One-class collaborative filtering (OCCF), or recommendation with one-class feedback such as shopping records, has recently gained more attention from researchers and practitioners in the community. The main reason is that one-class feedback in the form of (user, item) pairs are often more abundant than numerical ratings in the form of (user, item, rating) triples as exploited by traditional collaborative filtering algorithms. However, most of the previous work on OCCF do not consider the temporal context, which is known of great importance to users' preferences and behaviors. In this paper, we first formally define a new problem called time-aware OCCF (TOCCF), and then design a novel timeaware similarity learning (TSL) model accordingly. Our TSL is based on a novel time-aware weighting scheme and a seminal work on similarity learning, and is able to learn the item similarities more accurately. Empirical studies on two large real-world datasets show that our TSL model can integrate the temporal information effectively, and perform significantly better than several state-of-the-art recommendation algorithms.
One-class collaborative filtering (OCCF) [
In order to model the one-class feedback, two main lines of
techniques are usually adopted, which are parallel to that of
collaborative filtering, including memory-based OCCF and
model-based OCCF. For memory-based OCCF, the only
difference from that of memory-based CF is that the
similarity between two users or two items are estimated based
on the one-class feedback instead of the ratings. For
modelbased OCCF, the techniques are often different from that of
model-based CF, in particular of the underlying assumption
for the learning task of positive feedback only and the
prediction rule based on similarity learning. The most well-known
preference assumption for one-class feedback is probably the
pairwise preference assumption called Bayesian personalized
ranking defined on the difference between a purchased
product and an un-purchased one [
The aforementioned advances in modeling one-class
feedback in OCCF have indeed achieved great success in various
recommendation applications. However, we find that very
few work have explicitly studied the temporal effect in
OCCF, though it has shown to be very helpful in user behavior
modeling in CF [
In this paper, we first design a time-aware weighting scheme
for the reliability of the positive feedback, and then propose
a time-aware similarity learning (TSL) model by integrating
the weight as a confidence score into the similarity learning
model [
In time-aware one-class collaborative filtering (TOCCF), we have n users, m items and their positive feedback in the form of (user, item, time) triples, e.g., (u, i, tui), denoting user u has a positive feedback on item i at time tui. In TOCCF, our goal is to learn users’ preferences from the positive feedback and associated temporal information, and provide a personalized ranked list of items for each user u that he or she may like in the future.
Notice that in OCCF [
It is well known that similarity measurement is critical
in collaborative filtering, because it determines the
neighborhood of a certain user u and thus affects the preference
prediction of user u on other items. The state-of-the-art
approach [
1 p|Ni\{i}|
Vi·WiT′·, (1) where Ni = {i′|(u, i′, tui′ ) ∈ Tu}, and Vi·, Wi′· ∈ R1×d are latent feature vectors to be learned for item i and item i′, respectively. The similarity si′i or the latent vectors Vi· and Wi′· can be learned via some pointwise or pairwise loss functions in an optimization problem.
With the learned similarity si′i, the preference of user u
on item i can then be predicted as follows [
X i′∈Nu\{i}
si′i + bu + pi, where bu and pi are preference bias of user u and popularity bias of item i, respectively. 2.3
We introduce a confidence measurement for an observed positive feedback (u, i, tui), cui =
1 (tτ + 1) − tui where tτ is the largest time stamp (in day) in the training data, and thus (tτ + 1) is used to denote the current day. Notice that we use the inverse of the difference between the current time (tτ + 1) and the time the positive feedback is issued tui because a more recent feedback is more reliable, and is thus of high confidence.
Our proposed confidence measurement shown in Eq.(3)
looks similar to but is very different from the pairwise
confidence weight in BPRC [
Finally, we have a general time-aware weighting scheme, ωui = cui, if (u, i, tui) ∈ Tu, 1, otherwise, (4) which means that we will weight the known (i.e., observed feedback) only. It is thus a ponitwise confidence weight.
With the time-aware weighting scheme and the loss
function of FISM [
In order to verify the effectiveness of our proposed pointwise weighting scheme and time-ware similarity learning (TSL) model, we use two large public datasets, i.e., MovieLens 10M (we use ML10M for short) and Netflix, in our empirical studies. ML10M contains about 10 million numerical ratings from 71567 users and 10681 items, and Netflix contains about 100 million numerical ratings from 480189 users and 17770 items. In order to simulate the TOCCF problem setting with time-aware positive feedback, for each dataset, we first remove the (u, i, rui, tui) quadruples with rui ≤ 4, and then take the (u, i, tui) triples from the remaining data.
For the resulted time-aware one-class feedback of each dataset, we further split it according to the time stamp in order to generate a copy of training data, validation data and test data. We illustrate the data generation procedure in Figure 3. Specifically, we first use 60% feedback with the smallest time stamps for training; and then from the left 40% feedback, we randomly sample 20% feedback for validation and the remaining 20% for test. We put the statistics of the resulted datasets in Table 3.
For one-class feedback in TOCCF, we adopt several commonly used evaluation metrics in ranking-oriented item recommendation or information retrieval scenarios. In particular, we check the top-K performance using Precision@K, Recall@K, F1@K, NDCG@K and 1-call@K.
For neighborhood-based methods ICF(JI) and ICF(CS), we fix the size of nearest neighbors as 20. For factorizationbased methods BPR, FISM and our TSL, we fix the dimension of latent space as d = 20 and the learning rate γ = 0.01. The iteration number in BPR, FISM and our TSL are chosen from T ∈ {100, 500, 1000} and the value of tradeoff parameters are chosen from α ∈ {0.001, 0.01, 0.1} all through the NDCG@5 on the validation data, i.e., there are nine combinations of the value of the two types of parameters.
ICF(JI) ICF(CS) BPR FISM TSL Netflix
K
ICF(JI) ICF(CS) BPR FISM TSL 3.3
We report the main results in Table 2, from which we can have the following observations, • Two neighborhood-based methods, i.e., ICF(JI) and ICF(CS), are poor regarding the recommendation performance, which is caused by the intransitivity of the similarity measurements for the scarce positive feedback. Notice that the density of the training data of ML10M and Netflix are smaller than 0.2%. • Two factorization-based methods, i.e., BPR and FISM, perform much better than the neighborhood-based methods, which is expected because of the merit of transitivity via learned latent factors. • Our proposed time-aware similarity learning method, i.e., TSL, further improves FISM and BPR significantly, from which we can clearly see the value of the temporal information and the effectiveness of our weighting scheme to integrate the time context.
For real-world deployment of a recommendation model, we usually pay more attention to its top-K performance, because that will affect users’ behaviors most. For this reason, we also check the recommendation performance with different value of K ∈ {5, 10, 15}. We show the results of NDCG@K in Figure 2. Notice that the results on other topK performance are similar, and are thus not included due to space limitation. From Figure 2, we can see that the performance ordering on different value of K over two datasets is ICF(JI), ICF(CS) < BPR, FISM < TSL, which is consistent to that of Table 2. The results on NDCG@K again show the usefulness of the temporal context and effectiveness of our time-aware weighting scheme in similarity learning.
In this paper, we study an important recommendation
problem termed time-aware one-class collaborative filtering
(TOCCF), and propose a novel time-aware similarity
learning (TSL) model based on the seminal work of factored item
similarity model [
For future work, we are interested in generalizing our
timeaware similarity learning model to more advanced
similarity learning approaches for recommendation with social and
other auxiliary data [