473	0.214651	0.111707	0.400939	0.555546	0.765642	1.000000	0.038961	3.159079	0.987317
Ready_SVD	0.951652	0.750975	0.096394	0.047252	0.052870	0.067257	0.085515	0.074754	0.109578	0.051562	0.235567	0.520341	0.496288	0.995546	0.208514	4.455755	0.951624
Self_SVD	0.914393	0.717199	0.101697	0.042334	0.051787	0.068811	0.092489	0.072360	0.104839	0.048970	0.196117	0.517889	0.480382	0.867338	0.147186	3.852545	0.972694
Ready_Baseline	0.949459	0.752487	0.091410	0.037652	0.046030	0.061286	0.079614	0.056463	0.095957	0.043178	0.198193	0.515501	0.437964	1.000000	0.033911	2.836513	0.991139
Ready_SVDBiased	0.940413	0.739571	0.086002	0.035478	0.043196	0.057507	0.075751	0.053460	0.094897	0.043361	0.209124	0.514405	0.428420	0.997349	0.177489	4.212509	0.962656
Ready_Random	1.521845	1.225949	0.047190	0.020753	0.024810	0.032269	0.029506	0.023707	0.050075	0.018728	0.121957	0.506893	0.329799	0.986532	0.184704	5.099706	0.907217
Ready_I-KNN	1.030386	0.813067	0.026087	0.006908	0.010593	0.016046	0.021137	0.009522	0.024214	0.008958	0.048068	0.499885	0.154825	0.402333	0.434343	5.133650	0.877999
Ready_I-KNNBaseline	0.935327	0.737424	0.002545	0.000755	0.001105	0.001602	0.002253	0.000930	0.003444	0.001362	0.011760	0.496724	0.021209	0.482821	0.059885	2.232578	0.994487
Ready_U-KNN	1.023495	0.807913	0.000742	0.000205	0.000305	0.000449	0.000536	0.000198	0.000845	0.000274	0.002744	0.496441	0.007423	0.602121	0.010823	2.089186	0.995706
Self_TopRated	1.030712	0.820904	0.000954	0.000188	0.000298	0.000481	0.000644	0.000223	0.001043	0.000335	0.003348	0.496433	0.009544	0.699046	0.005051	1.945910	0.995669
Self_BaselineUI	0.967585	0.762740	0.000954	0.000170	0.000278	0.000463	0.000644	0.000189	0.000752	0.000168	0.001677	0.496424	0.009544	0.600530	0.005051	1.803126	0.996380
Self_IKNN	1.018363	0.808793	0.000318	0.000108	0.000140	0.000189	0.000000	0.000000	0.000214	0.000037	0.000368	0.496391	0.003181	0.392153	0.115440	4.174741	0.965327

76 KiB Raw Blame History

Self made SVD

Saving and evaluating recommendations

Embeddings

project task 5: implement SVD on top baseline (as it is in Surprise library)

Ready-made SVD - Surprise implementation

SVD

SVD biased - on top baseline

76 KiB

Raw Blame History