Ph.D VIVA Slides presenting Discriminate2Rec, a discriminative temporal interest-based content-based recommendation framework that employs a novel three-stage preference learning model that discriminates between items’ attributes based on their influence on user temporal preferences to improve both temporal and semantical attribute-level profile coherence for more accurate recommendation. We exploit different user-dependent and item/attribute-dependent temporal dynamics to infer positive and negative user-attribute temporal interest weights. Also, we introduce a negation modelling technique to model user-attribute negative interests, which allows us to learn attribute-level coherent user profiles.
Discriminative Temporal Interest-based Preference Learning Framework for Semantics-aware Content-based Recommendation
1. Discriminative Temporal
Interest-based Preference
Learning for Semantics-aware
Content-based Recommendation
Multimedia University
Faculty of Computing and Informatics
Naji A. Albatayneh
Ph.D. (IT) Candidate
Supervisor
Dr. Khairil Imran Ghauth
July 1st, 2022
Co-supervisor
Assoc. Prof. Chua Fang-Fang
2. Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 2
Outline
1. Introduction
Background & Motivation
Problem Statements & Research Questions
Research Objectives
2. Related Work
3. Proposed Method
Discriminate2Rec
Framework’s Components
4. Experimental Evaluation & Results
Data Sets, Evaluation Protocol & Evaluation Metrics
Experiment 1: Best Performing Interest Inference Algorithm
Experiment 2: Comparing Discriminate2Rec Against State-of-the-art
Methods
5. Conclusion
Contributions & Future Work
3. Background …(1/2)
Information Overload Problem
Information Retrieval
Search Engines (e.g., Google)
Allow users to create their own search query statements
Compare user’s query statement with other items
Information Filtering
Recommender Systems (e.g., Netflix)
Allow users to rate items and learn a user profile accordingly
Compare user’s profile with other items
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 3
Challenging Task
2 5 6.5 9 12.5 15.5 18 26 33 41
64.2 79 97
120
147
181
0
50
100
150
200
2010 2011 2012 2013 2014 2015 2016 2017 2018* 2019* 2020* 2021* 2022* 2023* 2024* 2025*
Data
in
Zettabyte
Year
Volume of data worldwide from 2010 to 2025
4. Background …(2/2)
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 4
Recommendation
Accuracy
User
Profile
Preference
Feedback
Learning
Method
Similarity
Measure
Quality
of Data
Recommendation Approaches:
Content-based Recommender Systems (CBRSs)
Collaborative Recommender Systems (CRSs)
Hybrid Recommender Systems (HRSs)
5. Motivation & Scope
The motivation behind this research is to overcome the problem of profile
incoherence in order to improve the recommendation accuracy in CBRSs.
The scope of this thesis is limited to improving the accuracy of
recommendation in CBRSs through improving semantical and temporal
coherence of user profile, which involves user preference learning, user
modelling and profiling, and temporal dynamics modelling techniques in the
field of CBRSs.
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 5
Profile
Coherence
Temporal
Coherence
Drift in user preferences
Semantical
Coherence
Uncertainty in Rating
Neglecting Negative Preferences
6. Problem Statements & Research Questions
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 6
Lack of effective
temporal dynamics
modelling techniques
for modelling the
dynamic time-drifting
nature of user-item
interactions in
CBRSs
Problem 1
Lack of effective user
interests inference
techniques for
inferring temporal
discriminative user-
attribute interest
weights
Problem 2
Lack of effective
negation modelling
techniques for
modelling and
learning negation in
user preference
feedback
Problem 3
Lack of effective
preference learning
and recommendation
techniques for
improving semantical
and temporal
attribute-level profile
coherence
How can CBRSs learn
user profiles while
assuring its
semantical and
temporal coherence
at attribute-level in
order to improve the
recommendation
accuracy?
Problem 4
Question 4
How can negation in
user preference
feedback be
effectively modelled
at attribute-level and
learned in a content-
based user profile to
improve its
coherence?
Question 3
How can positive and
negative temporal
discriminative user-
attribute interests be
inferred based on
user ratings, items
descriptive content
and temporal
dynamics?
Question 2
How can multiple
interconnected user-
dependent and item-
dependent temporal
dynamics be
modelled and learned
in a content-based
user profile in order
to improve the
accuracy of
recommendation?
Question 1
7. Research Objectives
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 7
To propose user
interest inference
techniques using user
preference feedback
and temporal
dynamics
Objective 1
To propose preference
learning and
recommendation
framework based on
user temporal
discriminative
interests and negation
modelling techniques
Objective 2
To evaluate the
accuracy of the
proposed
recommendation
techniques against
similar state-of-the-art
methods in terms of
RMSE, MAE, Precision,
Recall, and F1-
measure
Objective 3
Problem 1 Problem 2 Problem 3 Problem 4
9. Type Used techniques Limitations
[1]
(Saia et al.,
2016)
CBRS employing a semantic
measure to determine the
semantical coherence of a
candidate rated item with
other rated items in a
content-based user profile.
• No negation modelling.
• No temporal dynamics
modelling.
• Item-level semantical
profile coherence.
[2]
(Saraswat &
Chakraverty
, 2020)
CBRS Employing topic extraction
and modelling techniques on
user-generated content to
improve profile semantical
coherence.
• No negation modelling.
• No temporal dynamics
modelling.
• Item-level semantical
profile coherence.
9
Related Work …(1/6)
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 9
10. Type Used techniques Limitations
[3]
(H. Zhang
et al., 2020)
CBRS Exploiting hidden
relationships among items’
features for improving the
semantical coherence of user
profile.
• No negation modelling.
• No temporal dynamics
modelling.
• Item-level semantical
profile coherence.
[4]
(Gatzioura
et al., 2021)
CBRS Enriching user profile by
semantic features extracted
from playlists’ keywords.
• No negation modelling.
• No temporal dynamics
modelling.
• Item-level semantical
profile coherence.
10
Related Work …(2/6)
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 10
11. Type Used techniques Limitations
[5]
(Bag, Kumar,
Awasthi, & Tiwari,
2019;
…………………
[6]
Castro, Yera, &
Martínez, 2018;
…………………
[7]
Choudhary, Kant, &
Dwivedi, 2017;
…………………
[8]
Toledo, Mota, &
Martínez, 2015)
CRS Improving profile
coherence through
correcting user
noisy ratings in
order to cope with
the natural noise.
• No negation modelling.
• No temporal dynamics
modelling.
• No semantical
coherence.
• Item-level profile
coherence.
Related Work …(3/6)
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 11
12. 12
Related Work …(4/6)
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 12
High-level Architecture
(Boratto et al., 2017)
13. 13
Related Work …(5/6)
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 13
Profile Cleaner Component
(Boratto et al., 2017)
14. Type Used techniques Limitations
[9]
(Boratto et al.,
2017)
CBRS Removing
Incoherent Items
from user profile
based on temporal
and semantical
thresholds.
• Prone to lose much signal.
• No negation modelling.
• No temporal dynamics
modelling.
• Item-level profile coherence.
14
Related Work …(6/6)
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 14
The current state of CBRSs still lacks preference learning methods that take
into account multiple semantical and temporal aspects to mitigate the
problem of profile incoherence and, hence, maintain more semantical and
temporal attribute-level profile coherence.
15. Proposed Method
Introducing Discriminate2Rec
System’s Framework
Discriminative Interests Inference
Dynamic Profile Learner Component
Negation-based Profile Learner Component
Outline
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 15
Fully Published in
Albatayneh, N. A., Ghauth, K. I., & Chua, F. F. (2022) Discriminate2Rec: Negation-based dynamic
discriminative interest-based preference learning for semantics-aware content-based
recommendation. Expert Systems with Applications, 199, 116988
Partially Published in
Albatayneh, N. A., Ghauth, K. I., & Chua, F. F. (2018). Utilizing learners’ negative ratings in semantic
content-based recommender system for e-learning forum. Journal of Educational Technology &
Society, 21(1), 112-125
16. A novel three-stage preference learning and recommendation
framework
Discriminating between rated items’ attributes based on their
influence on user’s temporal preferences
Temporal dynamics in user preferences
User-dependent temporal effects
Item-dependent temporal effects
Negation modelling
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 16
Discriminate2Rec
19. Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 19
A basic algorithm to inferring static user-attribute interests
TAGOMMENDERS by (Sen et al., 2009)
Cannot yield realistic user-attribute interests as it lacks the awareness of
multiple user-dependent and item-dependent temporal dynamics.
We adapted this algorithm
Discriminative Interests Inference …(2/5)
20. Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 20
Modelling User-dependent temporal effects
Drift in User’s Baseline Ratings
Gradual:
Sudden:
estimating the time deviation associated with user’s rating
Discriminative Interests Inference …(3/5)
24. Negation is defined in terms of
vectors orthogonality in
vector spaces
Two given attributes/items are
considered entirely irrelevant
if their corresponding vectors
are orthogonal.
(Widdows & Peters, 2003)
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 24
Negation-based Profile Learner
25. Data Sets
Evaluation Protocol
Evaluation Metrics
Experiment 1
Best Performing Interest Inference Algorithm
Experiment 2
Comparing Discriminate2Rec Against State-of-the-art Methods
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 25
Outline
Experimental Evaluation & Results
Fully Published in
Albatayneh, N. A., Ghauth, K. I., & Chua, F. F. (2022) Discriminate2Rec: Negation-based dynamic
discriminative interest-based preference learning for semantics-aware content-based
recommendation. Expert Systems with Applications, 199, 116988
26. MovieLens Forum
Users (U) 300 425
Items (I) 5810 4082
Ratings (R) 39685 19310
Positive R 82.76% 43.64%
Negative R 17.23% 56.36%
Sparsity 97.72% 98.88%
Avg. R/U ± SD 132.28 ± 210.34 45.43 ± 52.18
Max. R/U 2179 93
Min. R/U 20 27
Rating Scale 0.5-5 1-6
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 26
Data Sets
27. k-fold Expanding Window Last-one-out (k-fold EW-LOO) Protocol
*sliding-window evaluation protocol (SW-EVAL) (Jeunen, 2019)
**last-one-out protocol (Q. Zhao et al., 2018)
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 27
Evaluation Protocol
Past
K=1
Present
Dataset
Fold 1
Fold 2
Fold 3
Fold 4
Fold 5
K=2 K=3 K=4 K=5
Legend
Dataset
Training set
Testing set
Inner Loop
(Apply Last-one-out protocol within each
window)
Outer
Loop
(Expanding
window)
Time
29. Compared Interest Inference Algorithms
Algorithm Specifications Implementation
Static No temporal dynamics modelling. Eq. (4.6) Page 82
DynU
Accounts only to gradual drifts in user’s baseline ratings by
modelling the time deviation associated with user rating. Eq. (4.7) Page 82
DynUA
Accounts to gradual drifts in user’s baseline ratings as well
as to attribute-dependent temporal effects by modelling
attribute’s popularity as a function that changes over time.
Eq. (4.8) Pages
82-83
DynUA+
Accounts to both gradual and sudden drifts in user’s
baseline ratings, and to attribute-dependent temporal
effects.
Eq. (4.9) Page 83
DynUA++
Accounts to gradual and sudden drifts in user’s baseline
ratings, and considering attribute-dependent temporal
effects as not a completely user-independent measure.
Eq. (4.10) Pages
83-84
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 29
Experiment 1 …(1/5)
30. Comparing Interest Inference Algorithms
RMSE on Forum Dataset
Algorithm
Forum
Without Negation With Negation
f =10 f =50 f =100 f =200 f =10 f =50 f =100 f =200
Static 0.9038 0.8889 0.8830 0.8802 0.8883 0.8789 0.8771 0.8763
DynU 0.8973 0.8857 0.8767 0.8739 0.8812 0.8719 0.8711 0.8680
DynUA 0.8967 0.8833 0.8752 0.8707 0.8789 0.8705 0.8682 0.8654
DynUA+ 0.8795 0.8724 0.8673 0.8638 0.8715 0.8627 0.8608 0.8539
DynUA++ 0.8793 0.8719 0.8669 0.8630 0.8710 0.8622 0.8597 0.8531
All algorithms benefit from the growing number of factorisation
dimensions as well as from the introduction of negation modelling in
user preferences
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 30
Experiment 1 …(2/5)
31. Comparing Interest Inference Algorithms
MAE on Forum Dataset
Algorithm
Forum
Without Negation With Negation
f =10 f =50 f =100 f =200 f =10 f =50 f =100 f =200
Static 0.7127 0.6974 0.6926 0.6911 0.6989 0.6904 0.6883 0.6851
DynU 0.7035 0.6941 0.6898 0.6889 0.6931 0.6867 0.6839 0.6812
DynUA 0.7004 0.6919 0.6884 0.6874 0.6898 0.6841 0.6821 0.6789
DynUA+ 0.6901 0.6870 0.6853 0.6814 0.6819 0.6797 0.6760 0.6717
DynUA++ 0.6890 0.6852 0.6833 0.6795 0.6807 0.6786 0.6751 0.6704
User-dependent temporal dynamics are shown to be consistently
more influential than attribute-dependent temporal dynamics
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 31
Experiment 1 …(3/5)
32. Comparing Interest Inference Algorithms
RMSE on MovieLens Dataset
Algorithm
MovieLens
Without Negation With Negation
f =10 f =50 f =100 f =200 f =10 f =50 f =100 f =200
Static 0.9210 0.9132 0.9083 0.8992 0.9036 0.8986 0.8970 0.8885
DynU 0.9143 0.9068 0.9010 0.8931 0.8965 0.8917 0.8898 0.8817
DynUA 0.9143 0.9050 0.8999 0.8915 0.8946 0.8899 0.8880 0.8799
DynUA+ 0.9053 0.8972 0.8917 0.8839 0.8860 0.8818 0.8795 0.8716
DynUA++ 0.9038 0.8956 0.8899 0.8826 0.8841 0.8809 0.8781 0.8705
* Sudden drifts in user’s baseline ratings (captured by the day-specific parameter) are
demonstrated to have the most significance on the accuracy improvements on both data
sets, achieving an average improvement of 0.0096 and 0.0062 on the Forum dataset in
terms of RMSE and MAE respectively, and an average improvement of 0.0083 and 0.0081
on the MovieLens dataset in terms of RMSE and MAE, respectively.
*
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 32
Experiment 1 …(4/5)
33. Comparing Interest Inference Algorithms
MAE on MovieLens Dataset
Algorithm
MovieLens
Without Negation With Negation
f =10 f =50 f =100 f =200 f =10 f =50 f =100 f =200
Static 0.7412 0.7331 0.7283 0.7195 0.7238 0.7188 0.7172 0.7089
DynU 0.7344 0.7265 0.7212 0.7134 0.7167 0.7119 0.7102 0.7019
DynUA 0.7340 0.7252 0.7197 0.7118 0.7145 0.7105 0.7083 0.6997
DynUA+ 0.7251 0.7175 0.7116 0.7042 0.7063 0.7020 0.6999 0.6921
DynUA++ 0.7236 0.7157 0.7098 0.7036 0.7043 0.7011 0.6983 0.6910
1. Accounting to gradual drifts in user’s baseline ratings by modelling the time deviation
associated with user rating has notable improvements.
2. Considering attribute’s popularity as not completely a user-independent measure,
hence modelling the change in user’s rating scale as a multiplicative factor influencing
the attribute’s popularity is beneficial.
1
2
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 33
Experiment 1 …(5/5)
34. Discriminate2Rec Against State-of-the-art Methods
Compared Methods
timeSVD (Y. Koren & Bell, 2015)
SCB-baseline (de Gemmis et al., 2015), exploiting TAGOMMENDERS
profiling technique (Sen et al., 2009)
SCB-ICoh (Boratto et al., 2017)
CB-DPM (Cami et al., 2019)
Discriminate2Rec (employing DynUA++ with negation modelling
and a LSA semantic vector space of dimensionality f=200)
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 34
Experiment 2 …(1/4)
35. Discriminate2Rec Against State-of-the-art Methods
RMSE & MAE on Both Data Sets
Method
Forum MovieLens
RMSE MAE RMSE MAE
SCB-baseline 0.9205 0.8910 0.9381 0.8503
SCB-ICoh 0.9114 0.8126 0.9199 0.8317
timeSVD 0.9007 0.8032 0.8992 0.8110
CB-DPM 0.8882 0.7011 0.8973 0.8024
Discriminate2Rec 0.8531 0.6704 0.8705 0.6910
Our approach using a static interest inference algorithm with and
without modelling the negation, still outperforms the state-of-the-art
approach denoted by SCB-ICoh (Boratto et al., 2017).
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 35
Experiment 2 …(2/4)
36. 0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Precision Recall F1 Precision Recall F1 Precision Recall F1 Precision Recall F1 Precision Recall F1
@N = 5 @N = 10 @N = 15 @N = 20 @N = 30
Performance
Value
Top-N Recommendation
Forum Dataset
SCB-baseline CB-DPM timeSVD SCB-Icoh Discriminate2rec
Discriminate2Rec Against State-of-the-art Methods
Precision, Recall & F1 @ (5, 10, 15, 20, 30) on Forum Dataset
+ 0.0519 + 0.0702 + 0.0763 + 0.0854 + 0.0876
Avg. Improvement 0.0743
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 36
Experiment 2 …(3/4)
37. 0.0
0.1
0.2
0.3
0.4
0.5
0.6
Precision Recall F1 Precision Recall F1 Precision Recall F1 Precision Recall F1 Precision Recall F1
@N = 5 @N = 10 @N = 15 @N = 20 @N = 30
Performance
Value
Top-N Recommendation
MovieLens Dataset
SCB-baseline CB-DPM timeSVD SCB-Icoh Discriminate2rec
Discriminate2Rec Against State-of-the-art Methods
Precision, Recall & F1 @ (5, 10, 15, 20, 30) on MovieLens Dataset
+ 0.0314 + 0.0484 + 0.0481 + 0.0666 + 0.0821
Avg. Improvement 0.0553
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 37
Experiment 2 …(4/4)
38. Conclusion
Contributions
Future Work
List of Publication
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 38
Outline
Conclusion & Future Work
39. A novel three-stage preference learning and recommendation
framework (Discriminate2Rec).
Modelling of multiple interconnected user-dependent and item-
dependent temporal dynamics.
Discriminating between rated items’ attributes based on their influence
on user temporal preferences.
Modelling negation based on user’s negative interests.
Improving semantical and temporal attribute-level profile coherence.
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 39
Conclusion
40. Objective 1 Objective 2 Objective 3
Introduction of
temporal dynamics
modelling
techniques for
developing user
discriminative
interest inference
techniques
Contribution
1
Development of
user temporal
discriminative
interest inference
techniques based
on user preference
feedback and
temporal dynamics
Contribution
2
Introduction of
negation modelling
techniques based
on orthogonal
negation in vector
spaces and user
negative
discriminative
interests
Contribution
3
A novel three-
stage preference
learning and
recommendation
framework based
on user temporal
discriminative
interests and
negation-based
modelling
Contribution
4
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 40
Contributions
41. Scalability and computational time complexity
Exploring new negation modelling techniques
Holistic modelling for user robust interest inference
Additional experiments and data sets
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 41
Future Work
42. Albatayneh, N. A., Ghauth, K. I., & Chua, F. F. (2022) Discriminate2Rec:
Negation-based dynamic discriminative interest-based preference
learning for semantics-aware content-based recommendation. Expert
Systems with Applications, 199, 116988. PREVIEW
Albatayneh, N. A., Ghauth, K. I., & Chua, F. F. (2018). Utilizing learners’
negative ratings in semantic content-based recommender system for
e-learning forum. Journal of Educational Technology & Society, 21(1),
112-125. PREVIEW
Albatayneh, N. A., Ghauth, K. I., & Chua, F. F. (2014). A Semantic
Content-Based Forum Recommender System Architecture Based on
Content-Based Filtering and Latent Semantic Analysis. In Recent
Advances on Soft Computing and Data Mining (pp. 369-378). Springer
International Publishing. PREVIEW
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 42
List of Publication
43. Thank You!
Discriminative Temporal Interest-based Preference Learning for Semantics-aware Content-based Recommendation 43
Naji A. Albatayneh
naji.albatayneh@gmail.com
Editor's Notes
The heterogeneity and the exponential growth of the content on the Web have imposed an Information Overload problem, where users nowadays face diffeculties in finding content suit their needs among all the available content on the Web, which indeed affects some aspects of their daily life such as their decision-making.
However, this has motivated the need for software systems that can assist users overcoming this problem.
The early solutions were offered by IR systems, such as Search Engines (e.g., Google)
These systems assume that users are aware of their information needs by Allowing them to create their own search query statements. However, such systems retrieve as many similar items as possible after Comparing user’s query statement with other available items.
………
Bottom-up semantics: known in the literature as distributional models, learn semantic representations of items and user profiles by analysing the context in which terms occur in large corpora of textual documents. LSA, RI
Top-down semantics: generally exploit external sources of knowledge in order to annotate and capture the semantics of items and user profiles. Many recent works employed this semantic approach, ranging from incorporating ontologies, LOD, Wiki.
A non-liner learning approach employs machine learning algorithms such as the Nearest Neighbour algorithm, Support Vector Machines (SVM), Decision Trees, Random Forest, and Neural Networks to learn user preferences.
In probabilistic learning approach, a Naïve Bayes probabilistic model can be generated based on a previously observed data in order to classify a given candidate item i and, accordingly decide whether to add it to the class of +ve or –ve preferences
The linear approach learns user preferences by inducing a linear prototype vector of weights indicating the degree of a given user’s interest in each attribute of the items rated by the same user.
note1
Thus, profile coherence is a central problem in recommender systems field since it significantly affects the accuracy of recommendation, but no approach has ever studied, from a content-based filtering perspective, how semantical and temporal attribute-level profile coherence can be improved by the discrimination between the attributes of the rated items based on their influence on user temporal preferences.
a basic algorithm to inferring static user-attribute interests as a weighted average rating of a given user for items that contain a given attribute was proposed in (Sen et al., 2009)
inferring static user-attribute interests as a weighted average rating of a given user for items that contain a given attribute was proposed in (Sen et al., 2009)
The system determines the thresholds based on the availability and richness of the data instances.
Both metrics measure how close are the predicted ratings by the recommender system to the actual ratings given by the user, in which, RMSE gives relatively higher weights to large errors compared to MAE
P: the ratio of the retrieved items that are relevant to user interests
R: the ratio of the items that are relevant to user interests and are retrieved
F1: the standard harmonic mean of Precision and Recall. It measures the overall decision support accuracy performance of a recommendation algorithm
sudden drifts in user’s baseline ratings which are captured by the day-specific parameter implemented in DynUA+ variant, are demonstrated to have the most significance on the accuracy improvements on both data sets, achieving an average improvement of 0.0096 and 0.0062 on the Forum dataset in terms of RMSE and MAE respectively, and an average improvement of 0.0083 and 0.0081 on the MovieLens dataset in terms of RMSE and MAE, respectively.
timeSVD (Y. Koren & Bell, 2015): a well-known collaborative filtering approach that incorporates multiple temporal information into a Matrix Factorisation (MF) method. A public implementation of timeSVD available in LibRec toolkit is exploited, in which we employed a factorisation dimensionality f=200. This method learns a temporal dynamics-aware MF user model based on a set of latent factors to capture user temporal interests.
SCB-baseline: a semantics-aware content-based recommender system baseline employing a high-level architecture proposed in (de Gemmis et al., 2015). A user-attribute interest-based user profiling technique proposed in (Sen et al., 2009) is exploited to learn static interest-based content-based user profiles, in which, static user-attribute interests are computed as a weighted average rating of a given user for items that contain a given attribute as defined in Eq. 3.3. We employ LSA-based semantic vector space of dimensionality f=200 and the cosine similarity measure.
SCB-ICoh (Boratto et al., 2017): a semantics-aware content-based recommender system that improves user profile coherence by employing both semantic and temporal thresholds to remove incoherent items from a content-based user profile, in which, the semantic similarity between a candidate rated item and all other rated items in a user profile is measured besides measuring the average temporal distance of the candidate rated item with all other items in user profile. Note that, this method removes the whole item’s descriptive content from a content-based user profile if the item is considered incoherent by the system. We employ LSA of factorisation dimensionality f=200 and the cosine similarity measure.
CB-DPM (Cami et al., 2019): a Dirichlet Process Mixture model that adapt both user interests and user preferences for temporal content-based recommendation based on a Bayesian nonparametric recommendation framework.
Discriminate2Rec: our proposed recommendation approach employing the best-performing interest inference algorithm (i.e., DynUA++ with negation modelling and a LSA-based semantic vector space of dimensionality f=200).
>>This is attributed to the significance of our attribute discrimination technique which maximises the utilisation of signal.
This possibly can be accomplished by exploiting user’s personal data and preferences extracted from multiple heterogeneous data sources, such as social media networks, mobile apps, and smart, personal and wearable devices. Such heterogeneous data can be processed, merged and modelled using natural language processing and machine learning techniques to infer more robust user interests
it would be interesting to estimate the natural limitation, known as Magic Barrier, of the proposed recommendation framework, which is imposed by the natural noise in user ratings due to several factors such as human uncertainty