A demanding operation regime of Reccomender Systems (RS) is that of extreme cold start followed by a 'warming up' phase, in which a new user begins to interact with the items in the catalogue, or a new item receives the first interactions by existing users. In the RS literature the majority of approaches and techniques have been proposed for a fully 'warm' RS; only a small subset of the research addresses the cold start and within that subset, the 'warming up' phase is an area that has received little attention. During warm up new user (new item) interactions begin to appear but they are too few for a collaborative filtering technique to work, while a smart content based filtering approach (those of extreme cold start) may not be able to capture the emerging personalization traits. In this paper, starting from a stereotype driven approach developed for pure cold starts, we formulate and discuss a dynamic model that uses the few arising user/item interactions to adjust simple personalization features that are embedded in the proposed model. We demonstrate how the little personalization introduced by the dynamic model improves substantially a range of performance metrics during warm up.
Dynamic User and Item 1. Introduction ommender System (RS) is that of cold start, namely when the RS is required to produce content recommendations to new unknown users, or when new content is added to the catalogue and the task is to recommend the new unknown content to its existing users. Diferent solutions have been proposed to address cold start, some data driven and some technique driven, as reviewed in [1]. The author in [2] demonstrated how creating rating agnostic stereotypes for both users and items lead to better zero interactions available from the new user or concernrecommendation metrics during extreme cold start (i.e. 10]. ing the new item); only deep learning architectures can achieve comparable levels of accuracy, serendipity and fairness during extreme cold start to that of stereotyped metadata features [3].
When a new user begins interacting with the items catalogue, or when new content begins to be rated by some users, the RS can make use of such information to update its extreme cold start algorithm in light of the new evidence. During such ”warm up” phase, too few interactions are available to switch to a collaborative ifltering approach, but such little interactions should not be omitted as they may give important personalization 0009-0008-1062-9001 (N. A. AlRossais)
In a generic RS framework, user consumes content/item and attributes an explicit or implicit liking via a rating tentially non linear or stochastic, e.g. a deep learning architecture) that models via the user’s describing feature vector , the item’s describing feature vector , via all previous ratings provided to item by all other
users ( ≠ ) as well as such other users’ features, and
ifnally all previous ratings
given by the user to other
items and all such other item’s features. In our previous
works [2, 3] we modeled further the functional form of
our RS as in equation (
Equation (
Sterotypization of the features was introduced via
standard algorithms that can take any metadata feature type
and transform it into a stereotyped version of the original
metadata with substantially smaller dimensions. Doing
so for both user and item features aids the sparsity of the
problem, and simultaneously creates a basis, ̃ and ̃, that
improves the training process and predictive power of
model (
= ( ̃, , ̃ , ( ̄ ̃ )) = ( ̃, ( ̄ ̃), ̃ , )
information is available about the new user, ( ̄ ̃ ) represents the typical user bias of all previously observed users belonging to the same stereotypes. Likewise in the new item problem, ( ̄ ̃) represents the typical item bias across all items belonging to the same stereotype.
In the present research we extend the models (
(+1) ) (+1) = ⋅ ( ̄ ̃ ) + (1 − )⋅ < >1,.., = (/
)
In (
(+1) represents the dynamic user bias. <
>1,.., is the average observed bias of the particular user , over its first reviews. When is small the observed bias is not trustworthy, and it receives a low weight compared to
(+1) . As the user advances in his interactions with the items a stronger personalization model’s parameters optimized during training. arises, and when grows to the value of the user bias is fully personalized. In this model and are also
Beginning from pure cold start experiments, as described in [2, 3], when a user (item) is left out to represent a new user (item), all its ratings and interactions are blanked out and temporal consistency is enforced during the experiments. Only the users/items/ratings that had been expressed at a time prior to the new user (item) joining (+1) = ( ̃,
(+1) , ̃ , )
(+1) = ⋅ ( ̄ ̃) + (1 − )⋅ < >1,..,
= (/
)
In (
(+1) is the dynamic item bias, expressed as a dynamic weighted sum of ( ̃) and the actual item average bias for its first reviews, < >1,.., . The model improves the item characterisation as its number of reviews increases, obtaining full characterization after interactions.
(
results
The ideal dataset to demonstrate the application of
models (
Following the works and findings of [ 2, 3] we train
two algorithms for both and of problems (
This is chosen as our benchmark, because in the previous work [2] it represented the best extreme cold start model among the various machine learning / statistical driven models discussed in [2]. The second is a deep neural network algorithm, (DNN) whose deep layers, coupled by dropout layers, and performance during extreme cold starts have been the subject of the research [3]. The main focus of the paper is not to detail which of the functional algorithms exhibits better performance metrics during cold start and during the warm up transition, instead the objective is to investigate the efect of the dynamic bias model, how that improves the warming up phase and whether the findings can be ascribed to any particular RS functional form, or if they are independent of the RS chosen.
In the results that follow we discuss the behaviour of
the extreme cold start model and of the dynamic bias
warm up model as the new user begins rating existing
items (or existing users begin rating the new item). In
particular we examine the performance metrics at several
increasing review counts, , i.e. how the metrics change
when we take away the first reviews of the user for the
new user case (or reviews to the item for the new item)
and use these to train the dynamic bias models (
When assessing performance only the interactions that follow, from + 1 , are utilized to compute the metric for both warming model as well as for the extreme cold start model.
In the experiments presented in this paper we restrict our attention to users that have more than 100 interactions with the catalogue (approximately 4,000 users), and items that have been interacted with at least 100 times (approximately 2,500 items).
In addition to well known accuracy metrics of single predictions and ranked lists we will also discuss how the warming up phase afects serendipity (SER) and fairness (FRN) of ranked lists. SER is the property of generating ranked lists that contain useful and unexpected items, therefore exposing users to larger parts of the catalogue. to the new item. Fairness (FRN), which is related to SER, focuses on the tail of the distribution of the items that are recommended the fewest, it measures how often the least recommended items are injected in some lists. The latter metric is a very actual research issue in RS, after ethical concerns raised by some industry RS [14].
extreme cold start model. Three distinct interesting facts be adopted as discussed in [2]. One of the most relevant
can be inferred from these results; firstly the dynamic and popular measure is the Normalized Discounted
Cupersonalization improves the Rmse recommendation per- mulative Gain (Ndcg) whose definition can be found in
formance by 3 to 4% for both the new user and new item [2] and references within. In this context it is suficient
experiments. Secondly, the improvement obtained is in- to remind that the higher the Ndcg, the more valuable
dependent of the functional form of the RS chosen in the content of the ranked list to the user (both in terms
our experiments XGB vs DNN. Each RS has its own cold of items present and how they are ranked).
start base prediction ability and the warming up of the Figures (
When moving from accuracy of single rating predic- this dataset, where an approximately 5% improvement tions to a description of how well a RS proposes ranked over cold start base model is obtained within the first lists of items to users, there are several metrics that can 5, 10 interactions lifting the Ndcg curve. In the new item case the improvement over pure cold start is more mulation of the recommendation algorithms, and that is distributed over the entire number of interactions. These indicating that they can be a source of improved accuracy two distinct behaviours may be related to the diferent independently of the RS functional shape chosen. statistical nature of users and items. As a future work the author plans to enrich the models
Our previous works [2, 3] introduced operative metrics presented with a collaborative filtering element, to be
for serendipity (SER) and fairness (FRN) of ranked lists, in modeled via a Markov chain. The dynamic bias presented
this context we refer to such definitions. Table (
The model proposed in this paper addresses the warming up of RS from cold start, it presents a unified construct for both the new user and new item problems in a stereotype driven framework. The model allows for a degree of personalization via dynamic user/item bias terms.
The findings discussed in the paper demonstrate how for the new user problem the adoption of the bias allows for a rapid improvement in accuracy metrics, while for the new item problem the improvement is more gently spread over the initial incoming interactions. The only metric that using dynamic biases is observed to under perform the pure content based stereotype model is the fairness (FRN) metric for the new user problem. Such small under-performance in fairness may be the price to pay for a rapid user specialization. The new user and new item dynamic bias are efective with two alternative for[6] J. Li, Y. Wang, J. McAuley, Time interval aware start advertisements: Improving ctr predictions via self-attention for sequential recommendation, in: learning to learn id embeddings, in: Proceedings Proceedings of the 13th international conference of the 42nd International ACM SIGIR Conference on web search and data mining, 2020, pp. 322–330. on Research and Development in Information Re[7] M. Vartak, A. Thiagarajan, C. Miranda, J. Bratman, trieval, 2019, pp. 695–704.
H. Larochelle, A meta-learning perspective on cold- [11] G. Behera, N. Nain, Collaborative filtering with temstart recommendations for items, Advances in neu- poral features for movie recommendation system, ral information processing systems 30 (2017). Procedia Computer Science 218 (2023) 1366–1373. [8] C. Finn, P. Abbeel, S. Levine, Model-agnostic meta- [12] N. A. ALRossais, D. Kudenko, isynchronizer: A tool learning for fast adaptation of deep networks, in: In- for extracting, integration and analysis of movieternational conference on machine learning, PMLR, lens and imdb datasets, in: Adjunct Publication of 2017, pp. 1126–1135. the 26th Conference on User Modeling, Adaptation [9] Y. Zhu, R. Xie, F. Zhuang, K. Ge, Y. Sun, X. Zhang, and Personalization, 2018, pp. 103–107.
L. Lin, J. Cao, Learning to warm up cold item embed- [13] T. Chen, T. He, M. Benesty, V. Khotilovich, Y. Tang, dings for cold-start recommendation with meta scal- H. Cho, K. Chen, R. Mitchell, I. Cano, T. Zhou, et al., ing and shifting networks, in: Proceedings of the Xgboost: extreme gradient boosting, R package 44th International ACM SIGIR Conference on Re- version 0.4-2 1 (2015) 1–4. search and Development in Information Retrieval, [14] A. A. Kodiyan, An overview of ethical issues in 2021, pp. 1167–1176. using ai systems in hiring with a case study of ama[10] F. Pan, S. Li, X. Ao, P. Tang, Q. He, Warm up cold- zon’s ai based hiring tool (2019).