Enriching Article Recommendation with Phrase Awareness

TL;DR

This paper introduces PhrecSys, a phrase-aware recommendation model that enhances content-based article recommendations by incorporating phrase-level features, improving prediction accuracy and interpretability.

Contribution

It proposes a novel phrase-based approach for article recommendation that improves feature informativeness and model interpretability over existing methods.

Findings

Phrase features improve click and view prediction accuracy.

Attention mechanisms highlight meaningful text spans.

Model provides interpretable article recommendations.

Abstract

Recent deep learning methods for recommendation systems are highly sophisticated. For article recommendation task, a neural network encoder which generates a latent representation of the article content would prove useful. However, using raw text with embedding for models could degrade sentence meanings and deteriorate performance. In this paper, we propose PhrecSys (Phrase-based Recommendation System), which injects phrase-level features into content-based recommendation systems to enhance feature informativeness and model interpretability. Experiments conducted on six months of real-world data demonstrate that phrase features boost content-based models in predicting both user click and view behavior. Furthermore, the attention mechanism illustrates that phrase awareness benefits the learning of textual focus by putting the model's attention on meaningful text spans, which leads to interpretable article recommendation.

Figures

Tables

Table 1. Table 1. Click and View article pairs

Dataset	Click			View
	Train	Val	Test	Train	Val	Test
CNA	28,950	4,825	4,825	635,495	97,832	89,498
TPN	17,041	2,841	2,841	137,029	21,305	20,122

Table 2. Table 2. Number of articles in each dataset

Dataset	CNA	TPN
Articles	257,705	54,033

Table 3. Table 3. Result of click and view predictions when 𝑙𝑒𝑛 = 8 𝑙𝑒𝑛 8 \mathit{len}=8

Model	CNA-click				TPN-click				CNA-view				TPN-view
	MRR	Acc	h@3	h@5	MRR	Acc	h@3	h@5	MRR	Acc	h@3	h@5	MRR	Acc	h@3	h@5
K-NRM	.340	.132	.393	.612	.282	.095	.281	.477	.263	.087	.259	.428	.293	.104	.309	.495
MV-LSTM	.334	.127	.385	.606	.335	.140	.363	.571	.476	.283	.527	.738	.484	.293	.574	.741
CDSSM	.384	.169	.460	.683	.371	.172	.419	.625	.563	.385	.667	.792	.515	.319	.623	.789
BiLSTM-SA	.396	.182	.474	.694	.382	.178	.438	.642	.723	.621	.798	.848	.695	.555	.759	.856
TextCNN	.392	.175	.478	.686	.384	.181	.440	.648	.585	.421	.673	.804	.562	.379	.666	.809
K-NRM + Phrase	.337	.136	.415	.618	.353	.135	.369	.502	.294	.114	.302	.479	.312	.115	.329	.556
MV-LSTM + Phrase	.388	.180	.454	.679	.360	.157	.410	.621	.699	.611	.772	.837	.657	.500	.759	.831
CDSSM + Phrase	.369	.164	.431	.650	.377	.175	.433	.631	.609	.437	.716	.845	.607	.437	.709	.839
BiLSTM-SA + Phrase	.400	.190	.464	.676	.404	.186	.455	.653	.727	.632	.791	.879	.706	.582	.798	.875
TextCNN + Phrase	.395	.180	.481	.700	.391	.184	.440	.650	.709	.625	.769	.860	.704	.588	.788	.866

Equations

$$\begin{array}{rl}Acc& =\frac{1}{|U|}\sum _{u=1}^{|U|}accuracy(u)\\ accuracy(u)& =\{\begin{array}{ll}1,& \text{if}Rank(scor{e}_{c,\hat{r}})=1\\ 0,& \text{otherwise}\end{array}\end{array}$$

$$\begin{array}{r}h@N=\frac{1}{|U|}\sum _{u=1}^{|U|}h(u,N).h(u,N)=\{\begin{array}{ll}1,& \text{if}Rank(scor{e}_{c,\hat{r}})\le N\\ 0,& \text{otherwise}\end{array}\end{array}$$

$$MRR=\frac{1}{|U|}\sum _{u=1}^{|U|}\frac{1}{Rank(scor{e}_{c,\hat{r}})},$$

Full text

article recommendation, content-based recommendation, phrase mining

††copyright: rightsretained††doi: 10.475/123_4††isbn: 123-4567-24-567/08/06††conference: AAAI 2019; January 27-28, 2019; Hawaii, USA††journalyear: 2019††article: 4††price: 15.00††ccs: Information systems Recommender systems††ccs: Information systems Document representation††ccs: Computing methodologies Natural language processing††ccs: Computing methodologies Information extraction

1. Introduction

Learning from humans has shown to be effective for designing better AI systems (Guu et al., 2017). Intuitively, humans cannot memorize lengthy article information, which serves as a clue for article recommendation systems. That is, after reading through an article, there is a high possibility that only a few parts of the article will attract user attention. However, deep learning models that adopt raw text with embedding might degrade the meaning of sentences, which violates this precept.

The phrase, a small group of words standing together as a conceptual unit, can serve as a useful component of these attractive parts, or chunks. Strong data-driven methods for extracting quality phrases have been developed recently (Shang et al., 2017; Liu et al., [n. d.]). With these, high-quality phrases can be extracted without additional human effort. However, despite the regular utilization of handcrafted keywords, article recommendation seldom leverages phrase mining. Therefore, we propose PhrecSys, which utilizes phrase-level features in content-based recommendation systems. We add phrase mining to state-of-the-art content-based recommendation models and compare their performance with the original models. Moreover, we use the attention mechanism (Lin et al., 2017) and visualize the changes in attention weights during training for a clear view of the model focus.

This work includes two contributions. First, to the best of our knowledge, we are the first to incorporate phrase mining techniques to boost content-based article recommendation systems. Second, we visualize the change of attention weights during training to show systems when given phrase features: systems learn to read in a more focused way. As a great output, these focused chunks also help us interpret how these systems are producing recommendations.

2. Related Work

As the recommended targets are news articles, and the system is attempting to recommend articles based on the article currently being read, content-based recommendation models are the most related. Content-based models inherently sidestep both the cold-start issue and the problem of proposing unrelated content, as the content is the primary information needed to make good recommendations. This is especially useful for article recommendation in environments where new articles arrive regularly. Conventionally, articles are treated as bags of words, whose similarity is calculated by scoring functions such as TF-IDF (Pazzani and Billsus, 2007) and BM25 (Robertson and Walker, 1994). Recent advances in textual feature extraction with NN models (Conneau et al., 2017; Le and Mikolov, 2014) make it possible to generate article representations using NN encoders (Chen et al., 2017). With the help of word embedding, it is now possible to aggregate various neural network architectures in the textural features for different tasks. Information retrieval models can also be used for recommendation applications (cds, 2014; Chenyan et al., 2017; Shengxian et al., 2015), additionally incorporating user feedback (Ni et al., 2017) and past behavior (Debnath et al., 2008) to generate personalized recommendations.

We use phrase mining as a technique to extract the concepts that stick in readers’ minds after reading. Phrase-level features have proved useful in many different natural language tasks. Liu et al. (Liu et al., 2016) use phrases to enhance the interpretability of domain-specific documents, and demonstrate effectiveness and efficiency via various experiments. Huang (Huang, 2018) proposes a phrase-level topic model which improves semantic coherence within each topic. Further, he conducts a human study which shows that the outputs of the phrase-level topic model are easier to understand. Wang et al. (Wang et al., 2017) propose a method to translate phrases in neural machine translation (NMT), enhancing the BLEU scores on a translation task from Chinese to English. In this paper, we leverage the power of phrases in briefing and interpreting to build a better article recommendation system.

3. System Overview

An overview of our PhrecSys is shown in Fig. 1 The first step is extracting high-quality phrases in each article. For the efficiency and scalability needed by an automatic general article recommendation system, we use Autophrase (Shang et al., 2017), a data-driven and distant supervised learning algorithm.

Phrases from Autophrase with scores no lower than 0.5 are collected as candidates for phrase labeling. In this step, the phrase with the longest length is given a higher priority, i.e., longest match. Finally, phrases are labeled in each article.

The labeling process yields phrase-based articles, that is, articles injected with phrase information. We apply Glove (Pennington et al., 2014) to train the phrase embeddings for the next step. Note that punctuation is dropped in articles.

The five content-based recommendation models listed below are utilized to evaluate whether the added phrase mining component really benefits recommendation performance: TextCNN, CDSSM, MV-LSTM, K-NRM, and BiLSTM-SA. Given the article currently being read and a pool of candidate articles, these models generate the recommendation list ordered by recommendation score:

TextCNN. Similar to (Kim, 2014), convolutional neural networks with different kernel sizes are used to mimic unigram, bigram, and trigram language models, followed by a max-pooling layer and an asymmetric cosine similarity layer to calculate the score between two articles.

CDSSM (cds, 2014). Uses a single convolutional neural network to learn low-dimensional semantic vectors followed by multi-layer perceptrons (MLP), after which the normalized dot product is used to calculate the score between two articles.

MV-LSTM (Shengxian et al., 2015). Incorporates Bi-LSTM to capture contextualized local information in each positional sentence and then uses the normalized dot product to generate an interaction tensor. K-Max pooling is used to obtain the top-$k$ important information, after which MLP is used to produce the final score.

K-NRM (Chenyan et al., 2017). A kernel-based neural model for document ranking which uses a translation matrix that models word-level similarities via word embeddings and utilizes kernels to extract multi-level soft-match features. A ranking layer is followed to generate ranking scores.

BiLSTM-SA (Lin et al., 2017). Uses Bi-LSTM followed by a self-attention mechanism comprised of two linear layers with a softmax layer to generate multiple probabilities for each timestep.

In addition to the recommendation task itself, PhrecSys can visualizes the learned focus of each article via the attention technique, which helps models learn to focus on important parts while dealing with long sequences such as news articles. We then observe how the model learns with features of different granularity. Here we select BiLSTM-SA (Lin et al., 2017) for attention visualization: we store the attention weights learned in each training epoch, and then plot the tokens in the current article with different color depths according to their attention weights. Given these different color depths, we can easily perceive which part of the article the model is focusing on for recommendation in the training process. Figure 2 shows an example; more details are provided below.

4. Dataset

Two datasets – Central News Agency (CNA)111http://www.cna.com.tw and Taiwan People News (TPN)222http://www.peoplenews.tw – were provided by our collaborating partner cacaFly333https://cacafly.com for experiments. TPN and CNA are Taiwanese news websites articles written in Mandarin.

Two types of user behavior – clicks and views – are logged. Specifically, a View log is recorded when the user reads the current article $a_{c}$. When the user scrolls to the bottom of the web page to reach the recommendations, the partner system interprets this to mean that the user has finished reading $a_{c}$; a Click log is recorded when the user clicks on any of the recommendations. We generate Click log sequences to predict the articles that the user will click on next, and View log sequences to predict the articles that will be read next, accordingly.

For a Click log sequence, a positive article pair ($a_{c}$, $\hat{a_{r}}$) is generated from the article sequence if $a_{i}$ appears in the View log and $a_{i+1}$ in the Click log, $1\leq i\leq n-1$. As recency is often taken into account for recommendation (Ahmed et al., 2012; Doychev et al., 2014), we generate no more than the latest 8 positive pairs per user for experiments if the user has read many articles. For the View log sequence, we select those articles in the article sequence with the View log, and pair them to generate it. That is, ($a_{c}$, $\hat{a_{r}}$) is a pair of successive articles that are both being read by the user. Along with each positive pair, there are $m$ negative pairs ($a_{c}$, $\check{a_{r}}$), where the $\check{a_{r}}$’s are the un-clicked recommended articles for $a_{c}$.

Let $l$ be the number of article pairs in each user’s log sequence. For each user’s Click or View log sequence, the last article pair ${(a_{c},a_{r})}_{l}$ is for testing, the second-to-last is for validation, and the rest are for training. Table 1 shows the statistics of the training, validation, and testing data.

5. Experiments and Results

In this section, we first describe the datasets, model settings, and evaluation metrics, after which we discuss the performance and present the visualization of attention weights during training on word-level and phrase-level systems.

5.1. Settings and Metrics

For each dataset, we pre-train 50-dimensional phrase embeddings and word embeddings with GloVe (Pennington et al., 2014) separately, using only the articles within each dataset. The total number of articles used to train word/phrase embeddings for each dataset are listed in Table 2. Below we describe the parameter settings for different recommendation systems. Note that hinge loss is used as objective function.

TextCNN. We set CNN filter number to 32, the maximum article length to 512, and $\alpha$ in the asymmetric cosine to 0.85. During training we used the Adam optimizer with a learning rate of 0.1.

CDSSM. We set the CNN filter number to 32 and selected 128 as the MLP output dimension for article semantic representation. During training we used the Adam optimizer with a learning rate of 0.01.

MV-LSTM. We set the LSTM hidden size to 32 and the top-$k$ value to 512. Two layer MLPs were applied at the end of model, and 64 and 1 were the number of output dimensions respectively. During training we used the Adam optimizer with a learning rate of 0.01.

KNRM. We set the kernel number to 32 and the sigma value to 0.05. During training we used the Adam optimizer with a learning rate of 0.001.

BiLSTM-SA. We set the LSTM hidden size to 32 for each direction, $d_{a}$ to 100 for the fully-connected layer inside self-attention, and the attention number $r$ to 15. During training we used the Adam optimizer with a learning rate of 0.01.

We adopted several evaluation metrics to fairly evaluate the recommendation performance. To evaluate the performance of recommendation, per the literature we used the mean reciprocal rank (MRR) (Voorhees et al., 1999), accuracy (Acc) (Devooght and Bersini, 2017), hit ratio at 3 (h@3), and hit ratio at 5 (h@5) (Kumar et al., 2017). The hit ratio h@k intuitively measures whether the test item is present in the top-$k$ list.

[TABLE]

[TABLE]

Assuming $\left|U\right|$ testing instances in the testing set, each testing instance $u\in U$ contained one positive pair ($a_{c}$, $\hat{a_{r}}$) and $m$ negative pairs ($a_{c}$, $\check{a_{r}}$). The average MRR was calculated as

[TABLE]

where $Rank(score_{c,\hat{r}})$ was the rank of the positive pair score among all scores of pairs in $u$.

5.2. Recommendation Performance

We examined the performance on two datasets (News_1, News_2), both of which exhibit two types of behavior – Click and View. The five models described in Section 1 and  5.1 were evaluated to compare their performance with and without phrase mining: the results are shown in Table 3.

For the prediction of the next clicked article (the article that the user clicks on next), most models are slightly enhanced by phrase mining. However, for the next viewed article (the article that the user reads next), the performance of all models is improved significantly when leveraging phrase-level features. This may be because text content might not be the first priority to users when clicking, as eye-catching titles or images may hold a special attraction. In this case, though phrases capture the key points of an article, their benefit is limited. For view prediction, in contrast, the text content is essential. We observe that a large amount of views come from websites other than the current news media and very often from search engines. In these cases, the users’ intention is clearly a specific topic that interests them and hence they tend to read contents that are logically related successively. Another reason is that as users read through the current article, its content becomes a important factor in the surfing history. The notable enhancement of phrase-level features shows that phrases can enhance the model’s ability to learn the relatedness of contents between articles.

5.3. Visualization

To visualize the focus points being learned by the model, we record the attention weights of composite units (words or phrases plus words) of the current article in each training epoch and observe their change over time. Figure 2 shows the effect of phrase information. Figure 2a is the article parsed and trained with phrase-level units, whereas Figure 2b is the article parsed and trained with word-level units. The attention weights of epoch 5 illustrate the overall result: when phrase information is provided, attention weights saturate to a reasonable number of units. In comparison, without phrase mining, weights are evenly distributed, with no obvious key units to be found in the visualization.

Comparing Figure 2a and Figure 2b from Epochs 1 and 3 to Epoch 5, we observe that the attention weights are both evenly distributed and show no particular focus on anything in the beginning. Later, in Epoch 3, training with phrase-level features causes weights to be more concentrated and accurately focused on meaningful parts; finally, in Epoch 5, training with phrase-level information successfully leads to fewer but more informative cues. We conclude that phrases contain more compact information and are more easily utilized for recommendation models.

6. Conclusion

In this paper, we use data-driven phrase mining to automatically extract high-quality phrases for news article recommendation. We show that the proposed approach yields improvements in both click and view prediction. Additionally, we visualize the learning process of both phrase-level and word-level models to illustrate that the merit of phrase mining for recommendation is its ability to put the focus on key units. With this advantage, the visualization itself can be interpreted and is of great value to public opinion analytics.

Acknowledgements.

The authors would like to thank our collaborating partner cacaFly, the leader of digital advertising in Taiwan, for providing user website log in this paper. Research of this paper was partially supported by cacaFly, under the collaborative project contract O5T-1060629-1Cb.