In this paper, we present an overview of our work towards utilization of multimodal implicit feedback in recommender systems for small e-commerce enterprises. We focus on deeper understanding of implicit user feedback as a rich source of heterogeneous information. We present a model of implicit feedback for e-commerce, discuss important contextual features affecting its values and describe ways to utilize it in the process of user preference learning and recommendation. We also briefly report on our previous experiments within this scope and describe a publicly available dataset containing such multimodal implicit feedback.
Recommender systems belong to the class of automated
content-processing tools, aiming to provide users with
unknown, surprising, yet relevant objects without the
necessity of explicitly query for them. The core of
recommender systems are machine learning algorithms
applied on the matrix of user to object preferences. In large
enterprises, user preference is primarily derived from
explicit user rating (also referred as explicit feedback) and
collaborative-filtering algorithms [
In our research, however, we foucus on small or medium-sized e-commerce enterprises. This domain introduce several specific problems and obstacles making the deployment of recommender systems more challenging. Let us briefly list the key challenges:
High concurrency has a negative impact on user loyalty. Typical sessions are very short, users quickly leave to other vendors, if their early experience is not satisfactory enough. Only a fraction of users ever returns.
For those single-time visitors, it is not sensible to provide any unnecessary information (e.g., ratings, reviews, registration details). Consumption rate is low, users often visit only a handful (0-5) of objects.
All the mentioned factors contribute to the data sparsity problem. Although the total number of users may be relatively large (hundreds or thousands per day), explicit feedback is very scarce. Also the volume of visited objects per user is limited and utilizing popularity-based approaches w.r.t. purchases is questionable at best. Furthermore the identification of unique user is quite challenging.
Despite these obstacles, the potential benefit of recommender systems is considerable, it can contribute towards better user experience, increase user loyalty and consumption and thus also improve vendor’s key success metrics.
Our work within this framework aims to bridge the data sparsity problem and the lack of relevant feedback by modelling, combining and utilizing novel/enhanced sources of information, foremost various implicit feedback features, i.e., features based on the observed user behavior.
Contrary to the explicit feedback, usage of implicit
feedback [
Our work lies a bit further from the mainstream of the
implicit feedback research. To our best knowledge, the vast
majority of researchers focus on interpreting single type of
implicit feedback [
In contrast to the majority of research trends, we
consider implicit feedback as multimodal and
contextdependent. As our aim in this direction is a long-term one,
we already published some of our findings [
We reserve Section 2.1 to the description of multimodal implicit feedback, Section 2.2 to the contextualization of user feedback and Section 2.3 to the problem of learning negative preference. For each problem, we describe relevant state of the art, current challenges as well as our proposed methods and models. Finally, we remark on the evaluation of proposed methods in Section 3 and conclude in Section 4. 2 2.1
Despite the large volume of research based on a single
implicit feedback feature, we consider implicit feedback to
be inherently multimodal. Users utilize various I/O devices
(mouse, keybord) to interact with different webpage’s GUI
controls, so there is an abundant amount of potentially Figure 2: An example of mouse movement-based feedback
interesting user actions. As the complexity of such collection on an e-commerce product detail page. Cursor
environment is overwhelming, we imposed some positions (red line) are sampled periodically. Based on the
restrictions: samples, approximated mouse in-motion time (green boxes)
Limit to the feedback related directly to some and travelled distance (blue line) are calculated. Cursor
specific object, i.e., collect the feedback only from motion log is also stored for later reasoning.
the object’s detail page.
Aggregate the same types of user actions on per this data gap was done quite recently in RecSys Challenges
session and per object basis. 2016 and 20172. Both challenges’ datasets focused on job
Focus only on user actions which can be recommendation and contained several types of positive
numerically aggregated, i.e., the desired feedback and negative user feedback. Although the dataset was not
features have numerical domain. made publicly available, some approaches proposed
See Figure 2 for an example of feedback features relevant methods to deal with multimodal implicit
derived from user actions. In the following experiments, we feedback, e.g., fixed weighting scheme [
bought this object).
Although multimodal implicit feedback is not a
mainstream research topic, we were able to trace some
research papers. One of the first paper mentioning implicit
feedback was Claypool et al. [
More recently Yang et al. [
However, the lack of publicly available datasets containing multimodal implicit feedback significantly hinders advance of the area. Some work towards bridging
Vast majority of the state-of-the-art approaches
transforms multimodal implicit feedback into a single
numeric output ̅, which can be viewed as a proxy towards
user rating. However, these methods mostly use some fixed
model of implicit feedback (i.e., predefined weights or
hierarchy of feedback features), or perform predictions
based on each feedback feature separately [
In contrast to the other approaches, we aim on estimating ̅ via machine learning methods applied on a purchase prediction task. Our approach is based on the fact that the only measurable implicit feedback with direct interpretation of preference is buying an object. Such events are however too scarce to be used as a sole user preference indicator. However, we can define a classification task to determine, based on the values of other feedback types, whether the object will be purchased by the user. The estimated rating ̅ is defined as the probability of the purchased class.
We evaluated several machine learning methods, such as
decision trees, random forests, boosting, lasso regression
and linear regressions. We also evaluated approaches based
on the more feedback the better heuristics, i.e., the higher
value of particular feedback feature implies higher user
preference. In order to make the domains of all feedback
features comparable, we utilized either standardization of
1 Please note that the dataset used for the experiments
contains also other feedback features such as number of
page prints, followed links count, several non-numeric
feedback features etc. These features seemed not relevant
for the current task, however they may be utilized in the
future.
2 [2016|2017].recsyschallenge.com
feedback values (denoted as Heuristical with STD in
results), or used empirical cumulative distribution instead
of raw feedback values (Heuristical with CDF). The
estimated rating ̅ is defined as the mean of STD or CDF
values of all feedback features for the respective user and
object. For more details on heuristical approaches please
refer to [
Although the user feedback may be processed directly, the perceived feedback values are significantly affected by the presentation of the page (i.e., device parameters) and also by the amount of contained information. Both the displaying device and the page’s complexity can be described by several numeric parameters, which we generally denote as the presentation context. See Figure 4 for some examples of relevant presentation context.
We can trace some notions of presentation context in the
literature. Yi et al. [
The presentation context differs significantly from more
commonly used user context [
In our work, we considered following presentation context features:
Volumes of text, images and links on the page.
Page dimensions (width, height).
Browser’s visible area dimensions (width, height). Visible area ratio.
Hand-held device indicator.
We evaluated several approaches utilizing biases for feedback feature values based on the current presentation context. Such approaches, however, were not very successful and in particular often did not improve over the baselines without any context at all. On the other hand, we signifficantly improve over the baseline methods while using presentation context features as additional input of the machine learning methods described in Section 2.1.1. 2.3
One of the open problems in implicit feedback utilization
is learning negative preference from implicit feedback.
Several approaches were proposed for this task including
uniform negative preference of all unvisited objects [
We propose to utilize negative preference as relations
among less and more preferred objects, i.e. to model
a partial ordering 1 < 2. This model is based on the
work of Eckhardt et al. [
We incorporate preferential relations into the
recommendation pipeline by extending collected relations
along the content-based similarity of both ignored and
selected objects (decreasing level of similarity effectively
decreases also the intensity of the relation). Afterwards, we
apply re-ranking approach taking output of some baseline
recommender and re-order the objects so that the relations
with higher intensity holds. Re-ranking algorithm considers
the relations according to the increasing intensity and
corrects the ordering induced by the relation. Thus, more
intense relations should be preferred in case of conflicts.
Details of the re-ranking algorithm can be found in [
In this section, we would like to report on the experiments conducted to evaluate models and methods utilizing multimodal implicit feedback. However, let us first briefly describe the dataset and evaluation procedures. 3.1
The dataset of multimodal user feedback (including
presentation context) was collected by observing real
visitors of a mid-sized Czech travel agency. The dataset
was collected by the IPIget tool [
In evaluation of the methods, we considered following tasks:
Purchase prediction based on the other feedback available for particular user-object pair. This scenario provides preliminary results for the methods aiming to estimate user rating ̅.
Recommending purchased objects. In this scenario, we employ leave-one-out cross-validation protocol on purchased objects (i.e., for each purchased object, all other feedback is used as a train set and we aim to recommend object, which was actually purchased by the user).
Recommending “future” user actions. In this scenario, we use older user feedback (usually 2/3 of available feedback per user) as a train set. During the recommendation phase, we recommend top-k objects to each user, while the objects from the test set visited by the user should appear on top of the list.
In several of our previous works (see, e.g., [
Results of several methods aiming to learn estimated rating ̅ based on various feature sets are displayed in Table 1 (purchase prediction task) and Table 2 (recommending purchased objects task). As we can see in Table 1, multimodal feedback significantly improves purchase prediction capability across all methods. Usage of presentation context can further improve the results, if used as additional input feature (Feedback + Context). However, if the contextual features are used as baseline predictors, the results across all methods are inferior to the results of Feedback + Context with just one exception. In several cases, the results are worse than using multimodal feedback alone. This observation indicates that some more complex dependence exists between implicit feedback, presentation context and user preference. Although it seems that the examined machine learning methods can partially discover this relation, another option to try is to hand-pick only several relevant contextual scenarios instead of the global model applied so far. Results of recommendation task also revealed a potential problem of overfitting on the purchase prediction task. Linear regression, although it performed the best in purchase prediction scenario, did not improve over binary feedback baseline. On the other hand, we can conclude that if a suitable rating prediction is selected, multimodal implicit feedback together with presentation context can improve the list of recommended objects.
Results of re-ranking approach based on preferential
relations are depict in Table 3. Re-ranking based on
preferential relations improved results of all evaluated
recommending algorithms, although the improvement was
rather modest in case of VSM. During evaluation, we
observed that in case of VSM, only the relations with
highest intensity should be applied to improve the results.
For matrix factorization approach, on the other hand, also
relations with very low intensities should be incorporated.
Another point is that the offline evaluation is naturally
focused on mere learning past user behavior and both VSM
and Popular SimCat are largely biased towards exploitation
in exploration vs. exploitation problem [
In this paper, we describe our work in progress towards utilizing multimodal implicit feedback in small ecommerce enterprises. Specifically, we focused on three related tasks. Integrate multiple types of feedback collected on the detail of an object into an estimated user rating ̅, incorporate presentation context into the previous model and utilize negative implicit feedback collected on category pages. We propose models and methods for each of the task and also provide evaluation w.r.t. top-k ranking.
Although the proposed methods statistically signifficantly improved over the baselines, the relative improvement is not too large, so our work is not finished yet. One of the important future tasks is to perform online evaluation as the offline evaluation was focused on the exploitation only. Further tasks are to propose context incorporation models specific for some context-feedback feature pairs, explore other possibilities to incorporate negative feedback and also to evaluate unified approach integrating all presented methods.
The work on this project was supported by Czech grant P46. Source codes and datasets for incorporation of presentation context can be obtained from http://bit.ly/2rJZzg3, source codes of preferential relations approach can be obtained from http://bit.ly/2symm17 and raw dataset can be obtained from http://bit.ly/2tWtRg2.