Modeling Contextual Information in Session-Aware Recommender Systems with Neural Networks, RecSys 2016 Boston, Bartłomiej Twardowski Presentation for a paper: http://dl.acm.org/citation.cfm?id=2959162 Abstract: Preparing recommendations for unknown users or such that correctly respond to the short-term needs of a particular user is one of the fundamental problems for e-commerce. Most of the common Recommender Systems assume that user identification must be explicit. In this paper a Session-Aware Recommender System approach is presented where no straightforward user information is required. The recommendation process is based only on user activity within a single session, defined as a sequence of events. This information is incorporated in the recommendation process by explicit context modeling with factorization methods and a novel approach with Recurrent Neural Network (RNN). Compared to the session modeling approach, RNN directly models the dependency of user observed sequential behavior throughout its recurrent structure. The evaluation discusses the results based on sessions from real-life system with ephemeral items (identified only by the set of their attributes) for the task of top-n best recommendations.

1. RecSys 2016 Boston
Modelling Contextual Information in Session-Aware
Recommender Systems with Neural Networks
Bartlomiej Twardowski
Warsaw University of Technology
18 August 2016
Bartlomiej Twardowski RecSys2016 Boston

2. 1 Problem Definition and Motivation
2 Explicit Session Modeling with Matrix Factorization
3 Automatic Session Modeling with NN
4 Experiments and Evaluation
Bartlomiej Twardowski RecSys2016 Boston

3. Research Motivation
Traditional RS require identified User and persistent Items.
Context-Aware RS (CARS) are not prepared for handling user
sessions data directly.
Time in recommender systems is discretized (CARS Tensor
Factorization) or used as a bias.
Successes of using Neural Networks in other fields:
FF/Conv NN in vision and image processing
RNN for Natural Language Processing
Bartlomiej Twardowski RecSys2016 Boston

4. Industry Motivation
Main objective
Capture a user short-term goals as fast as possible.
The 57.06% of all sessions are non-logged users and its fingerprint
in form of HTTP cookie or device hash does not allow us to
identify the user.1
Only 2.53% of all sessions converted to transaction. Most of the
sessions are window-shopping ones.
The 20.98% of all page views are interactions with search engine.
From all sessions, the 35.80% percent used search for finding the
right offer.
1
Presented statistics are based on Polish e-marketplace allegro.pl for 3228
M page views sample in January of 2016, where 310 M sessions was identified
by HTTP cookie or mobile device hash.
Bartlomiej Twardowski RecSys2016 Boston

5. Implicit Feedback: User-System Interactions
·
source:direct
ua:chrome1
cookie
user-id S1
search-terms:iPhone
S2
sort: price
search-terms:iPhone 5 black
V1id1
name2
desc1
seler1
attrib1: value1
. . .
V2
V3
S3category:Accessories
terms:iPhone 5 black
S4
sort: price
location: Warsaw
C+
1
add-item:id2
C+
2
add-item:id3
C−
3
remove-item:id2
V4
V5
B1 item:id3
Figure: Navigational representation of the session for a sample
e-marketplace session
Bartlomiej Twardowski RecSys2016 Boston

6. Research Problem - Key Assumptions
User is not identified by known id, just by the current
behaviour.
All users activities in a form of session are available.
Items are ephemeral, described by a set of attributes.
RS output is Top-N new, recommended items.
Bartlomiej Twardowski RecSys2016 Boston

7. User Session Definition
User sessions in this work are defined as uninterrupted sequences of
activity in the system. The session ends when the user is inactive
for more than a predefined number of minutes [1].
All sessions form a set S = {s1, . . . , sm}, where each session is
represented as a set of events ordered in time:
sm = {e
(1)
m , . . . , e
(t)
m }, where t is the time of the event occurrence
in session m. In turn, each event is described by contextual
information: e
(t)
m ∈ CE
1 × CE
2 × · · · × CE
k , where number of
attributes k depends on collected event e
(t)
m type.
Bartlomiej Twardowski RecSys2016 Boston

8. Ephemeral Items and Item Representation
The recommended items are ephemeral i.e. the item life-cycle
is too short or the availability is too dynamic to identify it
only by unique id, e.g. news, online auctions.
In such settings, known workarounds: Content-Based filtering
and items clustering.
Assuming that items will never come back - dealing with
constant cold-start problem.
I = {i1, . . . , in}. Each item is described by a set of defined
attributes in ∈ CI
1 × CI
2 × · · · × CI
p, where the number of all the
item attributes is p.
Bartlomiej Twardowski RecSys2016 Boston

9. Item and Event Encoding
All methods presented in this work require items and events being
represented by real-valued vectors.
Item encoding: CI
1 × CI
2 × · · · × CI
p → xi , xi ∈ RdI should exist,
where dI is the number of real-values in the encoded item
representation.
Similarly, the session event CE
1 × CE
2 × · · · × CE
k → xe, xe ∈ RdE ,
where dE is the dimension of the encoded event vector.
Bartlomiej Twardowski RecSys2016 Boston

10. Matrix Factorization Model
The final estimation is:
ˆy(xs, xi) = (xsQ)(xiP) + xsbp + xibq
where Q ∈ RdS ×d and P ∈ RdI ×d are matrices with d-dimensional
latent features for session and items variables respectively.
Bartlomiej Twardowski RecSys2016 Boston

11. Explicit Session Modeling
The session vector xs aggregates variables from all events within it.
Due to the fact that some assumptions have to be made about
how all events information should be encoded into single session
vector this method is considered as an explicit session modeling.
One solution, which is giving good results and is used in this work,
is to aggregate event data in a time decaying way
xsm = t
j=1
1
1+t−j x
(j)
es,m , where session vector size dS = dE .
Bartlomiej Twardowski RecSys2016 Boston

12. Recurrent NN and Feed Forward NN
Both, Recurrent Neural Network (RNN) and Feed Forward Neural
Network (FFNN) are used to predict Top-N recommendations for
the session.
The RNN is used to capture data dependency between session
events in time. It uses hidden state as the memory to handle
variable length data. In this case, the sequence of events in session.
The FFNN is used as a ranking score estimator. It uses the
representation of session context returned by RNN and the new
items data as an input.
Pairwise ranking loss func.: BPR[3], TOP-1[2], WARP[4], k-os
WARP.
Bartlomiej Twardowski RecSys2016 Boston

13. NN Architecture
Event EmbeddingItem Embedding
x
(t)
es
xi
RNN Layers
Merge/Dropout Layer
FF Layers
ˆy
(t)
s,i
Figure: Neural Network Layers Architecture.
Bartlomiej Twardowski RecSys2016 Boston

14. Datasets
dataset ALLEGRO AVITO
items 24360 4374
sessions 20904 31826
events 535871 767550
searches 366874 110204
density % 0.028 0.464
s. len. mean 25.634 24.117
s. len. std 30.282 14.877
Table: Dataset statistics
For further processing, event and items in dataset is encoded.
Bag-of-words minimal frequency is set to 10. This results in
dI =473/2710 and dE =716/5523 and for AVITO/ ALLEGRO
datasets respectively.
Bartlomiej Twardowski RecSys2016 Boston

15. Evaluated Methods
Used methods and its training parameters:
POP
CB: cos
MF: d=100, adagrad
NN: no embedding, single GRU layer with 200, MLP 400,
dropout 0.3, rmsprop
Bartlomiej Twardowski RecSys2016 Boston

16. Experiments Results
ALLEGRO AVITO
REC@20 MRR@20 REC@20 MRR@20
NN-BPR-IE 0.4131±.0051 0.1328±.0023 0.1601±.0028 0.0469±.0009
MF-BPR-IE 0.3849±.0031 0.1150±.0003 0.1894±.0008 0.0404±.0001
NN-TOPK-IE 0.3367±.0057 0.0885±.0020 0.1985±.0031 0.0579±.0004
MF-TOPK-IE 0.2862±.0024 0.0773±.0017 0.2699±.0018 0.0658±.0005
NN-TOPK-I 0.2863±.0048 0.0709±.0025 0.2003±.0012 0.0579±.0004
MF-BPR-I 0.3080±.0013 0.0864±.0012 0.1883±.0015 0.0408±.0000
MF-TOPK-I 0.2353±.0025 0.0586±.0008 0.2691±.0001 0.0660±.0004
CB 0.1858 0.0354 0.1528 0.0243
POP 0.0499 0.0129 0.0193 0.0037
Table: Experiment results for Recall and Mean Reciprocal Rank for Top-20
recommendations. A row label describes used algorithm, loss function and
contextual information (I - only items data, IE - items and events context).
The mean values with 95% CI are given.
Bartlomiej Twardowski RecSys2016 Boston

17. Experiments Results
5 10 15 20
N
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0.45
REC@N
Dataset: ALLEGRO
5 10 15 20
N
0.00
0.02
0.04
0.06
0.08
0.10
0.12
0.14
MRR@N
Dataset: ALLEGRO
NN-BPR-IE
MF-BPR-IE
NN-TOPK-IE
MF-BPR-I
NN-TOPK-I
CB
POP
5 10 15 20
N
0.00
0.05
0.10
0.15
0.20
0.25
0.30
REC@N
Dataset: AVITO
5 10 15 20
N
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.07
MRR@N
Dataset: AVITO
MF-TOPK-I
NN-TOPK-I
NN-BPR-IE
MF-BPR-IE
CB
POP
Figure: Results in Recall and MRR@Top-N Recommendations.
Bartlomiej Twardowski RecSys2016 Boston

18. Thank You.
Q&A.
Bartlomiej Twardowski
B.Twardowski@ii.pw.edu.pl, @btwardow
Bartlomiej Twardowski RecSys2016 Boston

19. References I
Gayo-Avello, D.
A survey on session detection methods in query logs and a proposal for future
evaluation.
Information Sciences 179, 12 (2009).
Hidasi, B., Karatzoglou, A., Baltrunas, L., and Tikk, D.
Session-based Recommendations with Recurrent Neural Networks.
Rendle, S., Freudenthaler, C., Gantner, Z., and Schmidt-thieme, L.
BPR : Bayesian Personalized Ranking from Implicit Feedback.
In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial
Intelligence (2009).
Weston, J., Bengio, S., and Usunier, N.
WSABIE: Scaling up to large vocabulary image annotation.
IJCAI International Joint Conference on Artificial Intelligence (2011).
Bartlomiej Twardowski RecSys2016 Boston