An Item based recommender system works by computing a similarity between items, which can exploit past user interactions (collaborative filtering) or item features (content based filtering). Collaborative algorithms have been proven to achieve better recommendation quality then content based algorithms in a variety of scenarios, being more efective in modeling user behaviour. However, they can not be applied when items have no interactions at all, i.e. cold start items. Content based algorithms, which are applicable to cold start items, often require a lot of feature engineering in order to generate useful recommendations. This issue is specifically relevant as the content descriptors become large and heterogeneous. The focus of this paper is on how to use a collaborative models domain-specific knowledge to build a wrapper feature weighting method which embeds collaborative knowledge in a content based algorithm. We present a comparative study for diferent state of the art algorithms and present a more general model. This machine learning approach to feature weighting shows promising results and high flexibility.
Recommender systems aim at guiding the user through the navigation of vast catalogs and in recent years they have become widespread. Among item based algorithms, content based are the most widely used, as they provide good performance and explainability. Content Based algorithms recommend items based on similarities computed via item attributes. Although being applicable in any circumstance in which at least some information about the items is available, they sufer from many drawbacks related to the quality of the item features. It is hard and expensive to provide an accurate and exhaustive description of the item. In recent years the amount of data which is machine-readable on the web has increased substantially, it is therefore possible to build heterogeneous and complex representations for each item (e.g. textual features extracted from web pages). Those representations however often comprise of high number of features and are sparse and noisy, to the point where adding new data hampers the recommendation quality. Feature Knowledge-aware and Conversational Recommender Systems (KaRS) Workshop 2018 (co-located with RecSys 2018), October 7, 2018, Vancouver, Canada.
Copyright for the individual papers remains with the authors. Copying permitted
for private and academic purposes. This volume is published and copyrighted by its
editors.
weighting, which can be considered as a generalization of feature
selection, is a useful tool to improve content based algorithms.
Traditional Information Retrieval methods like TF-IDF and BM25 [
The most common approach to tackle the cold-start item problem is to rely on content-based algorithms, whose accuracy is sometimes much poorer. In this paper we further investigate the cold-start item problem focusing on how to learn feature weights able to better represent feature importance from the user point of view. We provide a comparative study of state of the art algorithms and present a more general model demonstrating its applicability on a wide range of collaborative models. Moreover we describe a two step approach that: (1) exploits the capability of a generic collaborative algorithms to model domain-specific user behaviour and achieve stateof-the-art performance for warm items (2) embeds the collaborative knowledge into feature weights The rest of the paper is organized as follows. In Section 2 we briefly review the literature in the cold-start recommendation domain, in Section 3 our framework is presented and a comparison of the diferent algorithms is discussed in Section 4. Finally conclusions and future works are highlighted in Section 5 2
Various tools are at our disposal to assess the relevance of a feature.
We can distinguish feature weighting algorithms in three categories:
ifltering, embedding and wrappers [
Filtering methods usually rely on information retrieval. Methods like TF-IDF or BM25 are not optimized with respect to a predictive model, therefore the resulting weights are not domain-specific and can not take into account the rich collaborative information, even when available.
Embedding methods learn feature weights as a part of the model
training, examples of this are UFSM [
W = D + VT V where D is a diagonal matrix having as dimension the number of features, nF , and V ∈ RnL ×nF . The number of latent factors nL is treated as a parameter.
Wrapper methods rely on a two step approach, by learning
feature weights on top of an already available model, an example of
this is Least-square Feature Weights (LFW) [
sim(i, j) = fiT Dfj
All these algorithms, to our best knowledge, have never been subject to a comparative study. 3
Feature weighting can be formulated as a minimization problem whose objective function is: argmin
W 2
F S(CF) − S(W) + λ ∥D∥2F + β ∥V∥2F where S(CF) is any item-item collaborative similarity, S(W) is the similarity function described in Equation (1), W is the feature weight matrix which captures the relationships between items features and β and λ are the regularization terms, we call this model Collaborative boosted Feature Weighting (CFW). This model can either use the latent factors (CFW D+V), as in FBSM, or not (CFW D), as in LFW.
The advantages of learning from a similarity matrix, instead of using the user interactions, are several: • High flexibility in choosing the collaborative algorithm which can be treated as a black box • Similarity values are less noisy than user interactions • The model is simpler and convergence is faster
In this paper a two steps hybrid method is presented, in order to easily allow to embed domain-specific user behaviour, as represented by a collaborative model, in a weighted content based recommender. The presented model is easily extendable to other algorithms and domains. The learning phase is composed by two steps. The goal of the first step is to find the optimal parameters for the collaborative algorithm, to this end a collaborative algorithm is trained and tuned on warm items. The second step applies an s r e s u train A warm items validation test
B
C
embedded method to learn the optimal item feature weights that
better approximate the item-item collaborative similarity obtained
before.
We solve (4) via SGD applying Adam [
We performed experiments to confirm that our approach is capable of embedding collaborative knowledge in a content based algorithm improving its recommendation quality in an item cold-start scenario. 4.1
In order to evaluate our approach we used only item descriptors accessible via web, which we assume will be available for new items, excluding user generated content. The datasets are the following:
Netflix. Enriched with structured and unstructured attributes extracted from IMDB. This dataset has 250k users, 6.5k movies, 51k features and 8.8M ratings in 1-5 scale. The rating data is enriched with 6.6k binary editorial attributes such as director, actor and genres.
The Movies Database. 1: 45k movies with 190k TMDB editorial features and ratings for 270k users. This dataset has been built from the original one by extracting its 70-cores.
For all the listed datasets, features belonging to less than 5 items
or more than 30% of the items have been removed, as done in [
Step 1 - Collaborative algorithm: In this step the training of the collaborative algorithm is performed on warm items, which are defined as the union of split A and B, see Figure 1. Hyper-parameters are tuned with a 20% holdout validation.
Knowledge-aware and Conversational Recommender Systems (KaRS) Workshop 2018 (co-located with RecSys 2018), October 7, 2018, Deriving item features relevance from collaborative domain knowledge Vancouver, Canada.
Step 2 - Feature weights: In this case a cold-item validation is chosen as it better represents the end goal of the algorithm to find weights that perform well on cold items. The collaborative similarity will be learned using only split A with the hyper-parameters found in the previous step, see Figure 1. An embedded method is then used to learn the optimal item feature weights that better approximate the item-item collaborative similarity obtained before. The hyperparameters of the machine learning model are tuned using split B while set C is used for pure testing. 4.3
The collaborative similarity matrices used in our approach are
computed using diferent algorithms: KNN collaborative (using
the best-performing similarity among: Cosine, Pearson, Adjusted
Cosine, Jaccard), P3alpha [
Table 1 shows the recommendation quality of both pure content based and hybrid baselines, as well as CFW D evaluated on all collaborative similarity models. Table 2 shows the performance of the two components of CFW and FBSM on Netflix, results for the other dataset are omitted as they behave in the same way.
From Table 1 we can see that FBSM performs poorly, which indicates that while it has the power to model complex relations, it is more sensitive to noise and data sparsity than other algorithms. Learning from a similarity matrix, as LFW and CFW D+V does, results in much better results than FBSM. In Table 2 it is possible to see that the latent factor component was able to learn very little, this suggests that while rendering the model more expressive, it introduces noise and numerical instability. Note that the performance using of the diagonal component alone is higher than the one obtained by adding the V component. However, its efectiveness could be influenced by the feature structure and therefore it might be relevant in some specific cases. rom Table 1 we can see that by using only the diagonal and discarding the latent factor component, the performance improves significantly.
While LFW only learned feature weights using a SLIM similarity matrix, our results indicate that it is possible to learn from a wide variety of item based algorithms, even those not relying on machine learning. This means that machine learning feature weights can be used on top of already available collaborative algorithm with little efort. Using an intermediate similarity matrix while ofering additional degrees of freedom in the selection of the collaborative model, also simplifies the training phase and improves overall performance. 5
In this paper we presented diferent state of the art feature weighting methods, compared their performance and proposed a more general framework to efectively apply machine learning feature weighting to boost content based algorithms recommendation quality, embedding user domain-specific behaviour. We also demonstrate high flexibility in the choice of which collaborative algorithm to use. Future work directions include testing the proposed approach in diferent datasets and domains, as well as exploring the symmetric problem of using the collaborative similarity to discover item features or to reduce feature noise.