PageRank (Brin and Page, 1998) is a link analysis algorithm and it assigns a numerical weighting to each element of a hyperlinked set of documents, such as the World Wide Web, with the purpose of “measuring” its relative importance within the set. The algorithm may be applied to any collection of entities with reciprocal quotations and references. The numerical weight that it assigns to any given element E is referred to as the PageRank of E. PageRank normalizes the number of links on a document by not counting each of them as equal. PageRank can be defined as follows (Equation 1): P n+1 = : i

X j:kjout>0

pn Aij kojut + : j

X j:kjout=0 pjn + 1 N

N (1) where kjout = Pl Alj is the outdegree of node j, is the teleportation parameter, and n is the iteration number. The PageRank score Pi of node i can be interpreted as the average fraction of time spent on node i by a random walker who with probability follows the network’s links and with probability 1 teleports to a random node. We consider = 0:5 throughout this paper since it is the accepted choice for citation networks (Chen et al., 2007) . 3.2

CiteRank (Walker et al., 2007) was designed specifically for ranking papers in a citation network. CiteRank performs a random walk on an aggregated citation graph but initiates the walk from a recent paper chosen with the probability that depends on its age. Authors estimated parameters of the random walk by fitting papers’ CiteRank score to the number of citations accrued by papers over some time period. Let us suppose M is a transfer matrix with elements Mij = 1=Lj if paper j cites i and 0 otherwise. The probability that a researcher follows the citation links to encounter a paper is defined as in Equation 2:

C~ = I0:~ + (1 )M:~ + (1 )2M 2:~ + ::: (2) where I0 is an identity matrix, i =exp agei= is the probability of initially selecting paper i, agei is the age of the paper and is the characteristic decay time. In this paper, we consider = 0:5 and = 2:6 years, as specified by Walker et al. (2007). 3.3

Our proposed model Time-Attentive Ranking is based on the notion that ‘An inventor or patentee can find patents by following citations links back in time from a particular patent’. The number of paths that can be attenuated between patent pi and pj can be expressed as a contagion matrix M given by Equation 3:

N 1 MN; = X i i

A N i=1 (3) where An is the adjacency matrix of patents citing each other for a particular year tn and is the probability of following a citation link. The more paths there are from patent pi to pj , the higher the likelihood that an inventor will find pj by following citation chains from pi, which is similar to -Centrality (Bonacich, 1987) and Katz centrality (Katz, 1953) metrics. Since the existing contagion matrix does not account for time and hence weights each edge equally, the authors propose a retained adjacency matrix which is given by Equation 4: Rn; (i; j) = 8 N ni ; if picites pj and > < >:0; t(pi) = tni otherwise tn (4) where < 1 is the retention probability given to attach more weight to a recent patent and decrease the weight as the patent keeps ageing. The contagion matrix can then be written as Equation 5 (using Equation 3 and 4):

N 1 EM N; ; = X i i

R N i=1 (5) and hence the score of a patent pj at the end of a time period [ti; tN ] is given by EM N (j) = Pi EM N (i; j). For our experiments we consider the best possible settings by empirically setting = 0:1 and = 0:3. Elsewhere in the paper we refer to the EM score as T for uniformity and ease of comprehension. 4 4.1

For this study, we used two datasets: (1) European Patent (EP) collection from the MAtrixware REsearch Collection (MAREC)2 and (2) US patents dataset collected by Kogan et al. (2017) that spans the period between 01-01-1926 and 11-02-2010. To the best of our knowledge, there exists no study of a similar kind on the European Patents, which is why we chose to work with the former collection from MAREC. However, this presents a unique challenge of finding a respective gold-standard list of “milestone” patents, which is not available. Hence for this collection, we resort to a qualitative evaluation as described in later sections. To compare our proposed approach’s performance against the state-of-the-art and perform a quantitative evaluation, we repeated our experiments on the USPTO dataset as well. Additionally, we also performed validation of our model’s performance by a deep learning technique as suggested in Chung and Sohn (2020) to identify a patent’s grade in determining in value. 4.2

MAREC collection We only considered granted patents from the ‘EP’ collection. For uniformity,

In this section, we present a qualitative of the results followed by a comprehensive analysis of the same. We ran the experiments on networks that were spliced in time (yearly) over a ten year period (1998-2008). The same set of experiments were carried out on networks comprising of all patents, patents belonging to a certain sector and patents belonging to a certain category. On studying the indegree and out-degree of patents, we observed that the degrees are very skewed, i.e., only a handful of patents gets a large number of citations while most of the patents in the network have less than ten citations. Hence, this network follows a similar pattern to that of a scholarly article citation network (albeit with more skewness). Hence, the network-flow algorithms that can be employed with paper citation networks can also be adapted here. Moreover, there is a strong correlation between the in-degree and out-degree of the patents in both collections, which implies that highly-cited patents tend to be cited by other highly-cited patents and to cite other highly-cited patents (Ren et al., 2018) .

Qualitative Comparison of Top Patents : For an intuitive understanding of how the different network-flow metric scores affect the rank, it is important to observe the top-ranked patents according to the PageRank score P , CiteRank score C and Time-Attentive Rank score T . As mentioned earlier, each patent is endorsed with several classification codes that classify them into sectors, categories, sub-categories etc. The highest level of classification is according to sectors (A-H). Each sector consists of several categories (A61K, A61P, ...), each category consists of several sub-categories and so on. A single patent can belong to several sectors and categories.

From Table 2, we can observe significant changes in the ranking of the patents. The tables reveal that there are more recent patents granted after the year 2000 in the top-10 list ranked by Time-Attentive Ranking than that produced by either PageRank or CiteRank. To be precise, the Time-Attentive ranking method includes five patents granted after the year 2000 in the top-10 list, while for PageRank and CiteRank, it is four out of ten for both. Of course, the difference is even more pronounced as we go deeper in the lists, say, top-15, top-20 and so on, which we could not present here due to space constraints.

For comparison, the list of top five patents on the basis of citation count is presented in Table 1. One can observe that none of the three metrics ranks the patents from Table 1 in their top five list. In fact, within the top fifteen results, only patent EP0272189 and EP0364618 feature in the lists compiled according to PageRank and CiteRank scores, while patent EP0527247 is listed by PageRank only. The rest do not find a place in the top-10 of any network-based score list. This corroborates our initial hypothesis that is simply acquiring a high citation count does not indicate the importance of a patent.

Next, we perform the same set of experiments over individual sectors of patents. Similar to the trend shown by the complete network, for sector B which has the highest number of patents, we observe from Table 3 that Time-Attentive Ranking features the Aluminum-stabilized ceria catalyst compositions and method of making same.

Particulate filter for purifying exhaust gases of internal combustion engines Cyclone dust extractor COMPOSITE IC CARD A method and system for simultaneous bi-directional wireless communication between a user station and first and second base stations Process for the aerobic biological purification of water Carbon fiber composite material and process for producing the same PHOTORESIST STRIPPER COMPOSITION Use of complexes among cationic liposomes and polydeoxyribonucleotides as medicaments Imidazolylmethyl-pyridines.

Aluminum-stabilized ceria catalyst compositions and method of making same.

Particulate filter for purifying exhaust gases of internal combustion engines Cyclone dust extractor COMPOSITE IC CARD A method and system for simultaneous bi-directional wireless communication between a user station and first and second base stations Process for the aerobic biological purification of water Carbon fiber composite material and process for producing the same PHOTORESIST STRIPPER COMPOSITION Use of complexes among cationic liposomes and polydeoxyribonucleotides as medicaments Imidazolylmethyl-pyridines.

Aluminum-stabilized ceria catalyst compositions and method of making same.

Particulate filter for purifying exhaust gases of internal combustion engines Cyclone dust extractor COMPOSITE IC CARD Evaporator with electronic circuit printed on a first side plate A method and system for simultaneous bi-directional wireless communication between a user station and first and second base stations Carbon fiber composite material and process for producing the same Multiple signal transmission device.

Process for the aerobic biological purification of water PHOTORESIST STRIPPER COMPOSITION

Date more recent patents in their top five as compared 5.1.3

to their counterparts.

While it is interesting to study the complete network and find the most influential patents as identified by Time-Attentive Rank, it does not deliver us

P – 0.4981 0.3921

C a lot of information. On the other hand if we limit the patent citation network by categories, it could provide us some insights as to which technologies have been gaining momentum in the last few years of the patent data. The total number of categories in the patent database exceeds hundred. Not surprisingly, the distribution of patents against categories is also skewed. For brevity, we present only the results for the most popular category A61K.

From Table 4, we observe a certain peculiarity. None of the top five patents ranked by the Time-Attentive Ranking mechanism is a post-2000 patent. This is interesting because it implies that while Time-Attentive rank gives more weightage to recent citations, it does not bias towards recent patents, thus maintaining a balance between older and newer patents. So, the top-ranked patent in all three cases is the same indicating that EP0776864 is indeed the most important patent in category A61K.

hypothesis hinges on the ranking of patents over a network metric based score, it is imperative that the lists generated by PageRank and CiteRank and TimeAttentiveRank will be different in their ordering of elements (ranks). As the lists are quite long, their scores are not directly comparable, and for a given depth d the two lists may not even have the same set of elements, we will have to resort to indefinite ranking (Webber et al., 2010) . To this end, we employ rank-biased overlap (RBO) to measure the similarity and agreement between the two lists. The RBO values for the year 2008 compared over the complete list of ranked results is presented in Table 7. The Rank-Biased Overlap is defined as in Equation 6.

RBO(S; T; p) = (1 1 p) X pd 1:Ad

(6) d=1 where S and T are two indefinite ranked lists. p stands for user’s persistence, which determines how steep is the decline in weights: the smaller p, the more top-weighted is the metric. Ad, agreeement can be defined as the proportion of S and T that are overlapped at depth d. Rank-biased Overlap falls in the range [0; 1], where 0 means disjoint, and 1 means identical. While RBO is the agreement score between two indefinite lists, we are more concerned with the top-k elements in the lists and hence RBO@k provides us a better measure to compare the top-ranked elements. It is imperative to note that RBO > RBO@k. For our case, we empirically consider k = 20 and p = 0:9.

Tables 5, 6 and 7 present the RBO confusion matrix. We can clearly observe a pattern here. The overlap between CiteRank and TimeAttentive ranked lists are certainly more than the overlap (agreement) between PageRank and TimeAttentive Rank, which confirms our intuition that recent patents receive more preference in the weighted citation measures rather than unweighted citations of PageRank. 5.2

For this collection, we adopt a different approach for carrying out our experiments. The experiments on the MAREC patents were solely based on network flow metrics, which we could not assess quantitatively due to the lack of a standard baseline. Instead, for the US patents collections, we compare our approach against the state-of-the-art Re-scaled PageRank method proposed by Mariani et al. (2019) to identify milestone patents. As a second objective, we wanted to determine the value added by textual content in determining a patent’s worth. This objective stems from similar studies on patents where it was shown that exploiting the multimodal nature of patents yields better prediction performance (Chakraborty et al., 2020) . For this purpose, we adapt the deep learning approach proposed by Chung and Sohn (2020). Due to the incompatibility of NLP based approach proposed by Chung and Sohn (2020) and network flow metrics-based approaches such as the one by Mariani et al. (2019) and ourselves in this paper, we only adopt the deep learning approach (DEP-net) to determine a patent’s grade, which is another measure of patent’s importance. As per Chung and Sohn (2020), a patent’s quality is assigned one of three grades (A, B, or C) based on the average number of forward citations per year. The deep learning approach is briefly summarised below: • A patent grade (A, B, or C) is assigned based on a threshold determined by the average forward citations accrued per year by the patent. • Textual content (abstract and claim) from the patent data is extracted along with several other indices such as number of claims, number of inventors, number of backward citations, number of IPCs, etc. • Abstract and claims are transformed (vectorized) into word embeddings as matrices. • A deep neural network composed of Bi-LSTM layer is added to the CNN structure using multiple filters that fuses the four components (abstract, claims, indices, network-metric score) as input to train and evaluate a patent’s quality. Finally, we also evaluate the patent quality for test data.

It is to be noted that we add an extra component to the original model proposed by Chung and Sohn (2020), i.e., the network-metric score. Both the rescaled PageRank score (R) and the Time-Attentive Rank score are fed separately as inputs to the deep neural model. To simplify things, we retained the parametric setting of the neural model as proposed by Chung and Sohn (2020). Finally, the features from the abstracts, claims, indices and networkflow metrics are fused and used as inputs to the fully connected layer. The loss function was crossentropy, and the activation function was softmax. We label this model as DEP-netPlus (as we add value to the DEP-net model).

Strumsky and Lobo (2015) listed 175 patents carefully selected “on the basis of consultation with engineers, scientists, historians, textbooks, magazine articles, and internet searches”. The patents in the list “all started technological pathways which affected society, individuals and the economy in a historically significant manner” (Strumsky and Lobo, 2015) . These significant patents thus provide a good “ground-truth” set of patents that can be used to discern the ability of different metrics to uncover these significant patents. The complete list of these patents can be found in Appendix C of (Strumsky and Lobo, 2015) . Presence in the list of significant patents by Strumsky and Lobo is a binary variable: a patent is either in the list or not. We can therefore study the ability of the metrics to rank these outstanding patents as high as possible, in agreement with the objectives of this paper. While there are 175 significant patents in the Strumsky-Lobo list, we restrict our analysis to those patents that were issued within our dataset’s temporal span and remove those that are absent in our dataset. This leaves us with M = 112 significant patents. 5.2.2

In this section we inspect the top-ranked patents. For simplicity, we focus on the top-10 patents as ranked by PageRank P and Re-scaled PageRank R and our Time-AttentiveRank T scores (Table 8). From Table 8, we can observe that the top-10 patents by Re-scaled PageRank span a wider temporal range (1942–2010) than the top-10 by PageRank (1942–1996), which is a direct consequence of the age-bias removal. The same temporal span is retained by the Time-Attentive Rank as well. However, it is noteworthy that our proposed method can pick more (3) patents from Strumsky-Lobo’s list of significant patents. Among the ten top-ranked patents, two are from 2010 (the last year in the dataset) and received only one citation. This happens because only a few among the most recent patents received citations, which results in temporal windows with a large fraction of patents with zero citations. Thus, within such a temporal window, a patent can achieve large T score thanks to one single citation. A possible solution for this issue is to only include the patents whose temporal windows contain a certain minimal number of incoming citations. Another observation is that both the Re-scaled PageRank and Time-Attentive rank do not necessarily rank patent with grade A in a higher position, so the ranking is not solely dependent on the citation count but also on the network structure. 5.2.3

To illustrate the importance of including networkflow based metric as a component, we performed the patent grade classification as described in (Chung and Sohn, 2020) . We used the same dataset Process for amplifying, detecting, and/or-cloning nucleic acid sequences Process for amplifying nucleic acid sequences (*) Process for producing biologically functional molecular chimeras Method for the direct analysis of sickle cell anemia Bubble jet recording method and apparatus in which a heating element generates bubbles in a liquid flow path to project droplets Microorganisms having multiple compatible degradative energygenerating plasmids and preparation thereof Method of screening for cDNA encoding novel secreted mammalian proteins in yeast Software version management system Specific DNA probes in diagnostic microbiology SElectrophotography Optical element holding device, lens barrel, exposing device, and device producing method (*) Process for producing biologically functional molecular chimeras Electrophotography Carbon nanotube arrays Network service for modularly constructing a software defined radio Method of screening for cDNA encoding novel secreted mammalian proteins in yeast Process for amplifying, detecting, and/or-cloning nucleic acid sequences Mutant dwarfism gene of petunia Method for the direct analysis of sickle cell anaemia Process for amplifying nucleic acid sequences Process for amplifying nucleic acid sequences (*) Process for producing biologically functional molecular chimeras Electrophotography (*) Personal computer Carbon nanotube arrays Network service for modularly constructing a software defined radio Method of screening for cDNA encoding novel secreted mammalian proteins in yeast (*) Using recombinant DNA to produce an altered antibody Process for amplifying, detecting, and/or-cloning nucleic acid sequences Method for the direct analysis of sickle cell anemia

Date superiority of our model in capturing not only the ductor” technology collected within the temporal “importance” of a patent but also in evaluating the span of 2000 to 2015. We carried out the same prepatent’s grade. processing steps along with down-sampling of the data or certain classes to maintain uniformity. The results of experiments performed with an additional component, i.e., our proposed TimeAttentiveRank score to the deep learning model which we refer to as DEP-netPlus are presented in Table 9. 6

of three network-flow algorithms for this purpose From the table, we can clearly observe that the on two different datasets. On the second dataset, classification model is enhanced by the inclusion we used a deep-learning-based approach to fuse of a network flow metric that account for the netpatent content along with network flow metrics, to work effect due to citations. This also confirms the compare against state-of-the-art and discovered that our proposed approach results in better performance both in identifying “milestone” patents as well as improving the patent grade prediction. From the experimental results, we summarily concluded that raw citation count is not enough to capture the importance of a patent since it does not take into account the age of citations. When accounted for the same using a balanced metric like Time-Attentive ranking, we are guaranteed to identify potential patents that are likely to spur technological growth in the near future. We also identified top patents per category and sector, which can help in the identification of niche areas for innovation. Although patent retrieval is a recall-oriented task, these criteria may also help in re-ranking the results against a keyword search for patents.

As part of our future work, we would like to study the importance of geographical location on influential patents, such as the country they originated from, the citations received from other countries and so on. We also plan to experiment with the various granularity of time such as a month, year, 5-year period, and so on.